Skip to contentSkip to conversation

Crazy idea: Design open-source graphics chip

OT Timothy Miller 17 participants 46 messages Jan 28, 2004 – Feb 2, 2004

lore
Share & tools
lore
#0 lore raw

This is somewhat off-topic, so we shouldn't discuss it TOO much on-list, but I feel it's relevant to the state of affairs with Linux.

I haven't looked at what's available on opencores.org, but one of the biggest problems we seem to have is with getting high-quality graphics cards that are compatible with Linux in the sense that there are open specs and there's an open-source driver. Oh, and we'd like to have something decent.

I have personally designed a graphics engine. Actually, I would say that I did maybe 90% of the Verilog coding on it, and about 20% of the back-end (place, route, etc.) work. I also did 100% of the X11 software (DDX) and 0% of the kernel driver code. I wouldn't call it a masterpiece of engineering compared to the latest and greatest high-end 3D and CAD graphics chips, but it's a powerful workhorse used in most of the air traffic control graphics cards and medical imaging cards that my employer sells (10 megapixel displays are easy for us). Were you to read the manual on it, you'd think some of it was a bit unusual (such as the way you issue rendering commands), because it WAS my first ASIC ever. I did meet all of our performance goals. And I've come a long way since then. (Unfortunately, this may sound like a plug, but I have competing desires to be humble about what I did but also not to publically say something that might understate the value of my employer's products. I also feel a sense of pride in my accomplishment.)

That being said, I would LOVE to be involved in the design of an open-source graphics chip with the Linux market primarily in mind. This is a major sore point for us, and I, for one, would love to be involved in solving it. With an open architecture, everyone wins. We win because we have something stable which we can put in main-line Linux, and chip fabs win, because anyone can sell it, and anyone can write drivers for any platform.

Imagine ATI and nVidia competing on how they can IMPROVE the design over one another but being obligated to release the source code. I know... wishful thinking. But I know a variety of ways that chips and boards could be made with respectable geometries (90nm) and high performance. No more being at the mercy of closed-development graphics chip designers who make Linux an after-though if they even think of us at all.

Please forgive my off-topic intrusion.

#1 lore raw
Quoted 2 lines

No more being at the mercy of closed-development graphics chip designers who make Linux an after-though if they even think of us at all.

I don't know if we are at the mercy of closed-development code. In a way you are always at the mercy of someone. Say there was an open driver development push for an open GPU, someone would still have to code for it. Someone would make decisions and someone would disagree.

We would in effect be at the mercy of those. We are at the mercy of the kernel coders as we speak. Decisions could be made that affect you and me, right now. Suddenly it might be decided that AC97 is useless and support will not be continued. Unlikely but it could happen. So then what? I don't think i could just jump in and code that. Certainly there are people that could, but what if those patches were to be ignored? Forked?

Oh ... don't get me wrong, i think that the conceptual idea is awesome. Personally, i wouldn't know where to begin, but can the open source community compete with Nvidia and ATI? afterall this goes beyond software, it delves into hardware. Sure there are people with the knowledge, maybe even with the means, but i doubt the financial backing would be there from the get go.

But hey. I hope i'm wrong and open hardware is the next big thing. One request though. Make the cooler quiet please :)

... One afterthought on the mercy bit. I had issues with NVidia's 5328 drivers on 2.6 ... it was frustrating and all, but if i thought the path of least resistance i doubt i'd be running Linux. Then again... I quit running Windows because i couldn't take it anymore.

Signature · 9 lines
with kind regards,
  Christian Unger

 - < > - < > - < > - < > - < > - < > - < > - < > -
 
  Alt. Email:  [email protected]
  ICQ:         204184156
  Mobile:      0402 268904
  Web:         http://naiv.sourceforge.net
#2 lore raw
Quoted 3 lines · Christian Unger

Oh ... don't get me wrong, i think that the conceptual idea is awesome. Personally, i wouldn't know where to begin, but can the open source community compete with Nvidia and ATI? afterall this goes beyond software, it delves

Well I think the first problem is that the idea is currently too big. If someone were to do this sucessfully they would make the first open cards something like a Trident 8900C. Something small but usable for people who need it. The next cards would add onto it, and so on and so on until you got a base that would meet the 3D ATI/Nvidia needs. Trying to aim for the top at the beginning is a great way to crater.

Signature · 4 lines
Stephen John Smoogen		[email protected]
Los Alamos National Lab  CCN-5 Sched 5/40  PH: 4-0645
Ta-03 SM-1498 MailStop B255 DP 10S  Los Alamos, NM 87545
-- So shines a good deed in a weary world. = Willy Wonka --
#3 lore raw
Quoted 5 lines

someone were to do this sucessfully they would make the first open cards something like a Trident 8900C. Something small but usable for people who need it. The next cards would add onto it, and so on and so on until you got a base that would meet the 3D ATI/Nvidia needs. Trying to aim for the top at the beginning is a great way to crater.

In case of manufacturing, design, etc. the idea of release small, release early
would be too costly for it to survive, me thinks.

Unless we had software to simulate a graphics chip and its software. Like vmware but with emulation of hardware and software on that virtual hardware. Lots of CPU power would be required there, but then if we could write code for a virtual hardware emulator, "writing" such a chip and then designing it would be feasible.
This is intriguing.

Regards,
Maciej

#5 lore raw
In reply to Christian Unger
Quoted 5 lines · Christian Unger
Quoted 2 lines

No more being at the mercy of closed-development graphics chip designers who make Linux an after-though if they even think of us at all.

Oh ... don't get me wrong, i think that the conceptual idea is awesome. Personally, i wouldn't know where to begin, but can the open source community compete with Nvidia and ATI? afterall this goes beyond software, it delves into hardware. Sure there are people with the knowledge, maybe even with the means, but i doubt the financial backing would be there from the get go.

We cannot compete with Nvidia or ATI or 3Dlabs or Matrox or even S3.

The real question we have to ask ourselves is, what would be the market demand for a graphics card that is 3 generations behind the state of the art and over-priced, the only advantage being that it's a 100% open architecture?

I don't have $100k to have it fabricated, so we have to goad some company into doing it for us, and given the volumes, they'll have to charge way more than it's worth if you compare its capabilities against ATI et al.

I've got some great ideas for how to do this chip, but they're frankly nothing revolutionary. The obvious test bed is an FPGA. That imposes serious limitations on what kind of logic utilization and performance we can get. The ASIC version can be clocked faster, but we dare not put in untested logic. (And we can't afford the tools necessary to do the proper simulation.)

So, the big question: How many units a year would be sold for an underpowered, over-priced graphics card that just happens to be 100% open and 100% supported?

#4 lore raw
In reply to Stephen Smoogen

Quote from Stephen Smoogen <[email protected]>:

Quoted 11 lines · Christian Unger
Quoted 3 lines

Oh ... don't get me wrong, i think that the conceptual idea is awesome. Personally, i wouldn't know where to begin, but can the open source community compete with Nvidia and ATI? afterall this goes beyond software, it delves

Well I think the first problem is that the idea is currently too big. If someone were to do this sucessfully they would make the first open cards something like a Trident 8900C. Something small but usable for people who need it. The next cards would add onto it, and so on and so on until you got a base that would meet the 3D ATI/Nvidia needs. Trying to aim for the top at the beginning is a great way to crater.

A simple framebuffer connected to the parallel port would be trivial to make, and it would be suprisingly useful for simple applications such as word processing. Literally a handful of components soldered on to a piece of stripboard and well written drivers for the framebuffer console and X is all it would take. No need for anything remotely fancy at first.

John.

#6 lore raw
In reply to Timothy Miller
Quoted 4 lines

The real question we have to ask ourselves is, what would be the market demand for a graphics card that is 3 generations behind the state of the art and over-priced, the only advantage being that it's a 100% open architecture?

Err, well there are always the server and embedded markets, if the device was cheap enough.

Quoted 11 lines

I don't have $100k to have it fabricated, so we have to goad some company into doing it for us, and given the volumes, they'll have to charge way more than it's worth if you compare its capabilities against ATI et al.

I've got some great ideas for how to do this chip, but they're frankly nothing revolutionary. The obvious test bed is an FPGA. That imposes serious limitations on what kind of logic utilization and performance we can get. The ASIC version can be clocked faster, but we dare not put in untested logic. (And we can't afford the tools necessary to do the proper simulation.)

