Nvidia at GTC 2017 announced the most awaited news of its next generation Volta based GPU architectures by revealing Volta Tesla V100 GPU. We know it is by far the most advanced GPU Nvidia has ever created, but, what does it offer? We have the answers on what exactly Nvidia Volta Tesla V100 GPU is all about.
First of all, the new Nvidia Volta Tesla V100 GPU comes out in two different designs, a single slot and a dual slot version. Nvidia has not come up with the performance jump between single slot and dual slot version of Volta Tesla V100 GPU; however, they have listed few key differences between the two. The single slot variant has a 150W TDP, while the dual slot version has a 300W TDP. Both versions have 5120 Volta CUDA Cores based on 12nm FinFet process by TSMC. Both version of Volta Tesla V100 have 16GB of HBM2, instead of previously rumored GDDR6 technology.
Second, it is the specialized processors known as Tenser Cores. Nvidia says” Tenser Cores are the most important feature of the Volta GV100 architecture to help deliver the performance required to train large neural networks. Tensor Cores provide up to 12x higher peak TFLOPs [teraflops] on Tesla V100 for deep learning training compared to P100 FP32 operations, and for deep learning inference, up to 6x higher peak TFLOPs compared to P100 FP16 generations
Third, the performance jump between a Volta Tesla V100 GPU and comparatively Pascal P100 GPU of 300W. The difference is quite big, in terms of performance leap, design factor, type of DRAM technology and the nature of load that Volta Tesla V100 offers. According to Nvidia, Volta offers a 40% performance boost as compared to Pascal architecture.
In terms of design factor, Volta V100 GPU is huge, the chip measures 815 square millimeters while a P100 GPU based chip measured in at 610 square millimeters.
Fourth, it is the nature of workload Nvidia Volta Tesla V100 GPU offers. Currently the GPU version shown at GTC 2017 largely targets the usage of GPU in data-centers, artificial-intelligence and high-performance-computing markets. Nvidia has not shown yet the consumer friendly version of Volta Tesla V100 and we will remain skeptical as when Nvidia breaks the news for a personal computer based usage of Volta architecture GPUs.
Fifth, the hardware specification we have come to know so far on Nvidia Volta V100 GPU:
- 5,120 CUDA cores
- 640 Tensor Cores
- Comes in two different versions: Single Slot and Dual Slot version
- Single Slot has 150W TDP
- Dual Slot has 300W TDP
- Delivers 7.5 TFLOPS using 64-bit FP
- Delivers 15 TFLOPS using 32-bit FP
- Stocks a 16MB cache and 16GB HBM2 with 900GB per second bandwidth
- Chip has 6 MB of L2 cache
- Chip has a die size of 815 square millimeters
The Volta based first GPU is scheduled to arrive later this year. Meanwhile, let us know what you think about the beasty Nvidia Volta GPU, do you think it has more potential for consumer PCs or is it more inclined for research and data-center workloads? Let us know in comments section below.