Rent NVIDIA GH200 Superchip in the Cloud

NVIDIA GH200 Superchip is built on the Hopper architecture and ships with 96 GB of HBM3 memory at 4,000 GB/s bandwidth. Released in 2023, the card delivers 989 FP16 TFLOPS and 494.5 FP32 TFLOPS at a 700W TDP.

For machine learning researchers, those numbers translate into several practical limits: the VRAM ceiling dictates the largest large language model weights you can load (and the maximum batch size at a given sequence length), while memory bandwidth sets the upper bound for attention-heavy inference. Compute throughput matters most for dense matrix multiplications — pre-training, large-batch pre-training, and diffusion.

See the NVIDIA GH200 Superchip page for the full spec sheet and comparisons to related GPUs.

The short answer: NVIDIA GH200 Superchip runs at 989 FP16 TFLOPS with 4,000 GB/s of memory bandwidth. The longer answer depends on what you run.

For dense FP16 training with large batches, NVIDIA GH200 Superchip saturates tensor cores and delivers throughput close to peak FLOPS. For memory-bound serving on long-context foundation models, bandwidth dominates — the 4,000 GB/s figure matters more than headline TFLOPS. For scientific computing, FP32 at 494.5 TFLOPS is the relevant number and puts NVIDIA GH200 Superchip in line with the HPC expectations of its Hopper class.

Check the NVIDIA GH200 Superchip page for complete specifications and related GPU matchups.

Use cases where NVIDIA GH200 Superchip performs well: Large-scale AI training, HPC. The card's 96 GB of VRAM, Hopper tensor cores, and data-center positioning suit it for AI teams who need cloud-grade hardware without the premium of the newest accelerators.

A good rule of thumb: if your model fits in 96 GB of VRAM at your target precision, and your workload is bandwidth- or compute-bound rather than memory-bound, NVIDIA GH200 Superchip is likely a better economic choice than upgrading to a higher tier.

Review full specs and related comparisons on the NVIDIA GH200 Superchip page.