At a conference at SC23 in Denver, NVIDIA revealed for the first time the performance of Blackwell B100GPU is coming in 2024 and will succeed the Hopper generation (H100 and H200).
NVIDIA showed this in a short preview in the form of a slide Blackwell B100 will be able to burn its current flagship solution H200. There is no scale for the vertical axis, but that’s easy to guess Inference performance with GPT-3 models such as GPT-3 175B (renamed Da Vinci, 175 billion parameters). more than doubled by the new architecture. Obviously, the data should be viewed as a single forecast and not as a general value.
click to enlarge
According to rumors, NVIDIA will use TSMC’s 3-nanometer production process to produce the Blackwell B100: It is also possible that this will be the company’s first MCM (multi-chip module, multiple chips on the package that work together) project will be. A second slide seems to indicate that B100 will come with HBM3E memoryperhaps faster than the one implemented on H200.
click to enlarge
Recently, the US company announced the transition from a biennial to an annual rhythm in the development of new architectures for the world of data centers.
NVIDIA: before B100, here H200 with 141 GB HBM3E memory
In August, NVIDIA announced the GH200 Grace Hopper Superchip, a version of the Grace Hopper Superchip (ARM CPU + Hopper GPU) equipped with HBM3E memory. In these hours, the company led by Jensen Huang has carried out a reboot H200, updated version with HBM3E memory of the H100 GPU-based accelerator. In addition, the storage capacity increases from 80 141GBwith a range of 4.8 TB per second thanks to 6144-bit bus.
click to enlarge
The number 141 is special, and in fact there are physically 144GB on the card due to six 24GB stacks, but NVIDIA decided to keep the margin for some reason, perhaps because of production yield.
Although NVIDIA didn’t go into details, There doesn’t seem to have been any changes to the GPU compared to the H200Therefore, the performance improvements claimed by the company must be due solely to memory. The new accelerator will be available in the second quarter of 2024.
click to enlarge
NVIDIA also revealed it HGX H200, a system with 8 H200 solutions in SMX format, Update of the already existing HGX H100. Finally, another novelty is called out Quad GH200: As the name suggests, it is a solution that brings together a total of 4 GH200 Grace Hopper Superchip solutions in parallel 288 ARM cores, 1920GB LPDDR5X memory and 384GB HBM3 memory (in this case it is not the HBM3E updated version), so a total of 2.3 TB of storage. Among the first companies to offer this solution we find: as reported hereHPE (HP Enterprise) with Cray EX2500.
