B300, B200, H200 and H100 compared
The four NVIDIA data-center GPUs that Tera rents, by generation and by memory for each GPU and for each 8-GPU node.
Tera rents nodes with four NVIDIA GPU models: B300, B200, H200 and H100. The H100 and H200 are from the Hopper generation. The B200 and B300 are from the newer Blackwell generation. Each node has 8 GPUs.
How much memory does each GPU have?
- B300 (Blackwell Ultra): 288 GB for each GPU, 2,304 GB for a node.
- B200 (Blackwell): 180 GB for each GPU, 1,440 GB for a node.
- H200 (Hopper): 141 GB for each GPU, 1,128 GB for a node.
- H100 (Hopper): 80 GB for each GPU, 640 GB for a node.
These are the figures of the NVIDIA reference systems. Confirm the configuration of a node in the written offer.
Why does GPU memory matter?
A model and its working data must fit in GPU memory. More memory for each GPU lets you run a larger model on fewer GPUs, use larger batches in training and serve longer contexts in inference. A model that does not fit on one GPU is divided across several, which adds communication between the GPUs.
Which GPU fits which work?
- Training of large models, and inference of the largest models: memory for each GPU helps most, so look at B300 and B200.
- Fine-tuning and inference of mid-size models: H200 has the memory for many of these.
- Smaller models, and software that is already tuned for it: H100 is still in wide use.
How do you compare price across models?
Compare the cost of the work, not the cost of the GPU-hour. A GPU that costs more per hour but completes the job in fewer hours, or on fewer GPUs, can cost less in total. Run a short test of your own workload if you can. Then compare offers with the same term and prepayment on the pricing page, and read how to size a GPU cluster.
About Tera Compute. Tera Compute provides GPU rentals to startups: dedicated bare-metal nodes for AI teams. Tera also finds colocation for the GPUs that a startup owns. The company is based in the United States and is backed by Alliance. It is built by an operating team with deep experience in industrial-scale compute and data centers. Learn more at tera.gw.