GPUs & accelerators
Datacenter GPU modules and PCIe accelerators: compute, memory, interconnect, power and procurement, stated per part.
NVIDIA H100 SXM
The H100 SXM is NVIDIA's Hopper-generation compute GPU in the socketed SXM5 form factor, sold only as a component of an HGX baseboard (4 or 8 GPU) inside partner or DGX systems, never as a standalone card. It is the workhorse part behind most of the Hopper-era training and inference fleet still in production: 80 GB of HBM3 at 3.35 TB/s, a fourth-generation Transformer Engine, and 900 GB/s of NVLink to the other GPUs on the baseboard. As of mid-2026 it is a prior-generation part: NVIDIA's shipping frontier is Blackwell Ultra (B300) and Rubin has entered production, so H100 buys now are almost always driven by compatibility with an existing Hopper fleet, software stack, or by price rather than by peak performance. It remains a rational choice when a workload fits in 80 GB and the buyer can get it cheaper per FLOP than a Blackwell or H200 seat.
NVIDIA H200 SXM
The H200 SXM is the memory-upgraded revision of Hopper: the same GH100 compute die as the H100, but with 141 GB of HBM3e across six memory stacks instead of 80 GB across five, and 4.8 TB/s of bandwidth instead of 3.35 TB/s. Compute throughput per GPU is identical to H100, so the H200 exists to fix the memory ceiling, not to add FLOPS: bigger KV caches, larger single-GPU model shards, and fewer GPUs needed to hold a given model. NVIDIA markets it as drop-in compatible with H100-qualified systems at the same 700 W envelope, which is why it is the common upgrade path for operators with an existing Hopper fleet who are not ready to requalify for Blackwell. It sits above H100 and below Blackwell (B200/B300) in NVIDIA's current lineup, positioned as the memory-bound-workload answer within Hopper rather than a new architecture.
NVIDIA B200
The B200 is NVIDIA's first-generation Blackwell data center GPU: two reticle-limited dies joined by a 10 TB/s NV-HBI link and packaged as one SXM6 module with 180 GB of HBM3e. It sells as a bare module only inside an 8-GPU HGX B200 baseboard, never standalone, and that baseboard is the building block for DGX B200 and third-party OEM systems (Supermicro, Dell, HPE, Lenovo). In a private AI build it is the training and large-batch inference workhorse for shops that need FP4 throughput and 180 GB of HBM per GPU but do not need the 288 GB tier or the extra power headroom of B300. As of mid-2026 it sits one rung below Blackwell Ultra (B300) in NVIDIA's current lineup and OEMs are shifting new quotes toward B300, but B200 fleets already deployed remain the majority of installed Blackwell capacity.
NVIDIA B300
The B300 is NVIDIA's Blackwell Ultra data center GPU, a binned and re-engineered Blackwell die that trades the B200's 8-Hi HBM3e stacks for 12-Hi stacks to land 288 GB per GPU, and raises the power ceiling to 1,400 W to push dense NVFP4 throughput about 1.5x past B200. It ships on the same 8-GPU HGX baseboard pattern as B200 and is also the GPU building block of the GB300 NVL72 rack (72 GPUs, 36 Grace CPUs, fully liquid-cooled). In a private AI build it is the part to reach for when a single GPU's memory ceiling, not raw FLOPS, is what is blocking a large-context or reasoning-model deployment. As of mid-2026 it is NVIDIA's current top-of-stack Blackwell part, shipping since January 2026, and new OEM quotes are increasingly steering toward it over B200.
NVIDIA RTX PRO 6000 Blackwell Server Edition
The RTX PRO 6000 Blackwell Server Edition is NVIDIA's passively cooled, PCIe form factor Blackwell part built for standard server chassis rather than SXM baseboards. It carries 96 GB of GDDR7 on a 512 bit bus, no NVLink, and a 600 W ceiling, so it drops into off the shelf 2U to 6U rackmount servers from Dell, HPE, Cisco, Lenovo, and Supermicro at up to eight cards per node instead of requiring a liquid cooled SXM baseboard and NVLink fabric. This is the class of part GPU Smith specs for single node and small cluster private inference builds under about eight GPUs, where the memory capacity and FP4 throughput matter more than multi-node interconnect.
NVIDIA L40S
The L40S is NVIDIA's Ada Lovelace generation, air cooled PCIe accelerator: 48 GB of GDDR6 with ECC, 864 GB/s of bandwidth, no NVLink, and a 350 W ceiling in a standard dual-slot server card. It predates the Blackwell RTX PRO 6000 Server Edition by two years and remains in active production and wide OEM and channel distribution. It is the default air cooled PCIe workhorse GPU Smith specs into single node and small cluster inference builds under about eight GPUs when a model comfortably fits in 48 GB and budget or lead time rules out the newer, pricier Blackwell part.