Nvidia announces the Ampere A100 GPU, establishing a new performance record for the fastest GPU in production.

NVIDIA's flagship Ampere GPU, the A100, has set a new performance record as the fastest GPU ever recorded on OctaBench, the benchmark tool developed by OTOY that evaluates GPU performance using the Octane Renderer. The achievement was shared by Jules Urbach, CEO of OTOY, on July 23, 2020. The A100 represents the world's largest graphics chip based on the 7nm process node, having been unveiled in May with specifications and performance metrics that continue to break world records.

In the OctaBench benchmark, the NVIDIA A100 Tensor Core GPU posted a score of 446, demonstrating significant performance advantages over competing architectures. According to Jules Urbach, the Ampere A100 is approximately 43% faster than the Turing GPU in OctaneRender, notably achieving this result even with RTX off. The benchmark utilized the standard Linux OB4 benchmark with RTX disabled and was recompiled for CUDA11, with a reference baseline of 980=102 OB.

When compared to other high-performance GPUs, the A100's dominance becomes clear. The Tesla V100, the A100's predecessor, is approximately 20% slower on average, while the Titan V is 11% slower—a relatively narrow gap that may reflect the Titan V's use of the same GV100 GPU as the Tesla V100, which could be more optimized for datacenter and cloud-scale workloads. The Titan RTX shows a more substantial performance gap, registering 38% slower than the A100. The comparison highlights that Turing GPUs, typically optimized for gaming and GP-GPU applications, perform differently on this specific workload than the datacenter-focused architecture of Ampere.

The A100's impressive performance is underpinned by its massive scale and advanced design. At 54 billion transistors packed within a single die, the A100 is by far the largest 7nm chip produced to date. The current A100 configuration features a vastly cut-down design due to early yields, but as manufacturing improves, higher bin versions with more cores could emerge, further increasing performance in this benchmark. The potential for enhanced performance is significant, as these gains were achieved without RTX acceleration enabled in the test environment.

The broader implications of the A100's achievement extend to NVIDIA's consumer-focused Ampere lineup. As the market awaits the launch of consumer graphics cards based on Ampere architecture, the performance demonstrated in OctaBench suggests that GeForce RTX 30 series cards could come close to the performance of their HPC counterparts once RTX acceleration is enabled. The A100's dominance in this specific benchmark provides strong indication of the substantial generational leap Ampere will deliver across diverse workloads.