AMD CDNA Discussion Thread

xpea said:
As expected, it's not looking good...

https://twitter.com/x/status/1457903334199726083
Click to expand...

only 2.8x the competition ,
63f.jpg
 
Why does it matter that a single purpose product delivers more performance when the market is limited? Outside of goverments MI200 will be not used in any meaningfull numbers. Nextplatform said that China broke the sustained exaflops performance barrier within 35 MW. I dont see how Frontier should be on par with it when one OAM card has only 44% practical performance. That puts Frontier in the range of ~700 Petaflops at <=33MW.
 
Hopper and MI300 will be direct competitors launching at about the same time and on the same process node? Both being MCM?
 
MI300 is going to El Cap. So it should be ready by Q4 2022.
 
trinibwoy said:
I believe those were measured results from the HPL benchmark so not just paper. Now we can argue that HPL is not representative of anything but it's still the industry standard benchmark.
Click to expand...
AMD makes numerous comparisons to Nvidia's A100, claiming significant increases in computer performance and density. As always, take these announcements with a grain of salt as paper specs don't tell the whole story, but MI200 looks to be an absolute monster.
...
MI200 also adds FP64 matrix support, with a peak rate that's double the vector unit rate: 95.7 TFLOPS. Again, by way of comparison, the Nvidia A100 FP64 vector performance is 19.5 TFLOPS. That's on paper, of course, so we need to see how that translates into the real world. AMD claims performance is around three times as fast as the A100 in several workloads, though it's difficult to say if that will be the case across all workloads.
...
On the FP16 side of things, the performance isn't quite as high. Nvidia's A100 has 312 TFLOPS of FP16/BF16 compute, compared to 383 TFLOPS for the MI200, but Nvidia also has sparsity. Basically, sparsity allows the GPU to skip some operations, specifically multiplication by zero (which, so my math teacher taught me, is always zero). Sparsity can potentially double the compute performance of the A100, so there should be some use cases where Nvidia maintains the lead.
Click to expand...
AMD Instinct MI200: Dual-GPU Chiplets and 96 TFLOPS FP64 | Tom's Hardware (tomshardware.com)
 
troyan said:
That puts Frontier in the range of ~700 Petaflops at <=33MW.
Click to expand...
No, according to ORLN's lead engineer David Grant the system is dimensed for 26-28 MW.

https://www.exascaleproject.org/fro...n-and-construction-of-the-mechanical-systems/

Why are some forum members persistently fabricating inferior numbers (like 50 % lower bandwidth than the actual one, ~25 % higher power consumption than the actual one, purposely confusing theoretical and measured performance etc.) and making disastrous speculations based on the biased data? Some posts here sound more like a content created by competitor's marketing team than technological discussion. Everyone can make a mistake, but these "mistakes" are repeatedly CDNA-related and come from people, who are never mistaken when quoting Ampere numbers.
 
trinibwoy said:
I believe those were measured results from the HPL benchmark so not just paper. Now we can argue that HPL is not representative of anything but it's still the industry standard benchmark.
Click to expand...
It's only standard because it goes back 40 years. HPCG is considered a much more representative benchmark for modern workloads - ones that care much more about memory and cache than simple DPFP throughput. I guess AMD must've forgotten it existed though since they didn't release any figures for it. And obviously they have no reason to hide its HPCG performance as certain forum members who deride recent Chinese efforts as linpack machines are assuring us here it's good for real science™.
 
Qesa said:
It's only standard because it goes back 40 years. HPCG is considered a much more representative benchmark for modern workloads - ones that care much more about memory and cache than simple DPFP throughput. I guess AMD must've forgotten it existed though since they didn't release any figures for it. And obviously they have no reason to hide its HPCG performance as certain forum members who deride recent Chinese efforts as linpack machines are assuring us here it's good for real science™.
Click to expand...

It's true. HPCG benchmarks on MI200 would be interesting.
 
You must log in or register to reply here.
Back
Top