AMD Releases R600 GPU Documentation

[bridgman]

Do AMD have a Batman type of cave under the HQ with 1 meter concrete walls all around...

Not sure about AMD HQ, but the old ATI HQ in Markham does. So yes...

...where the tapes and harddrive a stored and guarded doors looking like bank vaults?

... and no. We use it for underground parking, UPSes for the data center, and a big bunker of diesel fuel for the backup generators

The walls are so thick because there seems to be an underground river flowing through that spot -- when we dug the big hole for the new building the hole mostly filled up with water over the weekend. Backups are all offsite, which is probably good - if anything really bad happened to the data center, I wouldn't want the same thing to happen to our backup tapes

[some-guy]

Great Job!
I'm almost completely sure my next card is going to be ATi

Do AMD have a Batman type of cave under the HQ with 1 meter concrete walls all around where the tapes and harddrive a stored and guarded doors looking like bank vaults?

Nope, they put it on 50 servers running intel cpu's xD

[Louise]

We think so

When I talk to people who are not in the GPU business everyone is always amazed how big and complex the chips are -- almost 670 million transistors for an HD38xx and climbing. CPUs are typically smaller (a Phenom is ~460 million), and if you look at die shots much more of the CPU is cache memory than on a GPU. Even an RV610, which pretty much only has 1 of everything (2 4-way SIMD blocks, 1 quad texture, 1 quad ROP), is still almost 200 million transistors.



I have to admit that 100,000 is a guess -- I have never actually counted it all; I know that the one time I did start counting I got close to 10,000 pages pretty quickly and hadn't made a big dent in the documentation tree. A lot of it is auto-generated or in databases, so at least it doesn't all have to be typed in.

Most of the time we get lucky and don't have to dig into the really detailed stuff, or we would need as many people to read it as it took to write it



The worst part is the "my head's gonna burst" feeling we get when starting on a new generation of GPUs. We're feeling it right now with the R6xx

These transistor numbers are crazy One thing is to be pack to many small things into such a small die, another is that they actually do something! Crazy!

I can't help but wonder how all the documentation is backed up

Do AMD have a Batman type of cave under the HQ with 1 meter concrete walls all around where the tapes and harddrive a stored and guarded doors looking like bank vaults?

[bridgman]

You guys have the coolest job in the world!

We think so

When I talk to people who are not in the GPU business everyone is always amazed how big and complex the chips are -- almost 670 million transistors for an HD38xx and climbing. CPUs are typically smaller (a Phenom is ~460 million), and if you look at die shots much more of the CPU is cache memory than on a GPU. Even an RV610, which pretty much only has 1 of everything (2 4-way SIMD blocks, 1 quad texture, 1 quad ROP), is still almost 200 million transistors.

I would never have thought that there are written 100.000+ pages of specs for a GPU.

I have to admit that 100,000 is a guess -- I have never actually counted it all; I know that the one time I did start counting I got close to 10,000 pages pretty quickly and hadn't made a big dent in the documentation tree. A lot of it is auto-generated or in databases, so at least it doesn't all have to be typed in.

Most of the time we get lucky and don't have to dig into the really detailed stuff, or we would need as many people to read it as it took to write it

It must be a strange feeling to have access to all this secret information!

The worst part is the "my head's gonna burst" feeling we get when starting on a new generation of GPUs. We're feeling it right now with the R6xx

[Louise]

We're really smart

Seriously, all good questions. It's not that we don't have information, it's that we have too much. There's maybe 100,000 pages of documentation for a typical GPU, not counting test programs or test reports. We (mostly Alex really) pick through the information and build up documents, sometimes by cherry picking from design docs, sometimes by writing explanations to fill the gaps.



We have a good mix of skills (I designed graphics hardware for a lot of years, Alex knows the 2d area & video area really well, we have hired an ex-ATI 3d developer as well, and we get help from folks in the community like Airlie and Glisse who have been working with the 3d innards for years. We also have direct access to the top hardware and software architects although we try not to suck up too much of their time.



Yes, hugely detailed specs -- but that documentation is primarily "block level" internals, not "here is a programming manual". We write the programming manuals, at least for the acceleration bits. For display & modesetting, we planned to leverage AtomBIOS to let the driver development move quickly, and also provide enough register spec information to support debugging and make sure developers could understand what AtomBIOS was doing and troubleshoot problems.

EDIT - I should mention that for this particular document I believe we made use of an outside tech writing firm (at least that's the impression I got from the revision history) to turn the raw documentation into nicely finished manuals. My primary contribution this time was noticing the document while playing with CAL, getting permission to publish it, and letting Alex know that another 50 pages could be chopped out of the R6xx programming guide

Thank you so much, for your detailed reply I most have read it at least 10 times, as it just sounds too amazing

I would never have thought that there are written 100.000+ pages of specs for a GPU. It must be a strange feeling to have access to all this secret information!

You guys have the coolest job in the world!

[bridgman]

There's enough information out there now to get AFR Crossfire running on 5xx, although (see the Ask ATI Dev thread) we haven't yet documented the block that handles inter-chip compositing to accelerate modes like SFR and SuparAA. I just didn't want to slow down the rest of the 3d docs so haven't even looked at it.

The big issue with Crossfire (and SLI) is that they require a lot of software work to intelligently split the work across the GPUs. Putting the results back together is the easy part; we only use the special hardware on the high end boards anyways.

[sabriah]

Phoronix: AMD Releases R600 GPU Documentation

AMD's Alex Deucher has today announced the availability of the documentation covering the R600 Family Instruction Set Architecture. This ISA (Instruction Set Architecture) documentation covers the unified shader block found on the Radeon HD 2000/3000 series and newer...

http://www.phoronix.com/vr.php?view=NjUyMA

Thanks AMD/ATI! Great move!


Now I wonder if that will eventually lead to Crossfire for Linux too. Anyone with better perspectives?

[remm]

Meanwhile nvidia is offering prizes for cuda coding and still gives reason to own multi-GPU cards in linux, they could be preparing code for public release right now just sitting on it until the need is there to release it.

Hey, that's great to know. So I have to use the proprietary driver on F9 (after having to use the nv piece of junk for weeks - nouveau is really not there for G80 cards like my 8600) just because they're "sitting on code". Great, then I'll be "sitting" on buying Intel or ATI for a very long time May Larabee come and do some serious damage to these arrogant bastards

[d2kx]

This is how I want the morning to start. Awesome

[oblivious_maximus]

WOOHOO!

This message was too short before this sentence.