AMD's Radeon HD 6970 & Radeon HD 6950: Paving The Future For AMD
by Ryan Smith on December 15, 2010 12:01 AM ESTCayman: The New Dawn of AMD GPU Computing
We’ve already covered how the shift from VLIW5 to VLIW4 is beneficial for AMD’s computing efforts: narrower SPUs are easier to fully utilize, FP64 performance improves to 1/4th FP32 performance, and the space savings give AMD room to lay down additional SIMDs to improve performance. But if Cayman is meant to be a serious effort by AMD to relaunch themselves in to the GPU computing market and to grab a piece of NVIDIA’s pie, it takes more than just new shaders to accomplish the task. Accordingly, AMD has been hard at work to round out the capabilities of their latest GPU to make it a threat for NVIDIA’s Fermi architecture.
AMD’s headline compute feature is called asynchronous dispatch, a long word that actually does a pretty good job of describing what it does. To touch back on Fermi for a moment, with Fermi NVIDIA introduced support for parallel kernels, giving Fermi the ability to execute multiple kernels at once. AMD in turn is following NVIDIA’s approach of executing multiple kernels at once, but is going to take it one step further.
The limit of NVIDIA’s design is that while Fermi can execute multiple kernels at once, each one must come from the same CPU thread. Independent threads/applications for example cannot issue their own kernels and have them execute in parallel, rather the GPU must context switch between them. With asynchronous dispatch AMD is going to allow independent threads/applications to issue kernels that execute in parallel. On paper at least, this would give AMD’s hardware a significant advantage in this scenario (context switching is expensive), one that would likely eclipse any overall performance advantages NVIDIA had.
Fundamentally asynchronous dispatch is achieved by having the GPU hide some information about its real state from applications and kernels, in essence leading to virtualization of GPU resources. As far as each kernel is concerned it’s running in its own GPU, with its own command queue and own virtual address space. This places more work on the GPU and drivers to manage this shared execution, but the payoff is that it’s better than context switching.
For the time being the catch for asynchronous dispatch is that it requires API support. As DirectCompute is a fixed standard this just isn’t happening – at least not with DirectCompute 11. Asynchronous dispatch will be exposed under OpenCL in the form of an extension.
Meanwhile the rest of AMD’s improvements are focusing on memory and cache performance. While the fundamental architecture is not changing, there are several minor changes here to improve compute performance. The Local Data Store attached to each SIMD is now able to bypass the cache hierarchy and Global Data Store by having memory fetches read directly in to the LDS. Meanwhile Cayman is getting a 2nd DMA engine, improving memory reads & writes by allowing Cayman to execute two at once in each direction.
Finally, read ops from shaders are being sped up a bit. Compared to Cypress, Cayman can coalesce them in to fewer operations.
As today’s launch is primarily about the Radeon HD 6900 series AMD isn’t going too much in depth on the compute side of things, so everything here is a fairly high level overview of the architecture. Once AMD has Firestream cards ready to go with Cayman in them, there will likely be more to talk about.
Facebook
Twitter
RSS
168 Comments
View All Comments
AnnonymousCoward - Wednesday, December 15, 2010 - link
First of all, 30fps is choppy as hell in a non-RTS game. ~40fps is a bare minimum, and >60fps all the time is hugely preferred since then you can also use vsync to eliminate tearing.Now back to my point. Your counter was "you know that non-AA will be higher than AA, so why measure it?" Is that a point? Different cards will scale differently, and seeing 2560+AA doesn't tell us the performance landscape at real-world usage which is 2560 no-AA.
Dug - Wednesday, December 15, 2010 - link
Is it me, or are the graphs confusing.Some leave out cards on certain resolutions, but add some in others.
It would be nice to have a dynamic graph link so we can make our own comparisons.
Or a drop down to limit just ati, single card, etc.
Either that or make a graph that has the cards tested at all the resolutions so there is the same number of cards in each graph.
benjwp - Wednesday, December 15, 2010 - link
Hi,You keep using Wolfenstein as an OpenGL benchmark. But it is not. The single player portion uses Direct3D9. You can check this by checking which DLLs it loads or which functions it imports or many other ways (for example most of the idTech4 renderer debug commands no longer work).
The multiplayer component does use OpenGL though.
Your best bet for an OpenGL gaming benchmark is probably Enemy Territory Quake Wars.
