SLYD
Read first

The 5-step pipeline.

How energy turns into a deployed, financed, offtake-matched cluster, and what SLYD does at every step.

How SLYD works →
Hardware

New and recovered GPU systems.

NVIDIA and AMD systems through documented manufacturer and qualified channel supply, with financing and deployment coordinated on the same platform.

Explore GPU hardware →
Marketplace

Compute, hardware, and power in one book.

Browse available GPU capacity by accelerator, configuration, region, and price, or bring supply to qualified demand.

Open marketplace →
Pre-qualify

Start with an indicative structure.

Tell us deal size, structure, and offtake. Any range is preliminary and subject to underwriting, diligence, and documentation.

Open Configure →
From the blog

GPU market trends and deployment playbooks.

Infrastructure best practices, hardware comparisons, and industry analysis from the SLYD team.

Read the blog →

Library

NVIDIA GPU servers

NVIDIA Hopper H100 and H200 systems

The Hopper decision is mostly about memory and form factor. This page keeps SXM and PCIe on separate rows, because they are separate products, and works through the system choices that follow from each.

What are NVIDIA H100 and H200 systems?

NVIDIA H100 and H200 are Hopper-generation data-center GPUs available through multiple system configurations. H200 generally addresses workloads that benefit from greater accelerator memory and memory bandwidth, while the correct system still depends on form factor, interconnect, server design, workload, software, power, cooling, condition, warranty, and verified availability.

Manufacturer specifications

Four products, not two

NVIDIA publishes separate specifications for the SXM and PCIe versions of each generation. Treating them as one product is the single most common source of misconfigured Hopper quotes.

Source: NVIDIA H100 and H200 product pages, reviewed 18 August 2026. NVIDIA footnotes the H200 table as preliminary specifications that may be subject to change, and marks the TF32, BFLOAT16, FP16, FP8, and INT8 Tensor Core rows on both generations as with sparsity.
Specification H200 SXM H200 NVL H100 SXM H100 NVL
GPU memory 141 GB 141 GB 80 GB 94 GB
Memory bandwidth 4.8 TB/s 4.8 TB/s 3.35 TB/s 3.9 TB/s
Form factor SXM PCIe, dual-slot air-cooled SXM PCIe, dual-slot air-cooled
Max thermal design power Up to 700W, configurable Up to 600W, configurable Up to 700W, configurable 350W to 400W, configurable
Interconnect NVLink 900 GB/s, PCIe Gen5 128 GB/s 2-way or 4-way NVLink bridge at 900 GB/s per GPU, PCIe Gen5 128 GB/s NVLink 900 GB/s, PCIe Gen5 128 GB/s NVLink 600 GB/s, PCIe Gen5 128 GB/s
FP8 Tensor Core (with sparsity) 3,958 TFLOPS 3,341 TFLOPS 3,958 TFLOPS 3,341 TFLOPS
FP16 Tensor Core (with sparsity) 1,979 TFLOPS 1,671 TFLOPS 1,979 TFLOPS 1,671 TFLOPS
FP64 Tensor Core 67 TFLOPS 60 TFLOPS 67 TFLOPS 60 TFLOPS
Multi-Instance GPU Up to 7 MIGs at 18 GB each Up to 7 MIGs at 16.5 GB each Up to 7 MIGs at 10 GB each Up to 7 MIGs at 12 GB each

At matching form factors the compute figures are the same across the two generations. The generational difference is memory capacity and bandwidth, which is why the H100 to H200 decision is usually settled by whether the workload is memory bound.

Workload fit

Which one the workload actually wants

Memory bound

H200 usually wins

Large-model inference, long context, and workloads where the model, activations, or KV cache force sharding on an 80 GB part. The 141 GB capacity and 4.8 TB/s bandwidth are what change the outcome, not the compute figures.

Compute bound

H100 remains competitive

At matching form factors NVIDIA publishes identical Tensor Core throughput. If the workload fits comfortably in 80 GB, the extra memory buys nothing and the decision moves to the total transaction.

Tightly coupled

SXM for multi-GPU training

Full NVLink bandwidth across an 8-GPU baseboard, with the higher power budget and the cooling design that comes with it.

Constrained site

PCIe for air-cooled racks

NVIDIA describes H200 NVL as suited to lower-power air-cooled enterprise rack designs with flexible configurations. Where the facility limits density, this is often the deployable option.

System options

How Hopper GPUs are packaged into servers

  1. HGX baseboard systems

    NVIDIA publishes HGX H100 and HGX H200 partner and NVIDIA-Certified Systems with 4 or 8 GPUs. This is the usual shape for training clusters.

