By: Adrian (a.delete@this.acm.org), August 30, 2022 1:12 am
Room: Moderated Discussions
Mark Roulo (nothanks.delete@this.xxx.com) on August 29, 2022 6:26 pm wrote:
> anonymous2 (anonymous2.delete@this.example.com) on August 29, 2022 5:08 pm wrote:
> > AVX-512 (ISA details murky) on Zen 4 but 2 cycles vs 1 on Intel so only 256b internally.
> >
> > Small win for those who want the ISA, but from a performance perspective limited value?
>
> This feels like a "we can *RUN* your code, but not terribly fast" feature. Many/most
> folks who care about numeric performance are moving the code to GPUs.
>
>
Only the folks with very deep pockets are moving the code that cares about numeric performance to GPUs.
Many years ago, I have moved a lot of code to GPUs, because they had much better values for Gflop/s per dollar and Gflop/s per watt.
Now I have mostly moved back to CPUs, because I could not afford to upgrade my old GPUs. Paying $10000 for a GPU is not something that individuals or small businesses can afford.
The Ryzen 9 7950X is expected to have a cost of around $0.50 per Gflop/s, which approximately matches the lowest computation cost that has ever been provided by GPUs.
Also the expected value for Gflop/s per watt is about the same as it was for the best GPUs before their transition to FinFET CMOS processes.
The only advantage of an up-to-date datacenter GPU is that it has a better Gflops per watt value, by a factor around 3. The much lower power consumption is essential for someone who buys CPUs and GPUs by the hundreds or by the thousands. On the other hand, for a small user the huge price of a datacenter GPU cannot be justified, while the gaming GPUs are good only for graphics and machine learning, not for the kinds of workloads that are normally described as caring about numeric performance.
> anonymous2 (anonymous2.delete@this.example.com) on August 29, 2022 5:08 pm wrote:
> > AVX-512 (ISA details murky) on Zen 4 but 2 cycles vs 1 on Intel so only 256b internally.
> >
> > Small win for those who want the ISA, but from a performance perspective limited value?
>
> This feels like a "we can *RUN* your code, but not terribly fast" feature. Many/most
> folks who care about numeric performance are moving the code to GPUs.
>
>
Only the folks with very deep pockets are moving the code that cares about numeric performance to GPUs.
Many years ago, I have moved a lot of code to GPUs, because they had much better values for Gflop/s per dollar and Gflop/s per watt.
Now I have mostly moved back to CPUs, because I could not afford to upgrade my old GPUs. Paying $10000 for a GPU is not something that individuals or small businesses can afford.
The Ryzen 9 7950X is expected to have a cost of around $0.50 per Gflop/s, which approximately matches the lowest computation cost that has ever been provided by GPUs.
Also the expected value for Gflop/s per watt is about the same as it was for the best GPUs before their transition to FinFET CMOS processes.
The only advantage of an up-to-date datacenter GPU is that it has a better Gflops per watt value, by a factor around 3. The much lower power consumption is essential for someone who buys CPUs and GPUs by the hundreds or by the thousands. On the other hand, for a small user the huge price of a datacenter GPU cannot be justified, while the gaming GPUs are good only for graphics and machine learning, not for the kinds of workloads that are normally described as caring about numeric performance.