Ryan Smith - Wednesday, December 15, 2010 - link
We use WolfMP, not WolfSP (you can't record or playback timedemos in SP).7Enigma - Wednesday, December 15, 2010 - link
Hi Ryan,What benchmark do you use for the noise testing? Is it Crysis or Furmark? Along the same line of questioning I do not think you can use Furmark in the way you have the graph setup because it looks like you have left Powertune on (which will throttle the power consumption) while using numbers from NVIDIA's cards where you have faked the drivers into not throttling. I understand one is a program cheat and another a TDP limitation, but it seems a bit wrong to not compare them in the unmodified position (or VERBALLY mention this had no bearing on the test and they should not be compared).
Overall nice review, but the new cards are pretty underwhelming IMO.
Ryan Smith - Thursday, December 16, 2010 - link
Hi 7Enigma;For noise testing it's FurMark. As is the case with the rest of our power/temp/noise benchmarks, we want to establish the worst case scenario for these products and compare them along those lines. So the noise results you see are derived from the same tests we do for temperatures and power draw.
And yes, we did leave PowerTune at its default settings. How we test power/temp/noise is one of the things PowerTune made us reevaluate. Our decision is that we'll continue to use whatever method generates the worst case scenario for that card at default settings. For NVIDIA's GTX 500 series, this means disabling OCP because NVIDIA only clamps FurMark/OCCT, and to a level below most games at that. Other games like Program X that we used in the initial GTX 580 article clearly establish that power/temp/noise can and do get much worse than what Crysis or clamped FurMark will show you.
As for the AMD cards the situation is much more straightforward: PowerTune clamps everything blindly. We still use FurMark because it generates the highest load we can find (even with it being reduced by over 200MHz), however because PowerTune clamps everything, our FurMark results are the worst case scenario for that card. Absolutely nothing will generate a significantly higher load - PowerTune won't allow it. So we consider it accurate for the purposes of establishing the worst case scenario for noise.
In the long run this means that results will come down as newer cards implement this kind of technology, but then that's the advantage of such technology: there's no way to make the card louder without playing wit the card's settings. For the next iteration of the benchmark suite we will likely implement a game-based noise test, even though technologies like PowerTune are reducing the dynamic range.
In conclusion: we use FurMark, we will disable any TDP limiting technology that discriminates based on the program type or is based on a known program list, and we will allow any TDP limiting technology that blindly establishes a firm TDP cap for all programs and games.
-Thanks
Ryan Smith
7Enigma - Friday, December 17, 2010 - link
Thanks for the response Ryan! I expected it to be lost in the slew of other posts. I highly recommend (as you mentioned in your second to last paragraph) that a game-based benchmark is used along with the Furmark for power/noise. Until both adopt the same TDP limitation it's going to put the NVIDIA cards in a bad light when comparisons are made. This could be seen as a legitimate beef for the fanboys/trolls, and we all know the less ammunition they have the better. :)Also to prevent future confusion it would be nice to have what program you are using for the power draw/noise/heat IN the graph title itself. Just something as simple as "GPU Temperature (Furmark-Load)" would make it instantly understandable.
Thanks again for the very detailed review (on 1 week nonetheless!)
Hrel - Wednesday, December 15, 2010 - link
I really hope these architexture changes lead to better minimum FPS results. AMD is ALWAYS behind Nvidia on minimum FPS and in many ways that's the most important measurment since min FPS determines if the game is playable or not. I dont' care if it maxes out 122 FPS if when the shit hits the fan I get 15 FPS, I won't be able to accurately hit anything.Soldier1969 - Wednesday, December 15, 2010 - link
I'm dissapointed in the 6970, its not what I was expecting over my 5870. I will wait to see what the 6990 brings to the table next month. I'm looking for a 30-40% boost from my 5870 at 2560 x 1600 res I game at.stangflyer - Wednesday, December 15, 2010 - link
Now that we see the power requirements for the 6970 and that it needs more power than the 5870 how would they make a 6990 without really cutting off the performance like the 5970?I had a 5970 for a year b4 selling it 3 weeks ago in preparation of getting 570 in sli or 6990.
It would obviously have to be 2x8 pin power! Or they would have to really use that powertune feature.
I liked my 5970 as I didn't have the stuttering issues (or i don't notice them) And actually have no issues with eyefinity as i have matching dell monitors with native dp inputs.
If I was only on one screen I would not even be thinking upgrade but the vram runs out when using aa or keeping settings high as I play at 5040x1050. That is the only reason I am a little shy of getting the 570 in sli.
Don't see how they can make a 6990 without really killing the performance of it.
I used my 5970 at 5870 and beyond speeds on games all the time though.