WHAT!? You are making the project out to be several orders of magnitude more difficult and expensive than it is.

Did you know that you can generate a 625-line TV signal with little more hardware than a Z80 CPU? Some 8-bits actually did that.

Quoted 3 lines

So, the big question: How many units a year would be sold for an underpowered, over-priced graphics card that just happens to be 100% open and 100% supported?

Quite a few. Think of the TV-connected embedded appliance market, for example. Displaying a static menu of choices isn't exactly very demanding.

John.

#38 lore raw
In reply to Timothy Miller
Quoted 33 lines · Timothy Miller
Quoted 9 lines · Christian Unger
Quoted 2 lines

No more being at the mercy of closed-development graphics chip designers who make Linux an after-though if they even think of us at all.

Oh ... don't get me wrong, i think that the conceptual idea is awesome. Personally, i wouldn't know where to begin, but can the open source community compete with Nvidia and ATI? afterall this goes beyond software, it delves into hardware. Sure there are people with the knowledge, maybe even with the means, but i doubt the financial backing would be there from the get go.

We cannot compete with Nvidia or ATI or 3Dlabs or Matrox or even S3.

The real question we have to ask ourselves is, what would be the market demand for a graphics card that is 3 generations behind the state of the art and over-priced, the only advantage being that it's a 100% open architecture?

I don't have $100k to have it fabricated, so we have to goad some company into doing it for us, and given the volumes, they'll have to charge way more than it's worth if you compare its capabilities against ATI et al.

I've got some great ideas for how to do this chip, but they're frankly nothing revolutionary. The obvious test bed is an FPGA. That imposes serious limitations on what kind of logic utilization and performance we can get. The ASIC version can be clocked faster, but we dare not put in untested logic. (And we can't afford the tools necessary to do the proper simulation.)

Quoted 4 lines

So, the big question: How many units a year would be sold for an underpowered, over-priced graphics card that just happens to be 100% open and 100% supported?

With the press Linux is getting from the IBM/Linux advertisements for the US football games, etc., methinks it won't be long before NVidia and all the rest go open-source, just to jump onto that band-wagon. They just need a smart way to protect their intellectual property.

Cheers,
Dick Johnson
Penguin : Linux version 2.4.24 on an i686 machine (797.90 BogoMips).
Note 96.31% of all statistics are fiction.

#7 lore raw
In reply to John Bradford
Quoted 5 lines · John Bradford
Quoted 1 line

The real question we have to ask ourselves is, what would be the market demand for a graphics card that is 3 generations behind the state of the art and over-priced, the only advantage being that it's a 100% open architecture?

Err, well there are always the server and embedded markets, if the device was cheap enough.

Ah, but it won't be. Low-volume ASICs are expensive. The chip itself would probably be around $150, not counting $100k NRE. Then you have to pay for the board, make up for the NRE, and make some profit to make it worth while. How much are YOU willing to pay?

Quoted 4 lines
Quoted 4 lines

I don't have $100k to have it fabricated, so we have to goad some company into doing it for us, and given the volumes, they'll have to charge way more than it's worth if you compare its capabilities against ATI et al.

Quoted 8 lines
Quoted 1 line

I've got some great ideas for how to do this chip, but they're frankly nothing revolutionary. The obvious test bed is an FPGA. That imposes serious limitations on what kind of logic utilization and performance we can get. The ASIC version can be clocked faster, but we dare not put in untested logic. (And we can't afford the tools necessary to do the proper simulation.)

WHAT!? You are making the project out to be several orders of magnitude more difficult and expensive than it is.

Did you know that you can generate a 625-line TV signal with little more hardware than a Z80 CPU? Some 8-bits actually did that.

Certainly. But when you can get perfectly good open-source drivers for an ATI Rage 128 and the board for $15 from a Taiwanese manufacturer, who's going to want what you're describing?

The thing you have to keep in mind is that in order for this open arch board to get developed, someone has to be willing to invest in fabricating it, and that means it has to be somewhat competitive and a significant performer.

From the mouth of someone who has done a graphics ASIC and numerous FPGA designs also in graphics and who has worked on graphics boards in air traffic control, medical, and workstation console markets and who has written X-server modules for Number 9 i128, Matrox G450, Permidia 2 and 3, Radeon 7500 and 9000, my own graphics chip, probably a number of chips I've forgotten AND who has been very performance and cost conscious the whole time: It is MORE complicated than I make it sound.

That doesn't mean it's not doable. :)

Quoted 8 lines
Quoted 3 lines

So, the big question: How many units a year would be sold for an underpowered, over-priced graphics card that just happens to be 100% open and 100% supported?

Quite a few. Think of the TV-connected embedded appliance market, for example. Displaying a static menu of choices isn't exactly very demanding.

This sort of thing is ALREADY available with open-source drivers.

Whatever we design is going to be EXPENSIVE. So, regardless of the fact that an ATI All-in-Wonder Radeon 9000 is over-powered for job you describe, that board will be cheaper than what we could produce.

Because of certain invariant costs, there is a performance point below which it is not worth it. Because of non-invariant costs, there is a performance above which it is not worth it. There may or not be a point where the compromize makes it worth doing.

Now, this all assumes that it's completely a hobbyist project. If we were to design something that was, in principle, a good performer, but we couldn't simulate, debug, and fabricate it, we MIGHT be able to convince some companies to do that FOR us. And they might even be able to enhance it in ways that would make it compete on performance.

But it's still going to be expensive.

#39 lore raw
In reply to Richard B. Johnson
Quoted 8 lines · Richard B. Johnson

With the press Linux is getting from the IBM/Linux advertisements for the US football games, etc., methinks it won't be long before NVidia and all the rest go open-source, just to jump onto that band-wagon. They just need a smart way to protect their intellectual property.

Indeed! And this may make the whole idea of an open-arch GPU a pipe-dream.

Honestly, we don't _need_ an open-arch GPU. We just need something whose register set is fully publically documented.

But an open-arch GPU would be NEAT, though. :)

Hmmm... If I understand this right, one of the reasons nVidia doesn't release open source drivers is that they don't own all of the IP in their cores. I wonder what that IP is and if the open-source community couldn't collaborate to produce LGPL-like replacements.

#8 lore raw
In reply to Timothy Miller

Quote from Timothy Miller <[email protected]>:

Quoted 1 line

[snip]

Quoted 7 lines
Quoted 2 lines

Err, well there are always the server and embedded markets, if the device was cheap enough.

Ah, but it won't be. Low-volume ASICs are expensive. The chip itself would probably be around $150, not counting $100k NRE. Then you have to pay for the board, make up for the NRE, and make some profit to make it worth while. How much are YOU willing to pay?

Yes, for real world devices there is always the point to be considered that you can buy a $15 card, and if your requirements are simple enough, simply ignore the bits that you don't need, and drive it with open source code. The cost of developing a much simpler and slightly cheaper solution outweighs the potential saving, so there is no real incentive to develop it.

However, if the much simpler but cheaper project already existed, and was as little as $1 cheaper to produce in volume, would embedded manufacturers use it? I suspect they would.

Quoted 4 lines

The thing you have to keep in mind is that in order for this open arch board to get developed, someone has to be willing to invest in fabricating it, and that means it has to be somewhat competitive and a significant performer.

Well, the cost of fabricating depends on the device. I was basically thinking of a 68000, an EPROM and a SIMM on a piece of stripboard, some ribbon cable and a DB-25 connector.

Maybe our goals are somewhat different :-)

John.

#9 lore raw
Quoted 5 lines · John Bradford

Well, the cost of fabricating depends on the device. I was basically thinking of a 68000, an EPROM and a SIMM on a piece of stripboard, some ribbon cable and a DB-25 connector.

Maybe our goals are somewhat different :-)

Very different. What you're describing is a dumb terminal.

What I'm describing is a PC console graphics card that will let someone play Quake III at a reasonable framerate.

Isn't that what most people want?

And the performance disparity between what you're describing and what I'm describing is enormous!

#42 lore raw
In reply to Timothy Miller
Quoted 6 lines · Timothy Miller

We cannot compete with Nvidia or ATI or 3Dlabs or Matrox or even S3.

The real question we have to ask ourselves is, what would be the market demand for a graphics card that is 3 generations behind the state of the art and over-priced, the only advantage being that it's a 100% open architecture?

