NVIDIA sells many data center parts under similar names. Most of the confusion comes from three things: the generation (Hopper or Blackwell), the packaging (a single GPU, a CPU plus GPU superchip, or a whole rack) and the market (global or export-restricted). This page sorts them. Prices are not quoted here. See the GPU pricing chart for live on-demand hours.
Hopper generation
- H100: the default. 80 GB of HBM3 on SXM, about 3.35 TB/s of memory bandwidth and NVLink 4 at 900 GB/s. Check the H100 price per hour.
- H800: a Hopper variant made for the China market, with reduced NVLink bandwidth compared with H100. Same compute family, weaker multi-GPU scaling. It is rarely listed in global on-demand catalogs.
- H200: the same Hopper compute with 141 GB of faster HBM3e at about 4.8 TB/s. Pick it when memory is the wall, not compute. See the H200 price per hour.
- GH200 Grace Hopper: a Hopper GPU and a 72-core Grace CPU on one module, linked by 900 GB/s NVLink-C2C. It pairs GPU memory with a large pool of LPDDR5X system memory, which suits memory-hungry inference and HPC more than plain training.
Blackwell generation
- B100: the first Blackwell part, positioned as a roughly 700 W drop-in for Hopper-class HGX platforms. Most catalogs went straight to B200 instead, so you will see B100 mostly in spec sheets.
- B200: the main Blackwell GPU, with about 180 GB of HBM3e in the HGX B200, FP4 support and NVLink 5 at 1.8 TB/s. Check the B200 price per hour.
- B300 (Blackwell Ultra): 288 GB of HBM3e per GPU and a stronger FP4 story. It is the step up when 180 GB is not enough. See the B300 price per hour.
- GB200 and GB200 NVL72: two B200 GPUs plus a Grace CPU on a superchip, and 36 of them in a liquid-cooled rack. Read GB200 NVL72 specs, price and how it differs from B200.
Which ones you can rent by the hour
On-demand hourly rows exist mainly for H100, H200, B200 and B300. H800 is region-restricted. GH200 shows up at a few providers. B100 and GB200 are mostly sold as quotes or reserved clusters. We only rank published on-demand hours, so those rows stay out of the price table. See how we rank an hour.
How to choose
- Default training and fine-tuning: H100.
- Model or KV cache does not fit in 80 GB: H200, then B200 or B300 for more.
- Very large serving with huge context and cheap system memory: look at GH200.
- FP4 inference or trillion-parameter scale: B200, B300, or reserved GB200.
- Not sure it fits? Run it through the VRAM fit tool.
For pairwise detail, see H100 vs H200, B200 vs H200 and the longer H100 vs H200 vs B200 guide.
FAQ: NVIDIA data center GPUs
What is the difference between H100 and H800?
Both are Hopper. H800 is the export-restricted version with lower NVLink bandwidth, so multi-GPU scaling is weaker.
Is GH200 the same as H100?
The GPU is Hopper-class, but GH200 adds a Grace CPU and a large LPDDR5X memory pool on the same module. It is a different product shape.
Is B100 better than H100?
Yes on paper, as a newer Blackwell part. In practice most clouds skip B100 and offer B200, so check live availability.
Which NVIDIA GPU should I rent for LLM training?
H100 is the default. Move to H200, B200 or B300 only when memory or FP4 throughput is the actual bottleneck.
Some outbound links in these articles may be affiliate links. Rankings on GPUBeacon stay independent.