Just half a year after the presentation of the computing accelerator Nvidia A 90 with Ampere architecture, Nvidia is launching a drilled-out version. At A 100 80 GB says it all: It’s the same GPU on the same SXM4 module, only Nvidia has screwed on the memory and changed it to fat 80 GByte doubled.

The A 100 80 GB should be in the first quarter 2021 will be available as FRU upgrade kits to upgrade existing installations. Nvidia expects first deliveries in January, larger quantities from February. The A 100 will still not be available individually in SXM4 format – only in the HGX baseboards, such as those in the DXG station are installed. There four A 100 can then be transferred to a pool of 320 GByte memory access.

First copies of DGX A 100 640 GB have already been delivered to partners like the University of Florida.

HBM2E instead of 6144 Bit Nvidia is clumsy when expanding the memory. Instead of unlocking the sixth memory controller and thus also the sixth HBM chip, Nvidia uses HBM2E chips with double the capacity of 16 GByte. Five of them together ensure a transfer rate of over 2 TByte / s – more precisely 2, 036 TByte / s. Even if Nvidia does not officially state any clock rates, they should be clocked with 1590 MHz and consequently just under 32 Work percent faster than on Nvidia’s “old” A 100 with 40 GByte.

The accelerated and enlarged memory gives the A 100 80 GB according to Nvidia a significant performance boost in some applications. Huge data sets that previously did not fit in local storage benefited most. This is the case, for example, when training an AI-controlled recommendation system (DLRM, Huge CTR Framework with a 450 GByte data set) now up to 2.6x as fast (+ 160%). Big data analysis on a TByte data set also runs significantly faster – factor Nvidia names 1.9 (+ 90%) here. HPC benchmarks such as Quantum Espresso, which at least also uses 1.6 Tbytes of data, ran 1.8x (+ 40%) so fast.

For the AI ​​inferencing, Nvidia is using 25 Percentage power out – on the speech recognition system RNN-T, in which 7 MIG instances à 10 GByte run in parallel, which is good in the memory of the A 100 80 GB fit. Nvidia claims to have increased energy efficiency to over 25 GFlops / watt.

Fast 400 G-Infiniband The supercomputer manufacturers should also be happy: Nvidia subsidiary Mellanox is bringing 400 Gbit / s a ​​new speed level for even faster data exchange between individual nodes. The cards should work with passive copper cables up to 1.5 meters in length and thus be easily usable within a rack. For longer distances, however, active copper cables or fiber optic cables are required. Atos, Dell, Fujitsu, Inspur, Lenovo and Supermicro are said to have already expressed interest. (csp)