I think I can do better than that by buying two-generation-behind cards off EBay.

The older Matrox, ATI, etc. cards have complete driver documentation (AFAIK) or at least are very well supported by completely open Linux / XFree86 drivers.

You can get a Matrox Millenium G400 on EBay right now for under $20. Or an ATI 7500 for under $30.

So, as cool as an open-source graphics chip would be, I think if you can't manufacture a complete AGP card with your new chip for less than $30, and better performance than those parts... don't bother.

Signature · 1 line
Torrey Hoffman <[email protected]>
#10 lore raw
In reply to Timothy Miller
Quoted 7 lines
Quoted 5 lines

Well, the cost of fabricating depends on the device. I was basically thinking of a 68000, an EPROM and a SIMM on a piece of stripboard, some ribbon cable and a DB-25 connector.

Maybe our goals are somewhat different :-)

Very different. What you're describing is a dumb terminal.

Hardly. It's nothing like a dumb terminal whatsoever.

It's a simple framebuffer, possibly with line drawing, and box filling capabilities. Nevertheless, it could be used as a general purpose X display, for spreadsheets, simple to moderate wordprocessing, (I.E. probably not DTP-like applications), status displays for various systems, etc.

So, it does have real world uses.

Quoted 4 lines

What I'm describing is a PC console graphics card that will let someone play Quake III at a reasonable framerate.

Isn't that what most people want?

In the embedded and server markets, I don't see it being a major requirement, actually.

Just because a standard graphics card is going to do all they want and be cheaper to develop, doesn't make it a requirement.

Quoted 2 lines

And the performance disparity between what you're describing and what I'm describing is enormous!

Your arguments seem to be based on the fact that fabricating an ASIC is out of the budget of most individuals, and that no large company would want to develop open source graphics hardware when they can buy $15 graphics cards. That argument is perfectly valid, but it's incomplete.

What _is_ within the budget of most interested individuals are things like general purpose CPUs, generic video sync generation ICs, SIMMs. The parallel port remains far easier to interface to than the PCI bus, and can easily provide enough bandwidth for experimenting with simple 640x480 framebuffer graphics type applications.

So, we can either do something interesting with the above, or sit around discussing how expensive it is to make a graphics card.

At least it provides a way for us to create the first generation of open graphics hardware cheaply, and experiment with various ideas.

Besides, this is just the first stage - once we have the graphics card, we can move on to other things like the 9-track tape drive discussed on LKML a while ago:

http://marc.theaimsgroup.com/?l=linux-kernel&m=105128749415083&w=2

John.

#43 lore raw
In reply to Torrey Hoffman
Quoted 4 lines · Torrey Hoffman

I think I can do better than that by buying two-generation-behind cards off EBay.

[snip]

Agreed. But would those not eventually run out? Certainly, there will always be a supply of used cards that are 2-generations behind, but eventually, we may get to the point where all the used cards have no public documentation.

But perhaps at that point, Linux will dominate and the manufacturers will feel pressured to open their register sets.

#11 lore raw
In reply to John Bradford
Quoted 11 lines · John Bradford
Quoted 7 lines
Quoted 5 lines

Well, the cost of fabricating depends on the device. I was basically thinking of a 68000, an EPROM and a SIMM on a piece of stripboard, some ribbon cable and a DB-25 connector.

Maybe our goals are somewhat different :-)

Very different. What you're describing is a dumb terminal.

Hardly. It's nothing like a dumb terminal whatsoever.

Quoted 7 lines

It's a simple framebuffer, possibly with line drawing, and box filling capabilities. Nevertheless, it could be used as a general purpose X display, for spreadsheets, simple to moderate wordprocessing, (I.E. probably not DTP-like applications), status displays for various systems, etc.

So, it does have real world uses.

But wouldn't it be painfully slow?

Quoted 4 lines
Quoted 4 lines

What I'm describing is a PC console graphics card that will let someone play Quake III at a reasonable framerate.

Quoted 8 lines
Quoted 1 line

Isn't that what most people want?

In the embedded and server markets, I don't see it being a major requirement, actually.

Just because a standard graphics card is going to do all they want and be cheaper to develop, doesn't make it a requirement.

Have you ever used a graphics card in VESA mode? Dragging a window around the screen and watching it repaint can be a very unenjoyable thing to watch. From what you've described, this is the sort of thing you'd get.

Quoted 11 lines
Quoted 3 lines

And the performance disparity between what you're describing and what I'm describing is enormous!

Your arguments seem to be based on the fact that fabricating an ASIC is out of the budget of most individuals, and that no large company would want to develop open source graphics hardware when they can buy $15 graphics cards. That argument is perfectly valid, but it's incomplete.

Quoted 5 lines

What _is_ within the budget of most interested individuals are things like general purpose CPUs, generic video sync generation ICs, SIMMs. The parallel port remains far easier to interface to than the PCI bus, and can easily provide enough bandwidth for experimenting with simple 640x480 framebuffer graphics type applications.

Interfacing with the PCI bus is easy enough in an FPGA. If all you want is a dumb framebuffer, you can fit that logic into a very small, inexpensive Xilinx part. All you need is a DAC and some memory chips, and you're set.

But even PCI can be very slow, particularly for image loads.

Quoted 10 lines

So, we can either do something interesting with the above, or sit around discussing how expensive it is to make a graphics card.

At least it provides a way for us to create the first generation of open graphics hardware cheaply, and experiment with various ideas.

Besides, this is just the first stage - once we have the graphics card, we can move on to other things like the 9-track tape drive discussed on LKML a while ago:

Ok, so, how about this idea:

- Small Xilinx FPGA, 16M of RAM, and a DAC on a board.
- AGP 2X
- Up to 2048x2048 resolution at 8, 16, and 32 bpp.
- Acceleration ONLY for solid fills and bitblts on-screen.

Given that so little is accelerated, there is no point in putting more than the viewable framebuffer on the card, hense the 16 megs. It would probably actually HURT performance to cache pixmaps on the card.

Oh, there's one thing I forgot. It would have to support VGA. There is a VGA core on opencores.org that we could use, but its logic area would probably push up the FPGA cost so that the board was in the $100 range.
Probably more.

<sigh>

#12 lore raw
Quoted 9 lines
Quoted 7 lines

It's a simple framebuffer, possibly with line drawing, and box filling capabilities. Nevertheless, it could be used as a general purpose X display, for spreadsheets, simple to moderate wordprocessing, (I.E. probably not DTP-like applications), status displays for various systems, etc.

So, it does have real world uses.

But wouldn't it be painfully slow?

[snip]

Quoted 4 lines

Have you ever used a graphics card in VESA mode? Dragging a window around the screen and watching it repaint can be a very unenjoyable thing to watch. From what you've described, this is the sort of thing you'd get.

OK, maybe it would be too slow for practical desktop use.

Quoted 13 lines
Quoted 9 lines

So, we can either do something interesting with the above, or sit around discussing how expensive it is to make a graphics card.

At least it provides a way for us to create the first generation of open graphics hardware cheaply, and experiment with various ideas.

Besides, this is just the first stage - once we have the graphics card, we can move on to other things like the 9-track tape drive discussed on LKML a while ago:

Ok, so, how about this idea:

Quoted 8 lines

- Small Xilinx FPGA, 16M of RAM, and a DAC on a board.
- AGP 2X
- Up to 2048x2048 resolution at 8, 16, and 32 bpp.
- Acceleration ONLY for solid fills and bitblts on-screen.

Given that so little is accelerated, there is no point in putting more than the viewable framebuffer on the card, hense the 16 megs. It would probably actually HURT performance to cache pixmaps on the card.

If we put 4 or more on each board, it could be useful for betting shops, stock markets, shop window displays, and other applications where you need to control a dozen or more screens, which basically contain textual information, but where 80x25 text mode just isn't enough. I.E. you might want the odd pie chart or different sized text or something.

Quoted 1 line

Oh, there's one thing I forgot. It would have to support VGA.

Maybe not, the primary market for this, (I.E. what makes it cost effective to produce, and therefore available for developers to use as their primary display), could be users who want to control many displays, and who would have a standard VGA card for the primary monitor. (Yeah, it would be kind of ironic if 99% of our amasing new graphics cards ended up in mahines with another card as the primary display, but then again, if it makes the open hardware available for developers to experiment with at a reasonable cost, it would be worth doing).

