NVIDIA’s Volta architecture
NVIDIA (NVDA) spends a lot of time and money on developing its core architecture, which it then leverages across various platforms. NVIDIA CEO Jensen Huang said that the company spent $3 billion over three years to develop its next-generation Volta architecture, which is specially designed for deep learning.
According to NVIDIA’s original roadmap, Volta was supposed to launch after Maxwell, but as its development took longer than anticipated, the company introduced Pascal architecture instead. The company unveiled the Volta-based V100 GPU (graphics processing unit) at the 2017 GPU Technology Conference.
What’s unique about Volta architecture?
The V100 is being built on TSMC’s (TSM) 12-nm (nanometer) Fin-FET (fin field-effect transistor) process node rather than the originally planned 10-nm node. TSMC’s 10-nm node has been delayed due to technical issues, indicating a slowdown of Moore’s Law.
However, as process shrinkage is just one aspect of its GPU, NVIDIA hasn’t been deterred from releasing it. Advanced Micro Devices (AMD) is building its next-generation Vega GPU on the 14-nm node due to the delay of the 10nm-node.
With the V100, NVIDIA has increased the die size to 815 square millimeters, compared with 610 in the p100. Process shrinkage and a larger die size enabled it to fit 21.1 billion transistors on a single V100 chip, compared with 15.3 billion transistors in the P100.
Other features of the V100
NVIDIA’s V100 chip will feature 84 streaming multiprocessors, 5,120 CUDA cores, and 320 texture units. The chip will feature an NVLink 2 connector that can connect six GPUs with a bidirectional link bandwidth of 25 Gbps (gigabits per second).
Volta: Designed for machine learning
The V100 chip will have a new type of core designed for machine learning operations. The chip will feature 672 tensor cores, providing 120 teraflops of tensor operations, and a fourfold increase in performance compared with the P100. The V100 will perform better than Google’s (GOOG) tensor processing unit.
Moreover, the V100 GPU will have a 16-GB HBM2 (high bandwidth memory) with 1.8 GHz (gigahertz) on a 4,096-bit bus, for 900 Gbps of bandwidth. As HBM memory is expensive, it is generally used in high-end data center GPUs.
The V100’s pricing
NVIDIA plans to launch the first Volta GPU on a DGX platform in calendar 3Q17. A Volta-based DGX system featuring eight Tesla V100 GPUs will be priced at $150,000. The announcement of the V100 has generated rumors about Volta-based gaming GPUs, which we’ll look at next.