NVIDIA Volta Tesla V100 Cards Detailed – 150W Single-Slot & 300W Dual-Slot GV100 Powered PCIe Accelerators
NVIDIA announced today two next generation cards based on its Volta graphics architecture and GV100 GPU. The new Tesla V100 accelerators will come in two different PCIe form factors, a 150W single-slot full height, half length design and a standard 300W dual-slot design. Both designs will house NVIDIA's next generation GV100 GPU featuring 5120 Volta CUDA cores with 16GB of HBM2.
NVIDIA Telsa V100 Accelerator - 150W Single-Slot and 300W Dual-Slot PCIe Cards
The GV100 Volta GPU that sits at the heart of each of these upcoming Tesla accelerators is a massive 815mm² chip with over 21 billion transistors built on TSMC's new 12nm FinFET manufacturing process. At 1455MHz the Tesla V100 delivers 15 TFLOPS of single precision compute and 7.5 TFLOPS of double precision compute at 300W. It's worthy of note that just like the P100, the V100 does not feature a fully unlocked GPU. The GV100 GPU houses 5376 CUDA cores but only 5120 are functional in the Tesla V100.
|Tesla Product||Tesla K40||Tesla M40||Tesla P100||Tesla V100|
|GPU||GK110 (Kepler)||GM200 (Maxwell)||GP100 (Pascal)||GV100 (Volta)|
|FP32 Cores / SM||192||128||64||64|
|FP32 Cores / GPU||2880||3072||3584||5120|
|FP64 Cores / SM||64||4||32||32|
|FP64 Cores / GPU||960||96||1792||2560|
|Tensor Cores / SM||NA||NA||NA||8|
|Tensor Cores / GPU||NA||NA||NA||640|
|GPU Boost Clock||810/875 MHz||1114 MHz||1480 MHz||1455 MHz|
|Peak FP32 TFLOP/s*||5.04||6.8||10.6||15|
|Peak FP64 TFLOP/s*||1.68||2.1||5.3||7.5|
|Peak Tensor Core TFLOP/s*||NA||NA||NA||120|
|Memory Interface||384-bit GDDR5||384-bit GDDR5||4096-bit HBM2||4096-bit HBM2|
|Memory Size||Up to 12 GB||Up to 24 GB||16 GB||16 GB|
|L2 Cache Size||1536 KB||3072 KB||4096 KB||6144 KB|
|Shared Memory Size / SM||16 KB/32 KB/48 KB||96 KB||64 KB||Configurable up to 96 KB|
|Register File Size / SM||256 KB||256 KB||256 KB||256KB|
|Register File Size / GPU||3840 KB||6144 KB||14336 KB||20480 KB|
|TDP||235 Watts||250 Watts||300 Watts||300 Watts|
|Transistors||7.1 billion||8 billion||15.3 billion||21.1 billion|
|GPU Die Size||551 mm²||601 mm²||610 mm²||815 mm²|
|Manufacturing Process||28 nm||28 nm||16 nm FinFET+||12 nm FFN|
For hyperscale datacenters NVIDIA has managed to cram that same 815mm² GV100 GPU into a card the size of a CD case. At half the power the 150W hyperscale Tesla V100 naturally won't be as fast as its 300W bigger brother but it's close. How close? NVIDIA isn't disclosing that information just yet.
NVIDIA's Volta Architecture & The GV100 GPU
NVIDIA's new Volta architecture manages to deliver 40% better performance/watt compared to Pascal and houses 7% more CUDA cores/mm² and 6% better performance/mm². This is thanks to a combination of the more efficient and higher density 12nm FinFET process as well as due to architectural refinements of the original Pascal architecture.
Each Volta SM -- Streaming Multiprocessor -- still houses 64 CUDA cores just like Pascal. However, volta features a slightly different SM partitioning. While in Pascal each SM was partitioned into two blocks, in Volta each SM is partitioned into four blocks. each with 16 FP32 cores, 8 FP64 cores, 16 INT32 cores and two brand new cores called Tensor cores.
This is another area where GV100 differs from GP100. Each Volta GV100 SM includes separate FP32 and INT32 cores which can simultaneously execute FP32 and INT32 operations at full throughput. Whilst GP100 only featured FP32 cores which were capable of executing either FP32 or INT32 operations at any given time.
Tensor cores are mixed precision FP32/FP16 4x4 arrays. Each array is able to accelerate the execution of what NVIDIA calls Tensor operations by a factor of 6 compared to traditional FP64 cores. This allows Volta to deliver 6x higher inferencing throughput per clock compared to Pascal and 12x the deep-learning throughput per clock.
The key architectural improvements from Pascal to Volta include :
- 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.