[bowman]

http://www.realworldtech.com/page.cf...WT090808195242

So is it G100, G200 or GT200 then?

Anyways, a focus on this GPU as a general purpose processor, ignoring the graphics aspect.

[armchair_architect]

Wow. Excellent, excellent writeup. It explains a lot of things that NVIDIA has already disclosed, but in an easy to understand way and coherent way. The article also reveals some significant details that weren't known before, as far as I know (though some were guessed at). That's going on my bookmark list, I have a feeling I'll be pointing people at it often.

[PeterT]

I found this observation particularly interesting:

In an interesting twist, some of NVIDIA's newest additions to CUDA actually violate the SIMT architecture and philosophy. As mentioned previously in this article, the Compute 1.2 specification includes warp voting functions, particularly the __any() and __all() functions. The __any() function takes a predicate as an argument and returns true if the input predicate is true for any of the 32 threads in a warp. The __all() function works similarly but requires that the predicate evaluate true for every thread in the warp. What is interesting about these two functions is that warps are now an architecturally visible feature of Compute 1.2 devices. This is perhaps one step down the path towards explicit vectors and the classic SIMD model.

I knew about the new voting functions, but I did not realize the extent to which they go against the so-called SIMT philosophy. It will be interesting to see how this plays out, since I believe that the fact that the actual complexity of the SIMD architecture is quite well-hidden in CUDA is one of the reasons for its rather quick adoption and success.

Another part that should be informative even to people somewhat familiar with the architecture is Dual 'Issue' on page 9. I never understood how the MAD/MUL 'dual' issuing was supposed to work, now it's clear.

[pcchen]

Basically, warpsize, all, and any are all used for optimization. You don't need to use them, since the automatic branch divergence handling in CUDA is pretty good. However, in some cases you can use them to get a better performance. Anyway, I think it's a nice thing, because you won't need them if you don't need them, and that's good

[Tim Murray]

Great article, Dave. It ranks up there with Rys' G80 piece or Wavey's Xenos articles as an excellent architectural overview.

and yeah, I'll probably be pointing people at it as a very deep hardware primer. thanks for making my life easier

[dkanter]

Thanks for the kudos everyone!

I had a great time doing this article and I'm glad there are folks out there who appreciate it.

A big thank you goes out to a lot of the people who spent hours talking with me and clarifying everything...

David

[wingless]

Originally Posted by dkanter

Thanks for the kudos everyone!

I had a great time doing this article and I'm glad there are folks out there who appreciate it.

A big thank you goes out to a lot of the people who spent hours talking with me and clarifying everything...

David

Mr. Kanter, I hope you know how seriously your words are taken and just how much the tech world appreciates your work. Thank you very much for the article. I've yet to see such a complete write-up on both the architecture and software side of the GPGPU front. Will we see an ATI write-up soon?

There have been some recent developments on ATI's side: http://www.techreport.com/discussions.x/15490

I would love to know what you think about ATI's single RV770 hardware and OpenCL, CAL, and/or Brook+. It seems a 4870X2 may not be ideal for GPGPU but a single HD 4870 may work well according to what you stated in this current article. Please let the B3D community know if we'll see an ATI write-up soon!

[dkanter]

Originally Posted by wingless

Mr. Kanter, I hope you know how seriously your words are taken and just how much the tech world appreciates your work. Thank you very much for the article. I've yet to see such a complete write-up on both the architecture and software side of the GPGPU front. Will we see an ATI write-up soon?

There have been some recent developments on ATI's side: http://www.techreport.com/discussions.x/15490

I would love to know what you think about ATI's single RV770 hardware and OpenCL, CAL, and/or Brook+. It seems a 4870X2 may not be ideal for GPGPU but a single HD 4870 may work well according to what you stated in this current article. Please let the B3D community know if we'll see an ATI write-up soon!

I'd love to write about RV770 and R700. I only have 24 hours in each day though : (

DK

[wingless]

Originally Posted by dkanter

I'd love to write about RV770 and R700. I only have 24 hours in each day though : (

DK

Sorry for my late response. I live in Houston and still don't have electricity. I put a charge on my old 2001 laptop at my brother's house and dial-up is working at least.

Please take your time. I think we'll have a lot of time before we see something new from AMD on this front.

I really wonder if AMD and Nvidia are planning on adding ECC to their cards to enhance their GPGPU flexibility. It seems like it would be the next logical step at making GPUs more marketable in the specialized computing industry. I simply can't wait to see what will come about next year, especially with Intel coming into the mix. The GPGPU market in 2009 will be really exciting I think.

In the meantime, I'm waiting for Photoshop CS4 to dazzle the masses when paired with the GT200(if it is bug free). I may have to pick up a GTX 260 Core 216 to mess around with.