AMD Vega GPU architecure uncovered

[CAT-THE-FIFTH]

My guess is around 460mm2.

A bit smaller than I expected,but still close to GP102 level size. If this is the top Vega GPU,then it will be dual use and have to cover similar markets to the GP100 too,so I wonder how much it will affect gaming performance??

[me-yeah]

The page is about firepro graphics cards. It's right there in the url. The specific technology stack it's talking about - Stream - only works on firepro graphics cards. Stream != OpenCL. OpenCL runs on all AMD GPUs (you need GCN for OpenCL 2.0). Stream is only applicable to firepro. These are different things.

And I'm afraid I simply can't parse the rest of your comment into something meaningful. AFAICT it boils down to a claim that cuda > opencl therefore nvidia > AMD?

Comparing OpenCL to CUDA is pointless - OpenCL is an open standard for compute that can run on any GPU/CPU/accelerator with appropriate drivers/software, while CUDA is a proprietary software stack that only runs on nvidia cards. Of course nvidia invest a lot more effort into pushing CUDA - it ties people to their graphics cards and makes them money. Besides, the first google result for 'AMD OpenCL' is http://developer.amd.com/tools-and-sdks/opencl-zone/ - compare that to the first result for 'nvidia cuda': http://www.nvidia.co.uk/object/cuda-...puting-uk.html ... I'm not seeing a huge amount of difference in the level of content available between those sites...

its the only page AMD has to explain stream technology, while on all the pages for RX and R9 cards they have the stream logo and the stream processor XXXX number in the specifications

https://developer.nvidia.com/cuda-zone

[scaryjim]

A bit smaller than I expected,but still close to GP102 level size. If this is the top Vega GPU ...

If it's the top Vega GPU that would suggest a shift in priority from the original plan, which said Vega 10 would be the smaller version, with a larger Vega 11 coming later?!

A 460mm² chip would be almost twice the size of Polaris 10. Given HBM memory interfaces are meant to use less silicon than GDDR5, that would imply more than twice the shaders of Polaris 10?

I'm concerned what that actually means for Vega - all the previous rumours were stacking up so nicely with the limited information we've been fed so far, in particular the 4096 shaders @ ~ 1500MHz giving 12.5TFlops compute. But that's a lot less than twice as many shaders as Polaris 10, so if the rumours and die size hold it's really burning silicon area for something other than shader count (maybe to enable the higher clock speeds?). But perhaps more worryingly, the implied performance was just behind GTX 1080. If they've had to go to a 460mm² die just to at best equal GTX 1080, they don't have much room to go after any derivatives of GP102.

My expectation would've been a ~ 400mm² Vega 10 with Vega 11 pushing up to a Fiji-esque ~ 600mm². I know it's a wait and see game, but my expectations of Vega have just taken a bit of a nose-dive...

[CAT-THE-FIFTH]

Vega 10 and Vega 20 slides are leaked:

http://videocardz.com/65521/amd-vega...lides-revealed

It looks like the GPU demoed is the smaller Vega 10,so hopefully it means it will be the gaming orientated GPU of the lineup.

That should mean it is more a Titan X competitor IMHO OFC.

Edit!!

Read it wrong - it looks like Vega 10 is the bigger chip.

[CAT-THE-FIFTH]

If it's the top Vega GPU that would suggest a shift in priority from the original plan, which said Vega 10 would be the smaller version, with a larger Vega 11 coming later?!

A 460mm² chip would be almost twice the size of Polaris 10. Given HBM memory interfaces are meant to use less silicon than GDDR5, that would imply more than twice the shaders of Polaris 10?

I'm concerned what that actually means for Vega - all the previous rumours were stacking up so nicely with the limited information we've been fed so far, in particular the 4096 shaders @ ~ 1500MHz giving 12.5TFlops compute. But that's a lot less than twice as many shaders as Polaris 10, so if the rumours and die size hold it's really burning silicon area for something other than shader count (maybe to enable the higher clock speeds?). But perhaps more worryingly, the implied performance was just behind GTX 1080. If they've had to go to a 460mm² die just to at best equal GTX 1080, they don't have much room to go after any derivatives of GP102.

My expectation would've been a ~ 400mm² Vega 10 with Vega 11 pushing up to a Fiji-esque ~ 600mm². I know it's a wait and see game, but my expectations of Vega have just taken a bit of a nose-dive...

