5 ms·

Performance doesn't scale with memory bandwidth, SM count or any other spec that is known to me. RTX 3090 and 3080 are not just similar cards from the same gene

by terafo 5y ago

Performance doesn't scale with memory bandwidth, SM count or any other spec that is known to me. RTX 3090 and 3080 are not just similar cards from the same generation. They are the same GA102 die with some configuration differences : 3090 does have 24gb VRAM while 3080 has only 10. On 3090 82 SMs of initial 84 are active and on 3080 it's 68 SMs. And they both perform only marginally better than 1660 Super, while usually the difference between this cards is in 3x range. It's likely just poorly written software.

klodolph 5y ago
“It’s likely just poorly written software” sounds completely glib to me, it sounds like an intuitive leap rather than a reasoned-out conclusion. Again, it sounds like an unusual workload. I don’t see why performance must scale with some spec in any particular straightforward way—I think we all remembered seeing workloads that would get faster once you added more CPU cores, then slower once you kept adding CPU cores—because back in the day, most of us didn’t have experience with systems that had lots of CPUs. GPUs are optimized to suit common workloads and, quite predictably, aren’t optimized to suit uncommon workloads. Modern graphics uses a mix of CPU and GPU work. The back-and-forth, especially returning results from the GPU to the CPU, has historically caused all sorts performance problems. We don’t have the explanation for why here, but I see no reason to conclude that it must be poorly written software just because the performance characteristics are unintuitive.
- terafo 5y ago
  > GPUs are optimized to suit common workloads and, quite predictably, aren’t optimized to suit uncommon workloads. That doesn't explain why there are such drastic performance differences between GPUs on same uarch. It shouldn't perform better on clearly worse GPUs. > We don’t have the explanation for why here, but I see no reason to conclude that it must be poorly written software just because the performance characteristics are unintuitive. I'm not saying that it `must` be poorly written software. I'm just saying that it's `likely`, given niche nature of this software. If performance is affected by some unusual factor they have to clearly indicate that.
_kbh_ 5y ago
Could the transfer speed over PCIE to the video card be the limiting factor then?. It would make sense why the M1 GPU bench very high as well as it likely has 200GB/s read to the shared memory compared to the ~32GB/s given by a PCIE4.0 16x slot.
- terafo 5y ago
  Mac Pro with w6900x uses PCIe gen 3, which is 2x worse than gen 4. And given other results in that thread, I think it's unlikely to be the bottleneck. If it was, people with ryzen CPUs and 30 series cards would have much higher performance than they have on practice.
  - _kbh_ 5y ago
    That is true, perhaps its a difference in drivers and operating systems?. it hard to tell if the benchmark is bad without having the same setup to test against because anything could impact it really.

by terafo 5y ago

klodolph 5y ago

“It’s likely just poorly written software” sounds completely glib to me, it sounds like an intuitive leap rather than a reasoned-out conclusion. Again, it sounds like an unusual workload. I don’t see why performance must scale with some spec in any particular straightforward way—I think we all remembered seeing workloads that would get faster once you added more CPU cores, then slower once you kept adding CPU cores—because back in the day, most of us didn’t have experience with systems that had lots of CPUs. GPUs are optimized to suit common workloads and, quite predictably, aren’t optimized to suit uncommon workloads. Modern graphics uses a mix of CPU and GPU work. The back-and-forth, especially returning results from the GPU to the CPU, has historically caused all sorts performance problems. We don’t have the explanation for why here, but I see no reason to conclude that it must be poorly written software just because the performance characteristics are unintuitive.

terafo 5y ago
> GPUs are optimized to suit common workloads and, quite predictably, aren’t optimized to suit uncommon workloads. That doesn't explain why there are such drastic performance differences between GPUs on same uarch. It shouldn't perform better on clearly worse GPUs. > We don’t have the explanation for why here, but I see no reason to conclude that it must be poorly written software just because the performance characteristics are unintuitive. I'm not saying that it `must` be poorly written software. I'm just saying that it's `likely`, given niche nature of this software. If performance is affected by some unusual factor they have to clearly indicate that.

_kbh_ 5y ago

Could the transfer speed over PCIE to the video card be the limiting factor then?. It would make sense why the M1 GPU bench very high as well as it likely has 200GB/s read to the shared memory compared to the ~32GB/s given by a PCIE4.0 16x slot.

terafo 5y ago
Mac Pro with w6900x uses PCIe gen 3, which is 2x worse than gen 4. And given other results in that thread, I think it's unlikely to be the bottleneck. If it was, people with ryzen CPUs and 30 series cards would have much higher performance than they have on practice.
- _kbh_ 5y ago
  That is true, perhaps its a difference in drivers and operating systems?. it hard to tell if the benchmark is bad without having the same setup to test against because anything could impact it really.