NVIDIA used its 2016 GPU Technology Conference this week to show off its new Pascal GP100 GPU and Tesla P100 graphics accelerator, which NVIDIA says delivers the highest absolute performance for high-performance computing, technical computing, deep learning, and other computationally-intensive data center workloads.
The Pascal architecture has five technological breakthroughs, NVIDIA notes:
- It has 150 billion transistors built on 16nm FinFET fabrication technology, making it the world’s largest FinFET chip ever built.
- It can deliver more than 5 teraflops of double-precision performance; for deep learning, NVIDIA notes, a Pascal-powered system offers a 12 times leap in neural network training performance compared to current-generation GPU architecture.
- The NVIDIA NVLink high-speed bidirectional interconnect scales applications across multiple GPUs, for five times greater acceleration in interconnect bandwidth compared to today’s best solution.
- CoWoS, or Chip-on-Wafer-on-Substrate, with HBM2 offers a unique approach to memory design for a 3x boost in memory bandwidth performance compared to the NVIDIA Maxwell architecture.
- New artificial intelligence algorithms take advantage of half-precision instructions for more than 21 teraflops of peak performance.