NVIDIA announces the declination PCIe of its accelerator Tesla V100. It still uses the enormous GV100, etched in 12nm FFN by TSMC and having no less than 21.1 billion transistors.
With a TDP reduced to 250W against 300W for the mezzanine version, performances are slightly revised downwards with a power announced to 7 Tflops in double precision and 14 Tflops in single precision in boost, a frequency going from 1455 to 1370 MHz approximately . However, 16 GB of HBM2 at 900 GB / s is still available.
Availability is announced for "later in the year", without further details. Nvidia points out that the US Department of Energy's V100-based Summit supercomputer, announced in 2014 and now scheduled for 2018, will have a power of 200 petaFLOPS (the initial range was 150-300
like for like, follow for follow
followed