  2. DGX systems

    NVIDIA publishes DGX H100 with 8 GPUs as its own complete system, with NVIDIA software and support attached.

  3. MGX and PCIe partner systems

    NVIDIA publishes MGX H200 NVL partner and NVIDIA-Certified Systems with up to 8 GPUs, and PCIe partner systems with 1 to 8 GPUs, which suits mixed and inference-oriented estates.

  4. Manufacturer implementations

    Dell publishes PowerEdge XE9680 with 8 NVIDIA H100 or H200 GPUs, HPE publishes Cray XD670 as a 5U chassis with 8 NVIDIA H200 Tensor Core SXM5 GPUs and a direct liquid cooling option, and Lenovo publishes ThinkSystem SR675 V3 with NVIDIA H100 and L40S GPUs.

Buyer diligence

New and qualifying recovered systems

Hopper is the generation where recovered supply is most common, which makes lot-level evidence more important than it is for a new-build purchase.

  • Exact GPU model and form factor, SXM or PCIe
  • GPU count and node configuration
  • CPU, memory, storage, NIC, and interconnect configuration
  • Firmware level and software compatibility
  • Power draw and connector requirements
  • Equipment condition and provenance
  • Inspection and test evidence, where it exists
  • Warranty provider, scope, term, and transferability
  • Availability confirmation with its date and expiration
  • Export, end-use, and geographic restrictions

SLYD does not represent that every recovered lot is tested, graded, warrantied, or ready for immediate delivery. Condition, testing, and warranty evidence are disclosed for the specific lot when they exist.

FAQ

Hopper server questions

What is the difference between NVIDIA H100 and H200?

Both are Hopper-generation data-center GPUs with the same published Tensor Core throughput at matching form factors. The difference is memory. NVIDIA publishes 141 GB and 4.8 TB/s of memory bandwidth for H200 SXM against 80 GB and 3.35 TB/s for H100 SXM, so H200 addresses workloads limited by accelerator memory or memory bandwidth.

Is H200 only available in an SXM form factor?

No. NVIDIA publishes both H200 SXM and H200 NVL. H200 NVL is a PCIe dual-slot air-cooled part with a configurable TDP of up to 600W, described by NVIDIA as suited to lower-power air-cooled enterprise rack designs, while H200 SXM has a configurable TDP of up to 700W.

What is the difference between SXM and PCIe Hopper GPUs?

They are different products with different published specifications, not packaging variants of the same part. SXM parts sit on an integrated baseboard with higher power budgets and full NVLink bandwidth. PCIe parts install in standard server slots at lower power with reduced interconnect bandwidth. Memory capacity, bandwidth, TDP, and NVLink figures all differ between them.

How many H100 or H200 GPUs go in a server?

NVIDIA publishes HGX H100 and HGX H200 partner and NVIDIA-Certified Systems with 4 or 8 GPUs, DGX H100 with 8 GPUs, and MGX H200 NVL partner and NVIDIA-Certified Systems with up to 8 GPUs. PCIe-based partner systems are published with 1 to 8 GPUs. The actual configuration comes from the specific server model.

Should I still buy Hopper, or move to Blackwell?

It depends on the workload, the facility, and the total transaction. Hopper systems have lower power and cooling requirements than current Blackwell platforms, which matters if the site is constrained, and a mature software ecosystem. Blackwell platforms offer more memory and interconnect bandwidth per system. The comparison should be run on the specific workload rather than by generation.

What should I check when evaluating recovered Hopper systems?

Confirm the exact GPU model and form factor, GPU count and node configuration, CPU, memory, storage and network configuration, firmware level, condition and provenance, any inspection or test evidence, warranty provider and transferability, quantity and location, and the export and end-use restrictions that apply. Evidence is evaluated for the specific lot rather than assumed.

Do NVIDIA's published Hopper performance figures assume sparsity?

The Tensor Core rows do. NVIDIA marks TF32, BFLOAT16, FP16, FP8, and INT8 Tensor Core figures for H100 and H200 as with sparsity, and additionally marks the H200 specification table as preliminary specifications that may be subject to change. Both caveats should travel with the numbers.

Sources for the specifications on this page

Specifications belong to the manufacturers named and are reproduced with their own caveats. They are not SLYD guarantees of performance, configuration, or availability.

Scope a Hopper requirement

Share the workload, memory requirement, node count, facility constraint, and condition tolerance. Those decide the form factor and the sourcing path.

Page updated: August 18, 2026

Reconnecting to the server...

Please wait while we restore your connection

An unhandled error has occurred. Reload 🗙