During its big Tuesday Keynote at GTC 2016, Nvidia announced the Tesla P100 GPU. Based on the new Pascal architecture, which uses 16nm FinFET technology, it packs in 15.3 billion transistors, uses HBM2 memory, and steps up performance with some vigor. The chip is 600 mm square and connects to the memory with 4,000 wires (Maxwell has 384).
It also includes NVLink, a new interconnect capable of 160 GB/s. NVLink is what Nvidia will use to connect the graphics card to the system instead of PCI-Express, which would not offer enough bandwidth. Nvidia claimed a 5x performance improvement with its new interconnect.
The Tesla P100 has 16 GB of CoWoS (Chip on Wafer on Substrate) 3D-stacked HBM2 memory, with an enormous 720 GB/s memory bandwidth. It is a new unified memory architecture. The chip has 14 MB of shared memory register files with an aggregate throughput of 80 TB/s, along with 4 MB of L2 cache.
At FP64, the Tesla P100 pulls off 5.3 TF of performance, jumping to 10.6 FP at FP32. At FP16, that number becomes a staggering 21.2 TF.
The Tesla P100 is in volume production today and will ship “soon." Jen-Hsun Huang said that really means it will show up in the cloud first, and ship via OEMs in Q1 2017.
Of course, as a consumer, you can’t have this chip yet. That’s no surprise. In the past, Nvidia also first implemented its new architectures and GPUs in scientific super computing before bringing it to the consumer market. Nvidia did not announce when we would be seeing a consumer-oriented Pascal chip.