GTC 2017 - NVIDIA removed the wraps off of their next-gen Volta GPU architecture today at GTC 2017, with the new Tesla V100 graphics card and DGX-1 with Tesla V100 system that costs $149,000.
We know that there are over 21 billion transistors on the GV100 GPU, offering 15 TFLOPs of single precision compute performance backed up by 7.5 TFLOPs of double precision compute. But there are two variations of Tesla V100 graphics cards: single-slot, and dual-slot. The single-slot Tesla V100 has a 150W TDP, while the dual-slot Tesla V100 ramps up the TDP to 300W.
NVIDIA hasn't confirmed what the different in performance is going to be for the 150W variant of Tesla V100, and that's a very interesting question. Performance wise, we're expecting a pretty huge 40% performance per watt advantage over the previous Pascal-based Tesla P100, on the new 12nm FinFET process.
NVIDIA notes the key architectural improvements that Volta has over Pascal as:
- New mixed-precision FP16/FP32 Tensor Cores purpose-built for deep learning matrix arithmetic;
- Enhanced L1 data cache for higher performance and lower latency;
- Streamlined instruction set for simpler decoding and reduced instruction latencies;
- Higher clocks and higher power efficiency.