So, what about a PCI card with four or eight 16MB framebuffers, and the basic acceleration and other specs you described above. Is that at least slightly feasible, do you think?

John.

#13 lore raw
Quoted 7 lines · John Bradford

If we put 4 or more on each board, it could be useful for betting shops, stock markets, shop window displays, and other applications where you need to control a dozen or more screens, which basically contain textual information, but where 80x25 text mode just isn't enough. I.E. you might want the odd pie chart or different sized text or something.

The market for secondary heads is too small. You can get an ATI Mach 64 PCI card for pennies and add it as a second head for what you're describing.

For an open-source graphics card to be marketable, it would have to be attractive as a primary head used in Linux workstations and servers, and it would have to be so in a PC market.

Quoted 5 lines
Quoted 3 lines

Oh, there's one thing I forgot. It would have to support VGA.

Quoted 9 lines

Maybe not, the primary market for this, (I.E. what makes it cost effective to produce, and therefore available for developers to use as their primary display), could be users who want to control many displays, and who would have a standard VGA card for the primary monitor. (Yeah, it would be kind of ironic if 99% of our amasing new graphics cards ended up in mahines with another card as the primary display, but then again, if it makes the open hardware available for developers to experiment with at a reasonable cost, it would be worth doing).

The irony is too much. Seriously.

Quoted 4 lines

So, what about a PCI card with four or eight 16MB framebuffers, and the basic acceleration and other specs you described above. Is that at least slightly feasible, do you think?

Adding extra heads is relatively easy, and you can keep the memory unified and do it all in one chip.

#14 lore raw
In reply to Timothy Miller
Quoted 4 lines · Timothy Miller

John Bradford wrote:

[...]

Quoted 6 lines
Quoted 6 lines
Quoted 3 lines

What I'm describing is a PC console graphics card that will let someone play Quake III at a reasonable framerate.

Isn't that what most people want?

Quoted 9 lines
Quoted 5 lines

In the embedded and server markets, I don't see it being a major requirement, actually.

Just because a standard graphics card is going to do all they want and be cheaper to develop, doesn't make it a requirement.

Have you ever used a graphics card in VESA mode? Dragging a window around the screen and watching it repaint can be a very unenjoyable thing to watch. From what you've described, this is the sort of thing you'd get.

I run X on an unaccelerated framebuffer (1280x1024 16bit color) every day.
I don't even _notice_ a difference from accelerated X for a number of uses, such as word processing, watching movies with mplayer, web browsing and programming. Dragging a window around is fine!
Simple opengl games like "frozen bubble" with software rendering are fine too,
on a 333MHz dual celeron.
The only stuff that don't work well is 3D-intensive stuff like quake and tuxracer.

(The unaccelerated xserver is running on the second head of a matrox G550. The primary head uses acceleration, but is often in use by another user.)

So a good 2D card is trivial - a video signal generator and memory on an AGP bus. Let the host processor do software rendering. Cheap, and I believe this is the sort of thing embedded uses might go for when they want to display mostly static stuff. (Web-based info kiosk and similiar).
Add a BIOS rom and you can even see what happens during boot on a pc.

The next step up is 2D acceleration, which is easy enough by sticking a generic microprocessor there. Maybe an inexpensive celeron/duron.

Then there's 3D, and enough of it to play quake. The first quakes ran
fine with software rendering and processors that were slow by today's
standards. Todays cheap processors are faster - I wonder if putting 2-4 of them on the card might be enough. They'd be able to access the memory directly, not limited to slow AGP/PCI speeds. And they'd be able to divide the work
between them, rendering separate parts of the screen.
[...]

Quoted 3 lines

- Small Xilinx FPGA, 16M of RAM, and a DAC on a board.
- AGP 2X
- Up to 2048x2048 resolution at 8, 16, and 32 bpp.

Why bother with 8-bit?

Quoted 6 lines

- Acceleration ONLY for solid fills and bitblts on-screen.

Given that so little is accelerated, there is no point in putting more than the viewable framebuffer on the card, hense the 16 megs. It would probably actually HURT performance to cache pixmaps on the card.

Oh, there's one thing I forgot. It would have to support VGA. There is

Why VGA? When you have a _driver_ , you don't need compatibility at all. (Just like soundcards - they don't need soundblaster compatibility for anything) The pc don't need vga - it can boot using the card's bios
Linux don't need vga - it will use the provided driver.
Apps don't need vga, they don't do that sort of thing anyway. They use the tty/X11/SDL/opengl.

Quoted 2 lines

a VGA core on opencores.org that we could use, but its logic area would probably push up the FPGA cost so that the board was in the $100 range.
Probably more.

Another reason to drop VGA then - money.

Helge Hafting

#35 lore raw
In reply to John Bradford
Quoted 10 lines · John Bradford
Quoted 7 lines
Quoted 5 lines

Well, the cost of fabricating depends on the device. I was basically thinking of a 68000, an EPROM and a SIMM on a piece of stripboard, some ribbon cable and a DB-25 connector.

Maybe our goals are somewhat different :-)

Very different. What you're describing is a dumb terminal.

Hardly. It's nothing like a dumb terminal whatsoever.

Quoted 7 lines

It's a simple framebuffer, possibly with line drawing, and box filling capabilities. Nevertheless, it could be used as a general purpose X display, for spreadsheets, simple to moderate wordprocessing, (I.E. probably not DTP-like applications), status displays for various systems, etc.

So, it does have real world uses.

Yes - but you want it:

1. to use the AGP to gain access to multiple offscreen pages
2. a DMA controler to copy the data
3. A simple emulation (either 8bit cpu based or better) of VGA/SVGA
4. room in the design for future processors.

Really - future processors:
1. including multiple vector multiply processors
2. general purpose CPU for control
3. LOTS of memory.

What you REALLY need to do (long term) is to move the entire X server into a graphics board (including Mesa/OpenGL/... but minus the network code, authentication, and resource database...).

It is my understanding that a LOT of the effort at speed is lost by using a single threaded process to handle the graphics. With a multiple cpu (not necessarily SMP mind you) performing the graphics transformations, you have a single rendering output step (another case for multiple cpus - 1 cpu: entire pixel render, 2: each takes 1/2 display, 4 - 1/4 display...). And with multiple dual ported graphics memory (port to pixel rendering cpu, port to frame buffer) you end up wit a very fast graphics display.

Limiting factor: it may be bigger than a single slot.

It would likely resemble the old SGI type of rendering engine, which used multiple boards, multiple staging memory, and multiported display.

BUT: it would be modular. Pay a little and you only get a frame buffer.

Add a general CPU - you get a basic X server (with slow 3D, but likely faster than currently done by the host processor)-- and pay more. Add a geometry engine (ie a processor/memory for Mesa) you get faster 3D operations...

Add multiple engines (each takes part of the display) you get speed...

It would likely require one AGP, but two PCI slots; and like the SGI engines, an internal connection between the two boards.

This would also allow the project to have price levels - a $20 AGP frame buffer wouldn't be bad at all (and not all that slow either...)

Add $40 for a general CPU... with the benifit of offloading the major X functions... and still have the ability to use the AGP. (BTY - the AGP is bi-directional... you should be able to copy images from the framebuffer)

Add $40 (might have to trade in the existing general CPU... so it could actually be ~$80) and you should get options for multiple geometry processors... at $10/20 each?

Note - the costs shown for the last upgrade is very likely wrong.

Quoted 4 lines
Quoted 4 lines

What I'm describing is a PC console graphics card that will let someone play Quake III at a reasonable framerate.

Isn't that what most people want?

Something like the above should do.

Quoted 1 line

[snip]

#15 lore raw
In reply to Helge Hafting

NOTE: My guilt meter for being so far off topic is starting to peg. Let's either find a better open forum to discuss this or just take it off list.

The only reason I haven't taken it completely off list so far is because SOME aspects of this discussion may be relevant to Free Software in general. It is the Linux mentality that spurred this idea and Linux users who would be the target market. But now we're just arguing trivialities.

So, the obligatory on-topic comment is simple: Any Linux-targeted graphics chip has to be quite sophistocated and cost-competitive to some extent in order for the idea to have any merit whatsoever. Core Linux users on this list are the most qualified to dictate what they would use it for.

Unfortunately, the comments I have gotten so far lead me to solutions that already exist.

Quoted 2 lines · Helge Hafting
Quoted 1 line

I run X on an unaccelerated framebuffer (1280x1024 16bit color) every day.

