[The Exiled]

nVidia launches Hopper H100: World's largest AI and Compute accelerator with 30x A100 performance

nVidia's new H100 is fabricated on TSMC's 4N process (an updated version of TSMC’s N5 technology that offers better power/performance characteristics and a very modest improvement in density. So even at just two years after Ampere, nVidia is making a full node jump and then some for GH100), and the monolithic design contains some 80 billion transistors. To put that number in scale, GA100 is "just" 54 billion, and the GA102 GPU in the RTX 3090 is a measly 28.3 billion. Despite being such a massive processor, and despite being fabricated on the absolute bleeding edge of microprocessor process technology, the TDP for the SXM (mezzanine) version of the chip is 700 Watts. The move to HBM3 memory improves the per-pin transfer rate from 2.43Gbps to 4.8Gbps, but otherwise the memory configuration is unchanged: five active stacks of HBM3 (with one dummy) giving a total capacity of 80GB per GPU and total bandwidth of 3TB/sec. The GPU's NVLink connection steps up to the 4th generation, too, resulting in 900GB/sec bandwidth between the GPU and its host system. For products that don't use NVLink, H100 also implements PCIe 5.0.

Izvor: HotHardware i AnandTech