News Nvidia Ampere A100 Takes Fastest GPU Crown in First Benchmark Result

[Admin]

The Ampere A100 has taken the Nvidia Ampere A100 Takes Fastest GPU Crown in First Benchmark Result as the fastest GPU ever benchmarked on OctaneBench.

Nvidia Ampere A100 Takes Fastest GPU Crown in First Benchmark Result : Read more

 

[bit_user]

11.2% faster than the Titan V is a pretty bad result, considering their specs.

Hopefully, the disappointing improvement is simply due to some scaling issues with the benchmark.

 

[DoofusOfDeath]

The GA100 silicon measures 826 millimeters-squared ...

That can't be right. In Imperial units, that's saying each GA100 chip is a (2.7 feet) x (2.7 feet) square.

 

[escksu]

The Ampere A100 has taken the Nvidia Ampere A100 Takes Fastest GPU Crown in First Benchmark Result as the fastest GPU ever benchmarked on OctaneBench.

Nvidia Ampere A100 Takes Fastest GPU Crown in First Benchmark Result : Read more

11.2% faster than the Titan V is a pretty bad result, considering their specs.

Hopefully, the disappointing improvement is simply due to some scaling issues with the benchmark.

We have to consider the fact that A100 is designed as a compute card for specialised applications such as AI. Its not exactly a gpu, not a quadro.

 

[みんぱ]

That can't be right. In Imperial units, that's saying each GA100 chip is a (2.7 feet) x (2.7 feet) square.

It's more like (1.13 in) by (1.13 in). One inch squared equals 645.16 millimeters squared. One foot squared equals 92,903.04 millimeters squared. Your example of (2.7 ft) by (2.7 ft) is larger than 677,000 millimeters squared (which is 820 times GA100's die area as stated in the article). Be educated.

 

[bit_user]

We have to consider the fact that A100 is designed as a compute card for specialised applications such as AI. Its not exactly a gpu, not a quadro.

I'm not familiar with the benchmark, but I got the impression it merely uses CUDA for computation. In that sense, the specs should be entirely predictive of the resultant performance.

The Titan V has 653 GB/sec of memory bandwidth and 14.9 TFLOPS of fp32 compute (at boost clocks).
The A100 has 1555 GB/sec of memory bandwidth and 19.5 TFLOPS of fp32 compute (at ??? clocks).

So, if the workload is compute-limited, then it should be at least 30.9% faster (possibly more, depending on whether the A100 figures I saw were base or boost). However, if it's limited by memory bandwidth, then it should be 138% faster! Either way, 11.2% is really disappointing. I think it likely points to bottlenecks in the software, but we don't really know. It's definitely not good news - that's for sure!

 

[DonGato]

That can't be right. In Imperial units, that's saying each GA100 chip is a (2.7 feet) x (2.7 feet) square.

Thats some silly math take the square root of that to get the dimensions so 28mmX28mm die size. so now when you see a square mm or cubic mm you take the square root or cube root on a calculator to see what each side is.

 

[Camikazi]

I get that this will almost definitely be faster than the competing Radeon but come on now, Fastest GPU Crown in a benchmark that only benchmarks Nvidia GPUs...

 

[astidcrisse]

That can't be right. In Imperial units, that's saying each GA100 chip is a (2.7 feet) x (2.7 feet) square.

According to your math, that would make a 677 329mm square chip

I think you're missing what "square" means

 

[bit_user]

I get that this will almost definitely be faster than the competing Radeon but come on now, Fastest GPU Crown in a benchmark that only benchmarks Nvidia GPUs...

I believe the main point of the article was to infer how much faster the gaming variants of Ampere GPUs could be than their Turing counterparts. However, the analysis is clearly problematic, for a number of reasons.