AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

AMD hasn't talked about tensor tasks a lot, but they have said they support them. Can't find the quote now but I think it was in the Financial Analyst Day broadcast

Yeah they did, but if they have a sizable advantage over Pascal (P100) ya know they will be touting it :)
 
Last edited:
View attachment 2001
Like the DeepBench results that were 50% faster than P100 despite the theoretical 21 vs 25 TFLOPs advantage only being 20%?

Carsten you may want to revise your response to me back last week :)

I thought, we (as in the people in this thread) had mostly agreed upon that being the usual cherrypicking long before.
Especially as it involves that test, along with the use of consumer Titan xp for Solidworks vs Frontier on their website instead of the much higher performing Quadro designed for professional visualisation and specific drivers/libraries - and to emphasise each company I agree do their own bit of cherry picking (Nvidia did it when comparing to their own previous products).
Cheers
 
Last edited:
Bear in mind that's just the cost of a DGX-1 divided by the number of GPUs, as you can only get V100 via that route for a quarter or so.

Yeah.
And to go with your context current price is specifically the Mezzanine solution and not PCIe that is cheaper.
As a general reference here were the prices in 2016 (clients may see different offered prices), and a big difference between PCIe and Mezzanine.

nvidia-pascal-table.jpg


Although I have to say I think many were offered the full node DGX-1 with an ok discount, the price it was offered to us last year in Q3 was pretty good relative to 'street price'.
P40 price looks a bit wrong though *shrug*, but context anyway was regarding the price differentiation between a DGX-1/NVLink P100 and the PCIe models.
Cheers
 
Last edited:
View attachment 2001
Like the DeepBench results that were 50% faster than P100 despite the theoretical 21 vs 25 TFLOPs advantage only being 20%?


Deep bench didn't test solely on tensor performance , nor did they even use the CUDA version which they should have used with the nV products, so no they weren't touting, in fact they downplayed it, saying its not about beating anyone, they were just happy being on the board, because they full well know, with the CUDA version of that bench nV would probably be ahead of Vega with Open CL (ROCm)

In the full context of what they stated with the deep bench benchmark table, they know they aren't going to compete, is what I got from it, anything else, is just ignoring what Raja stated. If AMD's own people are saying, this, why would you believe otherwise? You think its wise for them to downplay something when they know its probably the fastest growing and profitable market, a market that will grow to be in the multiple billions? What are the reasons to down play something like that, its not like nV will have an answer if AMD trumps them in the short term right?

And if anyone thinks AMD would do anything meaningful with Vega in DL markets, yeah that's a fools hope. They don't have the market cap, the money, the penetration is software, or software development to do anything in the market in the short to mid term with Vega, and in the long term its still in the air.

Can't sit here and say the billions nV has spent to push CUDA to colleges, to developers, for the past 5 generations, can be overturned by AMD because they have a product that can or cannot compete, but are in the market now. Just doesn't happen. Intel is having difficulty in getting any kind of head way in DL markets, and yet AMD with just getting some hardware out in the near future without any software support will take nV by storm..... AMD isn't even expecting this and they were point blank in saying so in their financial conference.

Sounds far fetched.
 
Last edited:
What I think is really funny is the ridiculous amount of nitpicking over one slide showing a single comparison.
And when the card comes out, I bet no one here will bother to pick this data and compare it to anecdotal results.
 
What I think is really funny is the ridiculous amount of nitpicking over one slide showing a single comparison.
And when the card comes out, I bet no one here will bother to pick this data and compare it to anecdotal results.
Because if Nvidia did the same against AMD or Intel most of us would be critical about using it, although Nvidia were carefully selective at times even with V100 in how they showed relative performance to P100, some of which comes down to the framework changes (such as Caffe2) or Nvidia libraries (cuDNN and TensorRT) and how they can be used now.
Cheers
 
Because if Nvidia did the same against AMD or Intel most of us would be critical about using it, although Nvidia were carefully selective at times even with V100 in how they showed relative performance to P100, some of which comes down to the framework changes (such as Caffe2) or Nvidia libraries (cuDNN and TensorRT) and how they can be used now.
Cheers

I said the amount of nitpicking was funny, not that said nitpicking was vendor-dependent.
 
