NVIDIA A100 ‘Ampere’ benchmarked – “the fastest GPU ever recorded”

NVIDIA A100 tested

Jules Urbach, the CEO of OTOY (a company specializing in holographic rendering in the cloud), shared first benchmark results of the NVIDIA A100 accelerator.

The GPU is the first, and so far the only, Ampere-based graphics card (or more precisely a compute accelerator). Although NVIDIA announced the immediate availability of the A100 for DGX100 systems, we have not seen any meaningful benchmarks of the A100 yet.

The A100 features NVIDIA’s first 7nm GPU, the GA100. This GPU is equipped with 6912 CUDA cores and 40GB of HBM2 memory. This is also the first card featuring PCIe 4.0 interface or SXM4 depending on the variant.

OctaneBench is a benchmark designed to test OctaneRender performance. The Render was the first commercially available raytracer that fully utilized the GPU. This software runs exclusively on NVIDIA graphics cards as it relies on CUDA technology, thus won’t be seeing any comparison to AMD Big Navi or Arcturus-based graphics cards.

Jules Urbach:

A record breaking ?
The @NVIDIA A100 has now become the fastest GPU ever recorded on #OctaneBench: 446 OB4*
#Ampere appears to be ~43% faster than #Turing in #OctaneRender – even w/ #RTX off!
(*standard Linux OB4 benchmark, RTX off, recompiled for CUDA11, ref. 980=102 OB)

NVIDIA A100 in OctaneBench, Source: Jules Urbach

The A100 scored 446 points. We are not sure which result is being compared to A100, but the fastest Turing-based graphics card in OctaneBench is GRID RTX 8000, which scored 328 points. The Volta-based Tesla V100, TITAN V, and Quadro GV100 are still holding up quite well to Ampere, showing 33 to 11% performance loss compared to A100.

Single-GPU benchmarks in OctaneBench, Source: Otoy

NVIDIA Compute Accelerator Series (Formely Tesla)
VideoCardz.com	A100 PCIe	A100 SXM	Tesla V100s	Tesla V100	Tesla P100
Picture
GPU	7nm GA100	7nm GA100	12nm GV100	12nm GV100	16nm GP100
Die Size	826 mm^2	826 mm^2	815 mm^2	815 mm^2	610 mm^2
Transistors	54 billion	54 billion	21.1 billion	21.1 billion	15.3 billion
SMs	108	108	80	80	56
CUDA Cores	6912	6912	5120	5120	3840
Tensor Cores	432	432	640	640	NA
FP16 Compute	78 TFLOPS	78 TFLOPS	32.8 TFLOPS	31.4 TFLOPS	21.2 TFLOPS
FP32 Compute	19.5 TFLOPS	19.5 TFLOPS	16.4 TFLOPS	15.7 TFLOPS	10.6 TFLOPS
FP64 Compute	9.7 TFLOPS	9.7 TFLOPS	8.2 TFLOPS	7.8 TFLOPS	5.3 TFLOPS
Boost Clock	~1410MHz	~1410MHz	~1601 MHz	~1533 MHz	~1480MHz
Bandwidth	1555 GB/s	1555 GB/s	1134 GB/s	900 GB/s	721 GB/s
Eff. Memory Clock	2430 MHz	2430 MHz	2214 MHz	1760 MHz	1408 MHz
Memory Config.	40GB HBM2e	40GB HBM2e	32GB HBM2	16GB / 32GB HBM2	16GB HBM2
Memory Bus	5120-bit	5120-bit	4096-bit	4096-bit	4096-bit
TDP	250W	400W	250W	300W	300W
Form Factor	PCIe 4.0	SXM4	PCIe 3.0	SXM2 / PCIe 3.0	SXM

Source: Jules Urbach