Nvidia announced H100 SXM, its flagship 4nm GPU with 80GB HBM3 memory for datacenter workloads.

NVIDIA's next-gen Hopper H100 GPU in SXM form factor has been spotted in the flesh for the first time. Until now, only renders of the H100 GPU had been publicly available, making this the first real pictures of the device. The next-gen Hopper H100 SXM GPU was photographed by Patrick Kennedy from ServeTheHome, with a board model of "PG520" and a 4nm H100 GPU die utilizing TSMC's very latest CoWoS packaging technology.

The H100 GPU features an 814mm² die with 6 memory modules around it, totaling 80GB of ultra-fast HBM3 memory. NVIDIA's new H100 GPU comes in two forms with different CUDA core counts: the SXM variant features 16896 FP32 CUDA cores, while the PCIe-based model provides 14592 cores. The unit pictured in these images represents the SXM variant with the full 16896 FP32 CUDA cores and 80GB of HBM3 memory.

Power consumption represents a significant increase from previous GPU architectures. The H100 operates at up to 700W, compared to the 400-450W range of the previous Ampere and Volta GPU architectures in data center form, representing an increase of up to 250-300W in power consumption.