What I think is really funny is the ridiculous amount of nitpicking over one slide showing a single comparison.
And when the card comes out, I bet no one here will bother to pick this data and compare it to anecdotal results.

It's pretty similar to the ridiculous marketing slides by Nvidia that almost never start at 0. We should be able to call bs for what it is. But then again, that was the "analyst" conference so I'm sure bs is always on the menu :p
 
Last edited:
And if anyone thinks AMD would do anything meaningful with Vega in DL markets, yeah that's a fools hope. They don't have the market cap, the money, the penetration is software, or software development to do anything in the market in the short to mid term with Vega, and in the long term its still in the air.
So Google will be dominating that market? Because it seems more likely the specialized TPU from Google and others will be used for the DL workloads while all the GPUs continue to handle more generalized workloads with DL on the side.

Sounds far fetched.
I quoted you stating that if AMD had an advantage they would be touting it. A picture of their CEO on a stage in front of a bunch of analysts touting DL performance would seem to be just that. Regardless of the content of the benchmark. The rest is just moving goalposts.
 
Apple just announced that the new iMac Pro will have a 11TFLOPs Vega GPU with up to 16GB VRAM. Sounds like a slightly lower-clocked version of the full 250W-ish cards we'll be getting for the PC.

They're coming with a Skylake-X. Shame they didn't go with a full AMD setup with Threadripper, though I guess when asking $5000 minimum for the setup it would be bad if it didn't bundle the CPUs with the highest single-threaded performance.

New iMacs are also carrying a bunch of new names for Polaris GPUs, like Radeon Pro 555 and 575.

It seems Metal 2 will only support AMD GPUs.





And here come all those straight-from-the-ass comments suggesting AMD graphics being in apple products is actually "bad" because apple pays only 2 cents per GPU to AMD and AMD loses money by selling to apple.
 
They're coming with a Skylake-X. Shame they didn't go with a full AMD setup with Threadripper, though I guess when asking $5000 minimum for the setup it would be bad if it didn't bundle the CPUs with the highest single-threaded performance.

They are going with Intel because of Thunderbolt 3. These Xeons will have lower single thread performance than most desktop Ryzen or Sky-X CPUs, mainly due to lower clock speed.
 
So Google will be dominating that market? Because it seems more likely the specialized TPU from Google and others will be used for the DL workloads while all the GPUs continue to handle more generalized workloads with DL on the side.
TPU is specfic to only part of the DL market, and that is what Volta's tensor units are focused on too. But Volta has more then just tensor units it can do the rest of the DL markets.
I quoted you stating that if AMD had an advantage they would be touting it. A picture of their CEO on a stage in front of a bunch of analysts touting DL performance would seem to be just that. Regardless of the content of the benchmark. The rest is just moving goalposts.


They didn't tout it, they down played it, do you want the transcripts and audio files linked? Showing a totally skewed benchmark in their favor and then saying its not about beating anyone, is saying what? That benchmark is BS. Just like anything AMD has shown for any launch. Just like nV has done in the past in certain circumstances and Intel has done at times too, Even google did with their TPU too.

Its like you trying to compare 25 tflops of 16bit precision to Volta's 32 bit flops to prove a point, you just don't do that, two totally different metrics which have no correlation with each other.

Its nice that you are trying to think outside of the box and trying to give reasons for Vega to be Volta's competitor in everything (because they both start with V), but I don't think it will hold a candle to Volta, and its going to have problems against Pascal.

lets go by the alphabetical synonymy shell we?

AMD might have been trying to shoot for that for marketing purposes, but damn it sure doesn't look likely to pan out does it?

What happens with Polaris vs. Pascal? Yeah don't need to look to far to see that fails.
 
Last edited:
22tflops of half precision @ boost clocks, that is slower *lower clocked or cut down* than the instinct version of their cards
It's a giant all in one device that mounts to a wall. Is it really a surprise a mobile part doesn't run as fast as a server part?

Its nice that you are trying to think outside of the box and trying to give reasons for Vega to be Volta's competitor in everything (because they both start with V), but I don't think it will hold a candle to Volta, and its going to have problems against Pascal.
That's not outside the box, it's how past devices worked with slight modifications. What's strange is you assuming AMD would release two different new architectures the same time. Then dismissing the marketing materials calling out Volta, similar release dates, display support, memory models, etc.
 
Back
Top