This is you. How many other people would be happy with this?

Quoted 2 lines

So a good 2D card is trivial - a video signal generator and memory on an AGP bus.

I wouldn't call it 'good', but you're essentially correct. Mind you, you can get the latest S3 chip for $30 or less. That's well documented, a lot faster than a dumb framebuffer, and has VGA built in.

Once you get below a certain level of performance in this open-source design, it's no longer worth doing because there are so many low-end alternatives that blow it away.

Quoted 3 lines

Let the host processor do software rendering. Cheap, and I believe this is the sort of thing embedded uses might go for when they want to display mostly static stuff. (Web-based info kiosk and similiar).

Not cheap. You can get cheaper with what's already out there!

Let me put it this way: To make it cheaper than what's out there, someone would have to do a run of _millions_ of some minimalist chip that sold for like $1. You could have a graphics card for $5. THAT would make it worth it. But we would never get the volumes necessary for that!

Quoted 1 line

Add a BIOS rom and you can even see what happens during boot on a pc.

A BIOS ROM doesn't help you if you don't have a VGA core, and a VGA core is not a trivial piece of logic.

I like Macs, Suns, and other UNIX workstations because they don't rely on this antiquated piece of logic to act as a console. The chip I designed doesn't do VGA, but that doesn't stop it from working nicely as a console in a Sun.

Quoted 3 lines

The next step up is 2D acceleration, which is easy enough by sticking a generic microprocessor there. Maybe an inexpensive celeron/duron.

<sigh> Think about the logic area required for that. For BASIC 2D acceleration, the amount of logic required for elementary operations is miniscule compared to the logic required for even the simplest of CPU cores. And the dedicated logic would be faster!

Quoted 6 lines

Then there's 3D, and enough of it to play quake. The first quakes ran
fine with software rendering and processors that were slow by today's
standards. Todays cheap processors are faster - I wonder if putting 2-4 of them on the card might be enough. They'd be able to access the memory directly, not limited to slow AGP/PCI speeds. And they'd be able to divide the work
between them, rendering separate parts of the screen.

My knowledge of 3D graphics is limited to linear algebra and mostly pure theory, but I do have SOME clue as to what existing 3D engines are like.
The issue of general purpose CPU's versus GPU's has been discussed
on the web at nauseum, and for what they do, modern GPU's are an order of magnitude faster than CPU's at what they do. And they're less expensive!

Quoted 2 lines

Why bother with 8-bit?

There are reasons. We can go into them off-list.

Quoted 3 lines

Why VGA? When you have a _driver_ , you don't need compatibility at all.

BOOT CONSOLE. You cannot get a boot console on a PC without a VGA core.
Once the kernel takes over, you're right, but until then...

Quoted 3 lines

Another reason to drop VGA then - money.

As soon as PC BIOS's don't require it, we can drop it.

#16 lore raw
Quoted 3 lines · Timothy Miller
Quoted 1 line

Another reason to drop VGA then - money.

As soon as PC BIOS's don't require it, we can drop it.

No PC BIOS recognizes a VGA. The PC/AT firmware uses int 0x10 to
communicate with the console and as long as there is a handler there, console output works. Most systems will actually run without a handler, too, but they'll usually complain to the speaker. The handler is provided by the ROM firmware of the primary graphics adapter.

Old PC/AT firmware actually did recognize a few display adapters, namely
the CGA and the MDA which had no own firmware. These days support for these option is often absent, even though the setup program may provide an option to select between CGA40/CGA80/MDA/none (the latter being equivalent to an option such as an EGA or a VGA, providing its own firmware).

Signature · 3 lines
+  Maciej W. Rozycki, Technical University of Gdansk, Poland   +
+--------------------------------------------------------------+
+        e-mail: [email protected], PGP key available        +
#25 lore raw
In reply to Timothy Miller
Quoted 5 lines · Timothy Miller
Quoted 2 lines

Why VGA? When you have a _driver_ , you don't need compatibility at all.

BOOT CONSOLE. You cannot get a boot console on a PC without a VGA core. Once the kernel takes over, you're right, but until then...

Why PC?

Quoted 3 lines
Quoted 1 line

Another reason to drop VGA then - money.

As soon as PC BIOS's don't require it, we can drop it.

Linuxbios?

Signature · 2 lines
Måns Rullgård
[email protected]
#17 lore raw
In reply to Maciej W. Rozycki
Quoted 14 lines · Maciej W. Rozycki
Quoted 3 lines · Timothy Miller
Quoted 1 line

Another reason to drop VGA then - money.

As soon as PC BIOS's don't require it, we can drop it.

No PC BIOS recognizes a VGA. The PC/AT firmware uses int 0x10 to
communicate with the console and as long as there is a handler there, console output works. Most systems will actually run without a handler, too, but they'll usually complain to the speaker. The handler is provided by the ROM firmware of the primary graphics adapter.

Quoted 6 lines

Old PC/AT firmware actually did recognize a few display adapters, namely
the CGA and the MDA which had no own firmware. These days support for these option is often absent, even though the setup program may provide an option to select between CGA40/CGA80/MDA/none (the latter being equivalent to an option such as an EGA or a VGA, providing its own firmware).

You're not entirely correct here. I attempted to write a VGA BIOS for a card which did not have hardware support for 80x25 text.

I first tried intercepting int 0x10. I quickly discovered that most DOS programs bypass int 0x10 and write directly to the display memory. As a result, very little of what should have displayed actually did.

Next, I tried hanging off this timer interrupt. I had two copies of the text display, "now" and "what it was before". I would compare the characters and render any differences. This worked quite well for DOS, but the instant ANY OS switched to protected mode, they took over the interrupt and all console messages stopped. Actually, the same was true for int 0x10.

Even just the DOS shell command-line tends to bypass int 0x10 and write directly to display memory.

Furthermore, 640x480x16 simply won't happen at all without direct hardware support. Some things rely on that (or mode X or whatever) for initial splash screens.

In the PC world, too many assumptions are made about the hardware for any kind of software emulation to work.

The suggestion that a general-purpose CPU on the graphics card could be used to emulate it is correct, but the logic area of the general-purpose CPU is greater than that of the dedicated VGA hardware. Furthermore, you can't just "stick a Z80 onto the board", because multi-chip solutions up the board cost too much.

#26 lore raw
In reply to Måns Rullgård
Quoted 2 lines · Måns Rullgård

Why PC?

Uh, the most common platform on which Linux is run?

Important point: unless the sales volumes of this card exceed a certain point, the project is pointless.

The volumes in, say, the Sun market are so low that you'd can't get decent a console graphics card for under $300.

But the Sun/Solaris market doesn't care about open source. So why bother?

Quoted 9 lines
Quoted 5 lines
Quoted 3 lines

Another reason to drop VGA then - money.

As soon as PC BIOS's don't require it, we can drop it.

Linuxbios?

Good point. As soon as every PC is running it, then that'll be helpful.

#18 lore raw
In reply to Timothy Miller
Quoted 6 lines · Timothy Miller

You're not entirely correct here. I attempted to write a VGA BIOS for a card which did not have hardware support for 80x25 text.

I first tried intercepting int 0x10. I quickly discovered that most DOS programs bypass int 0x10 and write directly to the display memory. As a result, very little of what should have displayed actually did.

Of course, but DOS is not BIOS and the assumption is we want to use the
adapter as a boot console and with Linux. The former is handled with appropriate firmware and the latter with a driver.

Actually I had an opportunity to use a few PC/AT headless systems (no
video adapter at all, although one could be placed in a PCI slot) with an option called "serial console redirection" in the firmware. Their BIOS setup program proved to work just fine over a serial line (unfortunately a VT100 terminal was assumed, so I had to type e.g. ^[OP to "press" <F1>, but it worked) as well any console output, including LILO (which had to be taught to use the regular console instead of accessing the serial port for I/O directly).

Signature · 3 lines
+  Maciej W. Rozycki, Technical University of Gdansk, Poland   +
+--------------------------------------------------------------+
+        e-mail: [email protected], PGP key available        +
#19 lore raw
Quoted 5 lines · Maciej W. Rozycki

Of course, but DOS is not BIOS and the assumption is we want to use the
adapter as a boot console and with Linux. The former is handled with appropriate firmware and the latter with a driver.

