Rent NVIDIA GB200 Superchip in the Cloud

The full NVIDIA GB200 Superchip spec sheet reads: Blackwell generation, 384 GB of HBM3e VRAM, 16,000 GB/s memory bandwidth, 4,500 TFLOPS FP16, 150 TFLOPS FP32, 2,700W power draw, released in 2024.

Memory is typically the constraint for large-model real-time serving — at 384 GB, NVIDIA GB200 Superchip comfortably handles mid-sized transformers in FP16 and much larger models in FP8/INT8. The 16,000 GB/s figure is particularly important for KV-cache-bound autoregressive decoding, where memory bandwidth caps tokens/second more than raw compute.

Full specs, benchmarks, and comparisons are on the NVIDIA GB200 Superchip page.

NVIDIA GB200 Superchip performance headline: 4,500 FP16 TFLOPS, 150 FP32 TFLOPS, 16,000 GB/s bandwidth, 384 GB VRAM.

Converted into practical benchmarks: model training a 7B-parameter LLM in FP16 with reasonable batch sizes typically saturates compute before bandwidth; real-time serving on the same model is usually bandwidth-bound and tracks the 16,000 GB/s figure. Diffusion image generation benchmarks sit between the two — compute-heavy steps utilise tensor cores well, while attention blocks still touch bandwidth.

See the NVIDIA GB200 Superchip page for the full spec sheet and comparisons to related GPUs.

NVIDIA GB200 Superchip is best for workloads where its 384 GB VRAM and Blackwell tensor cores are well-matched: Largest-scale AI training, multi-trillion parameter models.

If your workload needs significantly more memory (e.g., training frontier-scale models from scratch), NVIDIA GB200 Superchip is undersized and you'd want an H100/H200/B200 class card. If your workload needs less (e.g., small-scale serving on 7B-parameter models), cheaper cards like L4 or RTX 4090 may be more cost-efficient. For the middle band, NVIDIA GB200 Superchip is usually the sensible pick.

Review full specs and related comparisons on the NVIDIA GB200 Superchip page.