Published peak, not measured application throughput
Remember this
A nominal 70B BF16 payload is about 140 GB. That leaves only about 1 GB on a nominal 141 GB card before overhead, so “70B fits on one H200 in BF16” is not a sound serving plan.
When this is a sensible choice
Start here if…
Compare H200 when H100 needs an extra shard merely to fit the model or its context. A smaller tensor-parallel group can also remove communication, but only a workload test shows the latency effect.
Choose another configuration if…
If your job is already compute-bound at the same precision, extra HBM alone may not buy much. Evaluate the actual completed work per rental dollar.
Specifications with their boundaries attached
Architecture
Hopper
Memory
141 GB HBM3e
Memory bandwidth
4.8 TB/s per accelerator
Peak compute
989.5 TFLOPS BF16/FP16; 1,979 TFLOPS FP8, dense
Scale-up interconnect
NVLink 4: 900 GB/s bidirectional per GPU
Host attachment
PCIe 5.0 x16
Power
Up to 700 W
Partitioning
Up to seven MIG instances
Catalogue status
Documented product
Compute figures are theoretical peaks at the stated precision. Dense and structured-sparse rates must not be mixed. Bandwidth labelled bidirectional combines both directions. See the source documents.
Follow the bytes · conceptual topology
Three bandwidths, three different jobs
Local memory141 GB HBM3e
↔
Compute enginesExecute kernels on these bytes
① Memory bandwidth: 4.8 TB/s
Accelerator AOwn local memory
↔
Accelerator BOwn local memory
② Scale-up: NVLink, Infinity Fabric or PCIe NVLink 4: 900 GB/s bidirectional per GPU
Server AAccelerators + host
↔
Server BAnother fabric domain
③ Scale-out: NICs + switches + placement InfiniBand, RoCE, EFA or provider-specific transport
A 400 Gb/s NIC has a 50 GB/s raw line-rate equivalent before overhead. A 900 GB/s bidirectional NVLink figure counts traffic in both directions. Neither is the bandwidth at which a GPU reads its own HBM. This diagram explains the boundaries; it is not a wiring diagram for a particular cloud machine.
H200 NVL vs H200 SXM: what changes?
Both variants provide 141 GB of HBM3e and 4.8 TB/s of memory bandwidth per GPU. NVL is a PCIe card with a lower compute and power ceiling; SXM is a module used in HGX systems. Choose the server topology as well as the GPU.
Published specifications per GPU. These are ceilings, not measured model throughput.
Why do H200 NVL BF16 figures say both 835.5 and 1,671 TFLOPS?
NVIDIA labels its 1,671 TFLOPS BF16/FP16 figure “with sparsity.” The dense equivalent is half: 835.5 TFLOPS. Its 3,341 TFLOPS FP8 sparse peak similarly corresponds to 1,670.5 TFLOPS dense. Compare the same precision and sparsity convention; ordinary dense weights do not earn the sparse rate automatically.
NVIDIA H200 specification table and footnotes ↗, checked October 21, 2026 UTC. NVIDIA still labels the table preliminary. Confirm the purchased server’s power setting and two- or four-GPU NVLink bridge layout; eight cards in a chassis do not establish an eight-GPU NVSwitch domain.
320B · total parameter proxy · 18B active in the main model
320B is the model card’s total; 18B active is not its storage size. The vLLM recipe reports about 306 GiB for the native FP8 checkpoint. Hopper requires BF16 KV for this model; the documented ROCm path targets gfx950, not every Instinct GPU.
753B · total parameter proxy · active count not used for sizing
753B is the Hub’s rounded tensor count, not a claim that every tensor uses FP8. The 753 GB raw 8-bit estimate excludes scales, caches and workspace. Do not reuse Flash’s architecture or serving flags for this model. The vLLM recipe describes eight H200s for FP8 and eight B200s for full context. Its 0.29.0+ header conflicts with the 0.28.0 installation; pin a working build. A 32-sequence scheduling cap does not prove 32 concurrent million-token requests. The mixed NVFP4 variant is reported at about 465 GB, larger than an all-four-bit estimate.
Official repository initial upload is labelled 0828. Date evidence ↗
What is the memory floor?
Start with total parameters, then add the memory the workload needs. This arithmetic does not certify a serving configuration. All output sizes below are decimal GB.
Override device capacity with the memory exposed by your allocation, particularly for cloud B300 and partitioned devices. The starting 32 GB budget and 15% reserve are editable teaching assumptions. They are not measurements for the selected model. Mixed-precision tensors, quantization scales, vision encoders and draft models can increase the weight payload.
320 GB
Raw weights only 320B × 8 bits ÷ 8
119.85 GB
Budget per device 141 GB × (1 − 15%)
3 devices
Arithmetic lower bound round up ((weights + 32) ÷ budget)
WeightsCache + runtimeRemaining budget
This assumes perfectly balanced sharding. A result of three does not prove that a three-device parallel layout is supported. Check layer/expert divisibility, actual allocatable memory, precision kernels and the fabric before renting.
320B is the model card’s total; 18B active is not its storage size. The vLLM recipe reports about 306 GiB for the native FP8 checkpoint. Hopper requires BF16 KV for this model; the documented ROCm path targets gfx950, not every Instinct GPU. Read the model source ↗
For full training, also budget gradients, optimizer states, activations and communication buffers. The inference weight estimate above is insufficient.
Tokens per second, TTFT and TPOT are not certified for this hardware in our reference. Use the benchmark checklist to compare an exact model, software revision and workload.
Nearby memory capacities, different tradeoffs
These are comparison candidates selected by memory capacity, not performance rankings or drop-in replacements.
NVIDIA H200 NVL is a dual-slot PCIe GPU with 141 GB HBM3e, 4.8 TB/s memory bandwidth and up to 600 W configurable power. Its dense BF16/FP16 peak is 835.5 TFLOPS; compare its bridge topology and compute limits with H200 SXM below.
A rack-scale system, not a single PCIe card. This comparison uses 186 GB per GPU from GCP’s four-GPU VM configuration; the full rack includes 72 GPUs and 36 Grace CPUs.
Sources and review scope
Manufacturer and cloud documentation checked 2026-09-12. We reviewed specifications and the stated product boundaries; we did not run training or serving benchmarks on this device. Cloud memory and availability can differ by configuration.
Include your access date when citing a changing specification. Link to the specification section for hardware figures or the explanation for a sizing or workload decision.
Research method and limits
We compile manufacturer specifications, cloud documentation, model cards and serving recipes. The reference preserves source units and distinguishes individual devices from nodes and racks. Conflicting figures and unknown fields remain labelled.
Our contribution is the comparison, unit reconciliation, worked arithmetic and workload explanation. Published peaks are vendor specifications. Calculator results are estimates under the displayed assumptions. Neither is a measurement from our own accelerator lab.
For a manufacturer’s specification, consult the original source documents. Cite our page when using its analysis, and retain primary-source attribution for underlying figures. This curated reference does not establish market share, live availability or a universal performance ranking.
Found a discrepancy? Send a correction with the page URL, exact variant, disputed figure and supporting primary document.