Perhaps someone can tell us what the Linux kernel does before the console driver gets loaded.

If the console driver is a module, then all kernel init messages that appear before the module is loaded have nowhere to go.

#27 lore raw
In reply to Timothy Miller
Quoted 2 lines

A BIOS ROM doesn't help you if you don't have a VGA core, and a VGA core is not a trivial piece of logic.

We don't need a VGA core.

The primary markets for this card are going to be developers who don't care about using a standard VGA card for pre-kernel-loaded stuff, the embedded and server markets, where they can simply write a new system BIOS that emulates 80x25 text mode on the framebuffer, and the multi-head market, where the kernel or X will be responsible for the extra heads anyway.

Once the kernel or X has taken over the framebuffer, there is certainly no need for a VGA core.

As far as I am concerned, the first version of this card is going to be more or less an expensive proof-of-concept thing. It _will_ cost more than a brand new off-the-shelf VGA card, and it _will_ cost more than a second-hand VGA card with documented registers. That doesn't mean it isn't worth doing, though.

John.

#20 lore raw
In reply to Timothy Miller
Quoted 6 lines · Timothy Miller
Quoted 3 lines

Of course, but DOS is not BIOS and the assumption is we want to use the
adapter as a boot console and with Linux. The former is handled with appropriate firmware and the latter with a driver.

Perhaps someone can tell us what the Linux kernel does before the console driver gets loaded.

The kernel does lot of activities, but if you mean console output, then
it doesn't start before the console driver is initialized, unless a so-called initial console with a suitable driver is present, which may be firmware-driven (so the driver may be a trivial redirector to appropriate firmware callbacks).

Quoted 2 lines

If the console driver is a module, then all kernel init messages that appear before the module is loaded have nowhere to go.

If there's no better console available, e.g. because there's no suitable
hardware present in the system or no drivers have been loaded, then the dummy console is used -> drivers/video/console/dummycon.c.

And if you worry of the messages being lost, then you can always retrieve
them from the kernel log buffer -- use `dmesg' for example. ;-)

Signature · 3 lines
+  Maciej W. Rozycki, Technical University of Gdansk, Poland   +
+--------------------------------------------------------------+
+        e-mail: [email protected], PGP key available        +
#21 lore raw
In reply to Timothy Miller
Quoted 10 lines · Timothy Miller
Quoted 7 lines · Maciej W. Rozycki
Quoted 3 lines · Timothy Miller
Quoted 1 line

Another reason to drop VGA then - money.

As soon as PC BIOS's don't require it, we can drop it.

PC bioses don't need VGA and never did!
They use the int 0x10 handler provided by the graphichs card bios. When you make the card - you get to write that bios. No problem!

I once used a dec rainbow - a pc with an ibm-incompatible screen. The display memory was organized as a linked list of lines instead of an array of characters. It came with its own special version of msdos 2.11. Few ordinary dos programs would run on it, beause most tried to access the "standard" 80x25 array instead of using msdos for io. Those who did the right thing worked, though.

(An odd machine in other ways too - it had a z80 controlling the floppies and a 8088 controlling the screen and harddisk. Early 80's asymmetric multiprocessor :-)

Quoted 14 lines
Quoted 7 lines

No PC BIOS recognizes a VGA. The PC/AT firmware uses int 0x10 to communicate with the console and as long as there is a handler there, console output works. Most systems will actually run without a handler, too, but they'll usually complain to the speaker. The handler is provided by the ROM firmware of the primary graphics adapter.

Quoted 6 lines

Old PC/AT firmware actually did recognize a few display adapters, namely the CGA and the MDA which had no own firmware. These days support for these option is often absent, even though the setup program may provide an option to select between CGA40/CGA80/MDA/none (the latter being equivalent to an option such as an EGA or a VGA, providing its own firmware).

Quoted 7 lines

You're not entirely correct here. I attempted to write a VGA BIOS for a card which did not have hardware support for 80x25 text.

I first tried intercepting int 0x10. I quickly discovered that most DOS programs bypass int 0x10 and write directly to the display memory. As a result, very little of what should have displayed actually did.

Sure, but we're not interested in "most dos programs", are we? The pc bios bootup will work, it uses int 0x10.
lilo output will work.
linux kernel console output will work
X will work, either with the generic framebuffer driver, or with a proper driver written for the open hardware.

Quoted 7 lines

Next, I tried hanging off this timer interrupt. I had two copies of the text display, "now" and "what it was before". I would compare the characters and render any differences. This worked quite well for DOS, but the instant ANY OS switched to protected mode, they took over the interrupt and all console messages stopped. Actually, the same was true for int 0x10.

If you want DOS application compatibility or windows compatibility then you might need VGA. But you started out talking about open hardware for linux - and then you really don't need vga at all. Not even an initial 80x25 character array. A kernel without vga support (but some other console like fbcon) works fine.

Quoted 3 lines

Even just the DOS shell command-line tends to bypass int 0x10 and write directly to display memory.

Depends on what version of dos, but you can always get freedos for which source code is available - if dos matters to you. It is something I only ever use for flashing bios upgrades.

Quoted 4 lines

Furthermore, 640x480x16 simply won't happen at all without direct hardware support. Some things rely on that (or mode X or whatever) for initial splash screens.

Not in linux. Of course you can reserve the legacy vga memory region and just live with the loss of splash screens in dos.

Quoted 3 lines

In the PC world, too many assumptions are made about the hardware for any kind of software emulation to work.

Not in the pc world. The pc is only hardware.
The problem is the microsoft os world, but supporting that _isn't necessary_ when you don't plan on high volumes. I guess you could get windows going - it uses proper display drivers these days even if the installer doesn't. Install with vga card, swap driver, shutdown, swap cards, power-on or some such.

Quoted 5 lines

The suggestion that a general-purpose CPU on the graphics card could be used to emulate it is correct, but the logic area of the general-purpose CPU is greater than that of the dedicated VGA hardware. Furthermore, you can't just "stick a Z80 onto the board", because multi-chip solutions up the board cost too much.

Thanks for the information, seems I don't know enough about board manufacturing.

Helge Hafting

#28 lore raw
In reply to Timothy Miller
Quoted 8 lines · Timothy Miller
Quoted 2 lines · Helge Hafting
Quoted 1 line

I run X on an unaccelerated framebuffer (1280x1024 16bit color) every day.

This is you. How many other people would be happy with this?

Depends on what they do, of course. It _is_ fine for 2D office work, that surprised me quite a bit. I can even use mplayer and watch videos on this.

Someone interested in quake will of course not be satisfied though.

Quoted 5 lines
Quoted 4 lines

So a good 2D card is trivial - a video signal generator and memory on an AGP bus.

Quoted 8 lines

I wouldn't call it 'good', but you're essentially correct. Mind you, you can get the latest S3 chip for $30 or less. That's well documented, a lot faster than a dumb framebuffer, and has VGA built in.

Once you get below a certain level of performance in this open-source design, it's no longer worth doing because there are so many low-end alternatives that blow it away.

Good point. I had the impression you weren't going for the best performance anyway.

Quoted 8 lines

A BIOS ROM doesn't help you if you don't have a VGA core, and a VGA core is not a trivial piece of logic.

I like Macs, Suns, and other UNIX workstations because they don't rely on this antiquated piece of logic to act as a console. The chip I designed doesn't do VGA, but that doesn't stop it from working nicely as a console in a Sun.

The pc doesn't need VGA anymore than a Sun does. A pc without any vga runs the same stuff as the sun - various unixes and no legacy dos stuff. Don't confuse "pc" with "msdos/windows". If the sun's software selection satisfy you, then you're fine on a pc/x86 without vga (or other legacy hw) too.

[...]

Quoted 7 lines

My knowledge of 3D graphics is limited to linear algebra and mostly pure theory, but I do have SOME clue as to what existing 3D engines are like.
The issue of general purpose CPU's versus GPU's has been discussed
on the web at nauseum, and for what they do, modern GPU's are an order of magnitude faster than CPU's at what they do. And they're less expensive!

Seems the solution is to get one of those general purpose GPU's instead, and build the card around that. Not entirely open hw, but _fully documented_ with no hidden surprises.

Helge Hafting

#22 lore raw
In reply to Helge Hafting

Alright then, how about this: Assuming opencores has a PCI interface and a DDR memory controller, I could write a CRT controller. We can put that into an FPGA and see what happens.

