4 ms·

Latest Ampere datacenter rumors are 70% speedup compared to V100, meaning 25TFlops in FP32, whereas AMD in theory could to 15Tflops and in practice only 8Tflops

by bitL 7y ago

Latest Ampere datacenter rumors are 70% speedup compared to V100, meaning 25TFlops in FP32, whereas AMD in theory could to 15Tflops and in practice only 8Tflops due to their software. They would have to improve speed 3x just to match NVidia's next generation GPUs; I am not sure upcoming Big Navi would be sufficient.

trynumber9 7y ago
It's not comparable in efficiency but AMD is already shipping Vega II modules in the Mac Pro that claim "up to 28.3 FP32 TFLOPS". And regardless, the article is about 4K gaming. I don't doubt Nvidia will have some 7nm monster but there might be a time period before it launches where AMD briefly regains the crown.
- shaklee3 7y ago
  That's not equivalent. That's a card with two ASICs on it, with each being 14TFLOPS. It's identical to the V100, but 2-3 years later.
  - trynumber9 7y ago
    Nvidia doesn't make any cards with two ASICs on it. There is no equivalent.
    - shaklee3 7y ago
      Yes, they make the K40, which has two ASICs. They do not make a current generation card with two ASICs.
      - bitL 7y ago
        You probably meant K80. K80 = 2x K40 basically, i.e. 2x 12GB 4TFlops GPUs on a single board.
        shaklee3 7y ago
        Yes, thanks for the correction. I should also add that the reason it's not fair to compare these kinds of cards is the inter-processor link is usually extremely slow. Because of that, performance with that kind of card is roughly equivalent to just having two cards in the machine as opposed to having the chips on the same card.
shaklee3 7y ago
I've seen those A100 rumors, but I think they should be taken with a grain of salt. They did not specify where the 70% speedup will be, and it's highly likely it will not be fp32 increasing by that much.
3fe9a03ccd14ca5 7y ago
You’re being downvoted but it’s true. People don’t realize that the data center is now driving the GPU market volume, and AMD is not yet seriously competing in this space.

by bitL 7y ago

trynumber9 7y ago

It's not comparable in efficiency but AMD is already shipping Vega II modules in the Mac Pro that claim "up to 28.3 FP32 TFLOPS". And regardless, the article is about 4K gaming. I don't doubt Nvidia will have some 7nm monster but there might be a time period before it launches where AMD briefly regains the crown.

shaklee3 7y ago
That's not equivalent. That's a card with two ASICs on it, with each being 14TFLOPS. It's identical to the V100, but 2-3 years later.
- trynumber9 7y ago
  Nvidia doesn't make any cards with two ASICs on it. There is no equivalent.
  - shaklee3 7y ago
    Yes, they make the K40, which has two ASICs. They do not make a current generation card with two ASICs.
    - bitL 7y ago
      You probably meant K80. K80 = 2x K40 basically, i.e. 2x 12GB 4TFlops GPUs on a single board.
      - shaklee3 7y ago
        Yes, thanks for the correction. I should also add that the reason it's not fair to compare these kinds of cards is the inter-processor link is usually extremely slow. Because of that, performance with that kind of card is roughly equivalent to just having two cards in the machine as opposed to having the chips on the same card.

shaklee3 7y ago

I've seen those A100 rumors, but I think they should be taken with a grain of salt. They did not specify where the 70% speedup will be, and it's highly likely it will not be fp32 increasing by that much.

3fe9a03ccd14ca5 7y ago

You’re being downvoted but it’s true. People don’t realize that the data center is now driving the GPU market volume, and AMD is not yet seriously competing in this space.