Comments on: AMD Radeon and NVIDIA GeForce FP32/FP64 GFLOPS Table

By: Roger Dahl

Roger Dahl — Sun, 04 May 2014 02:59:15 +0000

This is awesome. Thank you.

It’s necessary to take these numbers with a grain of salt, and it’s difficult to estimate just how much salt is needed for any given architecture.

I think a copy of the table that is sorted on FP64 would be useful as well.

By: komar

komar — Sun, 09 Mar 2014 11:44:13 +0000

gtx 690@ ES BIOS 450W PT150%- +150 mhz + 605 mhz gddr5

baseclock 1065 mhz – boost 1215 mhz – 7.2ghz GDDR5

AIda : 6.498 Tflops/s FP32 / 0.275 Tflops/s FP64

By: Promilus

Promilus — Sat, 08 Mar 2014 07:41:27 +0000

@DrBalthar – sorry but I disagree. Test results of full FMA coverage on all ALUs would be the very same artificial as those of “theoretical peak performance” – it just doesn’t translate into real-world performance. Some apps uses more data parallelism, some more task parallelism, some uses both float and integer intensively and other are focused on fp64 only. Some shares a lot of calculation with CPU and some runs solely on GPU. FP MADD/FMA test results would mean absolutely nothing. For the very same reason there’s no “actual peak performance” benchmark for x86-64 CPUs. But you do have some peek of a real world performance with Linpack benchmark. Check top500.org, there’re both Rmax (linpack results) and Rpeak (theoretical performance) numbers. In mixed GPU and CPU clusters it’s ratio is close to 60-70%. On CPU only clusters it reaches 80-95%.

By: DrBalthar

DrBalthar — Fri, 07 Mar 2014 12:57:13 +0000

The F in FLOP stands for Floating point so integer and bit operation are irrelevant. Using just FMA, FPNADD still would be the most fair test as it is the only operation used so there wouldn’t be any difference or cheating. Actual peak performance doesn’t mean real life application performance. It just means actually validated numbers and not just marketing material.

By: Promilus

Promilus — Thu, 06 Mar 2014 19:03:34 +0000

“Well not really you can write very simple code that just does a MAD operation (as that’s the one the usually use for advertising FLOPs”)
No, it was FPMADD but since all new GPUs uses FMA which does basically the very same thing but with more accurate final result it’s just as good. And btw it wouldn’t be best possible benchmark and “peak” performance since many apps uses only certain type of calculations. Bitcoin miners iirc used plenty of integer and bit operations which were much higher on AMD GPUs than NVidia’s – define universal “actual peak performance” then? There’s no such thing.

By: DrBalthar

DrBalthar — Thu, 06 Mar 2014 13:44:00 +0000

Well not really you can write very simple code that just does a MAD operation (as that’s the one the usually use for advertising FLOPs).

By: Promilus

Promilus — Wed, 05 Mar 2014 15:31:59 +0000

@DrBalthar – define “actual peak performance”
I believe you are referring to typical/average real life performance but that one depends strongly by app you test it with.

By: DrBalthar

DrBalthar — Wed, 05 Mar 2014 12:49:09 +0000

So basically theoretical peak performance instead of actual peak performance

By: JeGX

JeGX — Wed, 05 Mar 2014 12:25:14 +0000

The values don’t come from a benchmark tool, it’s just a compilation from articles / reviews.

By: nuninho1980

nuninho1980 — Wed, 05 Mar 2014 12:18:33 +0000

What’s the FP32/FP64 benchmark tool?? Do you use AIDA64?