Quoted 100 lines · Helge Hafting
Quoted 11 lines · Timothy Miller
Quoted 8 lines · Maciej W. Rozycki
Quoted 3 lines · Timothy Miller
Quoted 1 line

Another reason to drop VGA then - money.

As soon as PC BIOS's don't require it, we can drop it.

PC bioses don't need VGA and never did!
They use the int 0x10 handler provided by the graphichs card bios. When you make the card - you get to write that bios. No problem!

I once used a dec rainbow - a pc with an ibm-incompatible screen. The display memory was organized as a linked list of lines instead of an array of characters. It came with its own special version of msdos 2.11. Few ordinary dos programs would run on it, beause most tried to access the "standard" 80x25 array instead of using msdos for io. Those who did the right thing worked, though.

(An odd machine in other ways too - it had a z80 controlling the floppies and a 8088 controlling the screen and harddisk. Early 80's asymmetric multiprocessor :-)

Quoted 13 lines
Quoted 6 lines

No PC BIOS recognizes a VGA. The PC/AT firmware uses int 0x10 to communicate with the console and as long as there is a handler there, console output works. Most systems will actually run without a handler, too, but they'll usually complain to the speaker. The handler is provided by the ROM firmware of the primary graphics adapter.

Quoted 6 lines

Old PC/AT firmware actually did recognize a few display adapters, namely the CGA and the MDA which had no own firmware. These days support for these option is often absent, even though the setup program may provide an option to select between CGA40/CGA80/MDA/none (the latter being equivalent to an option such as an EGA or a VGA, providing its own firmware).

Quoted 4 lines

You're not entirely correct here. I attempted to write a VGA BIOS for a card which did not have hardware support for 80x25 text.

I first tried intercepting int 0x10. I quickly discovered that most DOS programs bypass int 0x10 and write directly to the display memory. As a result, very little of what should have displayed actually did.

Sure, but we're not interested in "most dos programs", are we? The pc bios bootup will work, it uses int 0x10.
lilo output will work.
linux kernel console output will work
X will work, either with the generic framebuffer driver, or with a proper driver written for the open hardware.

Quoted 2 lines

Next, I tried hanging off this timer interrupt. I had two copies of the text display, "now" and "what it was before". I would compare the characters and render any differences. This worked quite well for DOS, but the instant ANY OS switched to protected mode, they took over the interrupt and all console messages stopped. Actually, the same was true for int 0x10.

If you want DOS application compatibility or windows compatibility then you might need VGA. But you started out talking about open hardware for linux - and then you really don't need vga at all. Not even an initial 80x25 character array. A kernel without vga support (but some other console like fbcon) works fine.

Quoted 2 lines

Even just the DOS shell command-line tends to bypass int 0x10 and write directly to display memory.

Depends on what version of dos, but you can always get freedos for which source code is available - if dos matters to you. It is something I only ever use for flashing bios upgrades.

Quoted 2 lines

Furthermore, 640x480x16 simply won't happen at all without direct hardware support. Some things rely on that (or mode X or whatever) for initial splash screens.

Not in linux. Of course you can reserve the legacy vga memory region and just live with the loss of splash screens in dos.

Quoted 2 lines

In the PC world, too many assumptions are made about the hardware for any kind of software emulation to work.

Not in the pc world. The pc is only hardware. The problem is the microsoft os world, but supporting that _isn't necessary_ when you don't plan on high volumes. I guess you
could get windows going - it uses proper display drivers these days
even if the installer doesn't. Install with vga card, swap driver,
shutdown, swap cards, power-on or some such.

Quoted 1 line

The suggestion that a general-purpose CPU on the graphics card could be used to emulate it is correct, but the logic area of the general-purpose CPU is greater than that of the dedicated VGA hardware. Furthermore, you can't just "stick a Z80 onto the board", because multi-chip solutions up the board cost too much.

Thanks for the information, seems I don't know enough about board manufacturing.

Helge Hafting

#41 lore raw
In reply to Frank Gevaerts
Quoted 4 lines · Frank Gevaerts

Some people agree:

http://www.opencores.org/projects/manticore/
http://www.opencores.org/projects/vga_lcd/

Yes, I was just about to mention manticore;). May I then mention the gpl cpu called leon; all the blueprints of the cpu is released under the gpl. Please look at esa's page (european space agency). Leon is a ieee 1754 conformant cpu which means it's compatible with sparc.

Signature · 1 line
b0ef
#23 lore raw
In reply to Timothy Miller

Quote from Timothy Miller <[email protected]>:

Quoted 3 lines

Alright then, how about this: Assuming opencores has a PCI interface and a DDR memory controller, I could write a CRT controller. We can put that into an FPGA and see what happens.

Well, there is a PCI to wishbone bridge:

http://www.opencores.org/projects/pci/

and a DDR memory controller:

http://www.opencores.org/projects/ddr_sdr/

but do we really need DDR RAM? For the small amount of RAM on the card, (8 or 16 MB at the most), surely the cost of standard static ram ICs wouldn't be too off-putting, and it would presumably simplify the design slightly.

(It's you that's got to build it, though, so it's your call :-) ).

John.

#45 lore raw
In reply to Timothy Miller

Hi,
Just one more crazy idea ...
Have you thought about cooperating with some kind of commercial company? ie. SiS tried/is trying to create a 3D GPU without much success ... but they might be able to help manufacturing GPUs/Cards even in bigger series and as long as they dont have to invest too much in development and get some income i think they can do it. And as far as i know a lot of windows users would appreciate opensource graphics cores because of the advantages of opensource drivers. I would also say that there would be enough people interested in writing drivers for windows so your card could be generally very well marketable. I know this is not windows mailng list, but speaking of marketability of the core windows has the greates market share and having it working in windows too would mean that you are getting at production quotes big enough for reaching low prices. So why not "abuse" windows to get us a good GPU?

Tomas Zvala

#24 lore raw
In reply to Timothy Miller
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"Timothy" == Timothy Miller <[email protected]> writes:

Timothy> Alright then, how about this: Assuming opencores has a Timothy> PCI interface and a DDR memory controller, I could write Timothy> a CRT controller. We can put that into an FPGA and see Timothy> what happens.

I suggest propose the following: spend the next few months designing, writing documents, and starting on the RTL. In that time PCI Express (PCI over high-speed serial) motherboards and fairly cheap next-generation Xilinx Virtex FPGAs with integrated SERDES and a free PCI Express core from Xilinx should be available.

PCI Express 8X gives you 16 Gb/sec of bandwidth in both directions (32 Gb/sec total) which should be enough to make UMA (ie no memory attached to the FPGA) palatable. So your proto board is looking like it has just power supplies, FPGA, and misc. video junk (ie DAC or digital flat panel support), so it should be reasonably cheap to design and fab. If you really want to, you could look at putting DRAM (RLDRAM?) down on the board but I don't think it's worth the cost and complexity.

(By the way, the Virtex FPGAs also have embedded PowerPC 405 cores that can run at ~400 MHz, which means a a lot of stuff -- exception paths, etc -- can be done in firmware if you want)

It's still in the thousands of dollars for this proto stage, but the Virtex should be fast enough to do something pretty interesting. At that point either the project takes off and you can look at doing a full custom chip (I don't think it's worth doing any "rapid chip" of low-NRE design, since your unit costs will be too high for a mass-volume graphics chip), or the project dies.

- Roland

#36 lore raw
In reply to Timothy Miller
Quoted 7 lines · Timothy Miller
Quoted 6 lines · John Bradford
Quoted 4 lines

The real question we have to ask ourselves is, what would be the market demand for a graphics card that is 3 generations behind the state of the art and over-priced, the only advantage being that it's a 100% open architecture?

Quoted 7 lines
Quoted 2 lines

Err, well there are always the server and embedded markets, if the device was cheap enough.

Ah, but it won't be. Low-volume ASICs are expensive. The chip itself would probably be around $150, not counting $100k NRE. Then you have to pay for the board, make up for the NRE, and make some profit to make it worth while. How much are YOU willing to pay?

Then why are companies doing ASICs with an OpenRISC core?

Quoted 3 lines

Whatever we design is going to be EXPENSIVE. So, regardless of the fact that an ATI All-in-Wonder Radeon 9000 is over-powered for job you describe, that board will be cheaper than what we could produce.

And it's probably overpowered when speaking about power consumption, too...

