24% improvement in single floating-point performance: NVIDIA paper exposes the next generation of GPUs

On December 15, the core research institute found an NV research paper, which mentioned that "GPU-N" seems to be an abbreviation of GPU-Next, speculating that the real body is Hopper. Parameters include 134 sets of SM (8576 CUDA), core frequency 1.4GHz, 60MB L2 cache, maximum 233GB HBM2e memory, with 3.5Gbps pin bandwidth or memory speed and 6144bit bit width, the total bandwidth can reach 6.3TB/s.

Edited by the Institute of Core Research

In terms of performance indicators, the single-precision floating-point is 24.2T and the half-precision is 779T, which is more reference-oriented single-precision than the NVIDIA A100, which is 19.5T of ampere large core. Interestingly, the increase in CUDA size is also proportional, not to mention the 59% increase in L2 and a 135% increase in memory bandwidth.

According to previous claims, Hopper is eventually expected to use an MCM multi-core package, and the dual core corresponds to 268 sets of SM+. In addition, Hopper will most likely end up in the professional market, and the protagonist of the game market will be 5nm Ada Lovelace. NVIDIA has given a clear roadmap at gtc+ that the next generation of Ampere will appear in 2022 and will wait and see.

(7831564)