Except Polaris 10 is not a part which has considerable FP64 capability - Vega 10 is most likely going to replace both Fiji and Hawaii based parts.



This means Vega 10 needs decent FP64 performance,which means additional space taken up for such functionality.

Remember,Fiji did not replace Hawaii in certain markets as it has a worse FP64 performance rate IIRC.

Unless the uarch changes have massively improved core utilisation,clockspeeds,etc the gaming orientated GP102 is not only slightly bigger,but should beat Vega10 IMHO OFC.

[scaryjim]

its the only page AMD has to explain stream technology, while on all the pages for RX and R9 cards they have the stream logo ...

Really? 'cause I can't see the Stream logo on http://www.amd.com/en-us/products/gr.../radeon-rx-480
Or on http://www.amd.com/en-us/products/graphics/desktop/r9, either.

You really need to stop comparing CUDA to Stream. They're not equivalent. The nearest equivalent AMD has to CUDA is OpenCL, but of course that's an open standard they support, rather than a closed ecosystem they control. Stream is simply a set of technologies that enhance the running of OpenCL on certain graphics cards. I've already given you the comparable AMD link to nvidia's cuda zone - it's their OpenCL zone at

http://developer.amd.com/tools-and-sdks/opencl-zone/

Note that the url is almost identical to nvidia's?

Seriously, I know I work in the industry somewhat, but I don't think the difference is that hard to understand, is it? Certainly I'd hope anyone who worked in a field where they needed to know about parallel compute would be able to understand the difference between OpenCLK, the platform and API they'd use to write compute tasks for AMD GPUs, and Stream, a set of optmising technologies that AMD provides for certain GPU classes? *shrug* Maybe I'm just crediting people with too much intelligence...

[me-yeah]

Really? 'cause I can't see the Stream logo on http://www.amd.com/en-us/products/gr.../radeon-rx-480
Or on http://www.amd.com/en-us/products/graphics/desktop/r9, either.

You really need to stop comparing CUDA to Stream. They're not equivalent. The nearest equivalent AMD has to CUDA is OpenCL, but of course that's an open standard they support, rather than a closed ecosystem they control. Stream is simply a set of technologies that enhance the running of OpenCL on certain graphics cards. I've already given you the comparable AMD link to nvidia's cuda zone - it's their OpenCL zone at

http://developer.amd.com/tools-and-sdks/opencl-zone/

Note that the url is almost identical to nvidia's?

Seriously, I know I work in the industry somewhat, but I don't think the difference is that hard to understand, is it? Certainly I'd hope anyone who worked in a field where they needed to know about parallel compute would be able to understand the difference between OpenCLK, the platform and API they'd use to write compute tasks for AMD GPUs, and Stream, a set of optmising technologies that AMD provides for certain GPU classes? *shrug* Maybe I'm just crediting people with too much intelligence...

either way, if vega architecure has been scaled up for parallel computing, they need to update some webpages, universities and business don't already have the time and money to book a super computer, so they will just turn to cuda for all the extra floating point math processors when running computer science simulations

stream processors not being used in standard desktop programs is'nt really a good sign to hold up

[CAT-THE-FIFTH]

Actually looking at that chart it seems Vega20 is replacing Hawaii,not Vega10. Hence,it does imply it is more a SP part,and makes the performance figures we seen so far more confusing.



The latest TPU reviews put the GTX1080 as being 71% faster at 1080P.

Something seems off,unless the performance figures we have seen are for parts which are not fully enabled,not hitting expected clockspeeds or the drivers are not fully functional.

Another consideration is whether Vega10 is hitting a TDP limit too. Its only 50% more than Polaris10,so it means it needs to be more efficient than Polaris 10 by a decent amount to even hit GP104 level performance.

[scaryjim]

... the gaming orientated GP102 is not only slightly bigger,but should beat Vega10 IMHO OFC.

Given those slides match the rumours I was talking about, GP102 is going to knock Vega 10 into a cocked hat... its best case scenario - unless they've done some serious work improving the overall efficiency in gaming - is to match GTX 1080. It simply doesn't have the grunt to do better. And if Vega 11 is going to be smaller...?

*shrug* yeah, really don't know what AMD are doing in the GPU arena now. If that's all true then they barely have an answer to GP104, let alone GP102. Very disappointed...

[scaryjim]

... universities and business don't already have the time and money to book a super computer, so they will just turn to cuda for all the extra floating point math processors when running computer science simulations ...

Can't speak for businesses but many universities (in the UK at least) have their own HPC clusters and most have allocated resource on a number of national HPC clusters. The people who use them are knowledgable enough to know the difference between the various technological options, and pick the one that best suits their needs and budget. I can guarantee you, however, than none of them will be using consumer-grade GeForce graphics cards as their processor of choice. If they are running cuda, it'll be on professional grade accelerators.

As a little aside, Intel's Knights <X> platform is actually making huge strides in research HPC thanks to a decent investment in developer tools and relations. It's another area that I'm sure AMD would be investing more in if they weren't quite so financially restricted...

[Nemz0r]

AMD are saying Half 1 2017

I do hope sooner rather than later. Need a card or two for a deep learning workstation and as things are right now the CUDA ecosystem is miles ahead and a Titan X(P) has a lovely 12GB of memory, but... gimped FP16 performance. An 8 GB consumer Vega card would have essentially 1/3 more memory to play with and a massive 25 TFLOPS of FP16 performance, plus AMD's new support for machine learning is promising. I hope we see them soon.

The Vega competition on ve.ga painted the worst case scenario: the T&Cs said winners would be selected on 5th February 2017 and they would receive their Vega graphics card prize "within 28 weeks" so... 20th August 2017.

[CAT-THE-FIFTH]

Given those slides match the rumours I was talking about, GP102 is going to knock Vega 10 into a cocked hat... its best case scenario - unless they've done some serious work improving the overall efficiency in gaming - is to match GTX 1080. It simply doesn't have the grunt to do better. And if Vega 11 is going to be smaller...?

*shrug* yeah, really don't know what AMD are doing in the GPU arena now. If that's all true then they barely have an answer to GP104, let alone GP102. Very disappointed...

There is no indication of a smaller Vega GPU now it appears- I still think Vega10 is basically the GPU from the XBox Scorpio and to save on R and D AMD is sharing the design with desktop cards(or maybe it was the other way round??),but if AMD cannot even beat a GTX1080 slightly it is very dissapointing.

[me-yeah]

Can't speak for businesses but many universities (in the UK at least) have their own HPC clusters and most have allocated resource on a number of national HPC clusters. The people who use them are knowledgable enough to know the difference between the various technological options, and pick the one that best suits their needs and budget. I can guarantee you, however, than none of them will be using consumer-grade GeForce graphics cards as their processor of choice. If they are running cuda, it'll be on professional grade accelerators.

As a little aside, Intel's Knights <X> platform is actually making huge strides in research HPC thanks to a decent investment in developer tools and relations. It's another area that I'm sure AMD would be investing more in if they weren't quite so financially restricted...

you might need all of that for a particle physic's simulation, but not for air tunnels and engineering programs after you have finished in CAD

AMD have a tool to convert cuda code to "stream" code https://developer.amd.com/wordpress/...-US-Letter.pdf

[kompukare]

Something seems off,unless the performance figures we have seen are for parts which are not fully enabled,not hitting expected clockspeeds or the drivers are not fully functional.

Well, the Vega architectural overview points to major changes compared to current GCN designs so I would say the driver team will have plenty to optimise.
Strange that according those leaks all three Vegas (11, 10 and 20) have 64CU. That doesn't make sense.

[CAT-THE-FIFTH]

Well, the Vega architectural overview points to major changes compared to current GCN designs so I would say the driver team will have plenty to optimise.
Strange that according those leaks all three Vegas (11, 10 and 20) have 64CU. That doesn't make sense.

Maybe one of the two 14NM Vega GPUs has extra FP64 functionality and the other is a pared down gaming GPU?? This is pretty much what the GP102 is - a pared down GP100 with the same number of shaders and GDDR5X instead of HBM2.

[kompukare]

Maybe one of the two 14NM Vega GPUs has extra FP64 functionality and the other is a pared down gaming GPU?? This is pretty much what the GP102 is - a pared down GP100 with the same number of shaders and GDDR5X instead of HBM2.

Possible, except the slide list Vega11&10 as 1/16 DP, and Vega 20 as 1/2 DP. The Vega 20 makes sense, but Vega 11&10 having the same only makes sense if there is something else to distinguish them as they are both listed as 64CU:

Well, aside from the 225W vs 300W.