Now think of a SoC with an OpenRISC core and an integrated graphics controller...

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]

In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds

#29 lore raw
In reply to Timothy Miller
Quoted 6 lines · Timothy Miller

Ok, so, how about this idea:

- Small Xilinx FPGA, 16M of RAM, and a DAC on a board.
- AGP 2X
- Up to 2048x2048 resolution at 8, 16, and 32 bpp.
- Acceleration ONLY for solid fills and bitblts on-screen.

Sounds OK to me.

Quoted 9 lines

Given that so little is accelerated, there is no point in putting more than the viewable framebuffer on the card, hense the 16 megs. It would probably actually HURT performance to cache pixmaps on the card.

Oh, there's one thing I forgot. It would have to support VGA. There is a VGA core on opencores.org that we could use, but its logic area would probably push up the FPGA cost so that the board was in the $100 range.
Probably more.

Why support legacy VGA? It makes things more complex and expensive, and doesn't give us much, especially for a SoC.

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]

In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds

#30 lore raw

Quote from Geert Uytterhoeven <[email protected]>:

Quoted 15 lines · Timothy Miller
Quoted 6 lines

Ok, so, how about this idea:

- Small Xilinx FPGA, 16M of RAM, and a DAC on a board.
- AGP 2X
- Up to 2048x2048 resolution at 8, 16, and 32 bpp.
- Acceleration ONLY for solid fills and bitblts on-screen.

Sounds OK to me.

Quoted 5 lines

Given that so little is accelerated, there is no point in putting more than the viewable framebuffer on the card, hense the 16 megs. It would probably actually HURT performance to cache pixmaps on the card.

Quoted 7 lines
Quoted 4 lines

Oh, there's one thing I forgot. It would have to support VGA. There is a VGA core on opencores.org that we could use, but its logic area would probably push up the FPGA cost so that the board was in the $100 range.
Probably more.

Why support legacy VGA? It makes things more complex and expensive, and doesn't give us much, especially for a SoC.

It makes the card usable in a standard machine before the kernel has booted. I don't care about that at all from a technical viewpoint, as there are better ways to deal with it, (re-write the system BIOS to use the framebuffer for the BIOS configuration, for example), but I can see that there may be a marketing reason for including VGA - more people are likely to buy the card. Whether this is significant depends on the market we try to sell it in, and whether cost per unit would be increased by the extra complexity, or decreased by the higher volume of production possible if more units are going to be sold.

A cheap cludge would be an optional second GPU on the card just to do the required VGA modes, with an analogue video pass-through. That would make the VGA cards more expensive than a single GPU which incorporated VGA, but add almost nothing in cost or complexity terms to the non-VGA cards.

John.

#31 lore raw
Quoted 5 lines · John Bradford

A cheap cludge would be an optional second GPU on the card just to do the required VGA modes, with an analogue video pass-through. That would make the VGA cards more expensive than a single GPU which incorporated VGA, but add almost nothing in cost or complexity terms to the non-VGA cards.

The DEC TGA series cards did something like that, using a Cirrus Logic chip for the VGA support.

Signature · 2 lines
Måns Rullgård
[email protected]
#44 lore raw
In reply to Timothy Miller

Hi!

Quoted 18 lines

The real question we have to ask ourselves is, what would be the market demand for a graphics card that is 3 generations behind the state of the art and over-priced, the only advantage being that it's a 100% open architecture?

I don't have _100k to have it fabricated, so we have to goad some company into doing it for us, and given the volumes, they'll have to charge way more than it's worth if you compare its capabilities against ATI et al.

I've got some great ideas for how to do this chip, but they're frankly nothing revolutionary. The obvious test bed is an FPGA. That imposes serious limitations on what kind of logic utilization and performance we can get. The ASIC version can be clocked faster, but we dare not put in untested logic. (And we can't afford the tools necessary to do the proper simulation.)

Quoted 3 lines

So, the big question: How many units a year would be sold for an underpowered, over-priced graphics card that just happens to be 100% open and 100% supported?

It might be very usefull for embeded folks from opencores...
Pavel

Signature · 1 line
64 bytes from 195.113.31.123: icmp_seq=28 ttl=51 time=448769.1 ms         
#32 lore raw
In reply to Geert Uytterhoeven
Quoted 8 lines
Quoted 6 lines

Ok, so, how about this idea:

- Small Xilinx FPGA, 16M of RAM, and a DAC on a board.
- AGP 2X
- Up to 2048x2048 resolution at 8, 16, and 32 bpp.
- Acceleration ONLY for solid fills and bitblts on-screen.

Sounds OK to me.

Disregarding what i said 5 min ago on modular RAM (sorry ... delusions of grandure).

Would it have to be AGP? afteall the data volumes pumping through that would probably be not that high, and with the PCI express being backwards compatible, this would evade a dying bus.

Signature · 11 lines
with kind regards,
  Christian Unger

-<< Contact Details >>- < > - < > - < > - < > - < > -<< Naiv Status >>- < > -

  Alt. Email:  [email protected]      |   Stable:    0.2.3 r3
  ICQ:         204184156                         |   Latest:    0.3.0
  Mobile:      0402 268904                       |   Current:   0.3.1
  Web:         http://naiv.sourceforge.net       |   Focus:     File Handling

  "You don't need eyes to see ... You need vision" (Faithless - Revenrence)
#37 lore raw
In reply to Geert Uytterhoeven
Quoted 13 lines · Geert Uytterhoeven
Quoted 11 lines · Timothy Miller
Quoted 8 lines · John Bradford
Quoted 4 lines

The real question we have to ask ourselves is, what would be the market demand for a graphics card that is 3 generations behind the state of the art and over-priced, the only advantage being that it's a 100% open architecture?

Err, well there are always the server and embedded markets, if the device was cheap enough.

Quoted 7 lines
Quoted 4 lines

Ah, but it won't be. Low-volume ASICs are expensive. The chip itself would probably be around $150, not counting $100k NRE. Then you have to pay for the board, make up for the NRE, and make some profit to make it worth while. How much are YOU willing to pay?

Then why are companies doing ASICs with an OpenRISC core?

I don't see how that relates. And to answer your question, maybe because there's a demand for it. I don't know.

Quoted 7 lines
Quoted 5 lines

Whatever we design is going to be EXPENSIVE. So, regardless of the fact that an ATI All-in-Wonder Radeon 9000 is over-powered for job you describe, that board will be cheaper than what we could produce.

Quoted 5 lines

And it's probably overpowered when speaking about power consumption, too...

Now think of a SoC with an OpenRISC core and an integrated graphics controller...

That would be cool. :)

#33 lore raw
In reply to Geert Uytterhoeven
Quoted 11 lines · Geert Uytterhoeven
Quoted 6 lines · Timothy Miller

Ok, so, how about this idea:

- Small Xilinx FPGA, 16M of RAM, and a DAC on a board.
- AGP 2X
- Up to 2048x2048 resolution at 8, 16, and 32 bpp.
- Acceleration ONLY for solid fills and bitblts on-screen.

Sounds OK to me.

To you. But if you are the only customer, that doesn't make for very large sales volumes.

Quoted 7 lines
Quoted 7 lines

Given that so little is accelerated, there is no point in putting more than the viewable framebuffer on the card, hense the 16 megs. It would probably actually HURT performance to cache pixmaps on the card.

Quoted 8 lines
Quoted 4 lines

Oh, there's one thing I forgot. It would have to support VGA. There is a VGA core on opencores.org that we could use, but its logic area would probably push up the FPGA cost so that the board was in the $100 range.
Probably more.

Why support legacy VGA? It makes things more complex and expensive, and doesn't give us much, especially for a SoC.

It's all about console support in a PC.

BTW, What is SoC?

#34 lore raw
Quoted 8 lines · Timothy Miller
Quoted 7 lines · Geert Uytterhoeven
Quoted 4 lines · Timothy Miller

Oh, there's one thing I forgot. It would have to support VGA. There is a VGA core on opencores.org that we could use, but its logic area would probably push up the FPGA cost so that the board was in the $100 range.
Probably more.

Quoted 6 lines
Quoted 2 lines

Why support legacy VGA? It makes things more complex and expensive, and doesn't give us much, especially for a SoC.

It's all about console support in a PC.

BTW, What is SoC?

System-on-a-Chip, i.e. a complete solution in one chip.

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]

In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds