Raw compute of NVIDIA Tesla V100 versus its generation peers

回答

Peak performance on NVIDIA Tesla V100: 125 FP16 TFLOPS, 15.7 FP32 TFLOPS, 900 GB/s memory bandwidth. Those figures cap theoretical throughput, but real-world performance varies based on kernel efficiency, batch size, and model shape.

For pre-training, expect near-peak utilisation on well-optimised frameworks (PyTorch with Flash Attention, DeepSpeed, Megatron-style tensor parallelism). For serving, KV-cache bandwidth is usually the bottleneck — which is why the 900 GB/s figure often predicts latency better than FP16 TFLOPS.

On ML benchmarks NVIDIA Tesla V100 lands in the tier you'd expect from its Volta generation, with strong performance-per-watt given the 16 GB VRAM capacity.

Deploy NVIDIA Tesla V100 on Vast.ai for $0.13/hr — currently the only tracked provider offering this GPU.