Today we come across the first real photo of a Nvidia H100 in its SXM format, which is distinguished by the use of a graphics chip Nvidia GH100 based on the Hopper architecture manufactured by TSMC using the manufacturing process of 4nm.
For the rest there is nothing new, so we simply remind you of the specifications of this GPU that we could see last week on sale in Japan. at a price close to 35,000 euros.
For this price you buy a GPU that integrates no less than 80 billion transistors thanks to the use of the most advanced manufacturing process of the moment, and it is none other than the 4nm from TSMC. your graphics chip Nvidia GH100which maintains the monolithic design, offers a configuration of 18432 CUDA cores accompanied by none other than 80 GB of HBM3 memorycapable of offering a bandwidth of 3 TB/sit’s a 50 percent more compared to the Nvidia A100.
In terms of performance, we are talking about 60 TFLOPs FP64/FP321,000 TFLOP TF32 and 2,000 TFLOP FP16, implying triple the performance of the Nvidia A100. Regarding the performance of the FP8, multiply it x6 after reaching 4,000 TFLOPs. Due to this increase in power, the interface is also used PCI Express 5.0 to achieve a bandwidth of 128 GB/s (vs. 64 GB/s) and the 4th generation of NVLink interconnect technology which offers a bandwidth of 900 GB/s (+50%).
All this has a high energy cost, since we are talking about a TDP of 700W, a huge leap from the 400 W consumed by the top-end Nvidia A100 with 80 GB of HBM2e memory. That is, the power consumption grows by 75% in exchange for offering a performance improvement of up to 600% and while the bandwidth increases by 50%.