NVIDIA

NVIDIA H200 PCIe 141GB

Hopper DatacenterDatacenterHopperPCIe 5CUDA

Operating mode

Choose the operating mode for this hardware

Use this to bias workload recommendations toward responsiveness, background autonomy, lighter serving, or multi-GPU scale-out.

Current mode

Balanced

Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.

See Full AI Tier List for NVIDIA H200 PCIe 141GB →

About this GPU for AI

The NVIDIA H200 PCIe is the high-memory Hopper accelerator in a standard PCIe 5.0 form factor, providing 141 GB of HBM3e and 4,800 GB/s bandwidth — the same memory subsystem as the SXM flagship — while using the same 756 TFLOPS FP16 compute core as the H100 PCIe. This makes it unique: you get H200-class memory (enough to run 70B models at FP16 with massive KV cache) in any PCIe-compatible server. It is particularly compelling for long-context inference workloads where memory capacity is the binding constraint rather than compute.

Beyond LLMs

AI Capability Matrix

What AI tasks this GPU can handle — from text generation to image and video creation.

Capability	Status	Representative Model	Detail
LLM Chat (7B)	Runs natively	Llama 3.1 8B Q4	—
LLM Coding (30B)	Runs natively	Qwen 3 30B Q4	—
LLM Large (70B)

hbm-memorymassive-vrampcie-form-factorhigh-bandwidthdatacenter-grade

Specifications

Compute

FP16756 TFLOPS

INT81512 TOPS

ArchitectureHopper

Memory

VRAM141 GB

Bandwidth4800 GB/s

General

FamilyHopper Datacenter

SegmentDatacenter

InterconnectPCIe 5

Compute PlatformCUDA

MSRP$30,000

Key Features

141 GB HBM3e — 4,800 GB/s bandwidth (matches H200 SXM memory spec)756 TFLOPS FP16 with sparsity / 1,512 INT8 TOPSFP8 Transformer EnginePCIe 5.0 x16, 350W TDPMIG support: up to 7 isolated instancesNo NVLink — multi-GPU via PCIe peer-to-peer

For AI Workloads

Strengths

141 GB HBM3e fits 70B models at FP16 with extensive KV cache — ideal for long-context inference
4.8 TB/s bandwidth matches H200 SXM — fastest possible decode speed in the PCIe form factor
Standard PCIe 5.0 form factor accessible without SXM infrastructure investment
Strong single-GPU option for organizations serving large models at scale without multi-GPU complexity

Considerations

Same 756 TFLOPS FP16 as H100 PCIe — no compute improvement over the previous generation in PCIe form
No NVLink limits multi-GPU scaling — less efficient than SXM for tensor parallelism across cards
Very high price for a PCIe card; cost-per-TFLOPS is significantly worse than SXM alternatives
Blackwell B200 PCIe and B100 variants are already available, offering higher compute at similar VRAM tiers

Architecture

Hopper

Hopper is NVIDIA's datacenter-focused architecture succeeding Ampere. Built on TSMC 4N, it introduces the Transformer Engine with automatic FP8/FP16 mixed-precision training, HBM3/HBM3e memory, and NVLink 4.0 for multi-GPU scaling. The H100 flagship delivers up to 3x the AI training performance of A100.

AI Relevance

The Transformer Engine automatically manages FP8 precision for optimal training speed without accuracy loss. With up to 141 GB HBM3e (H200), Hopper GPUs can hold the largest open-weight models entirely in GPU memory, making them the workhorse of AI datacenters.

Process: TSMC 4NPlatform: CUDATensor Cores: Gen 4Precisions: FP64, FP32, TF32, FP16, BF16, FP8, INT8

Recommendations by Workload

Chat

Qwen 3.5 122B A10B

Qwen 3.5 122B A10B matches Chat and keeps a practical fit profile. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.

Decode 162.1 tok/s · 131K ctx · llama.cppEST.

90.6 GB / 141.0 GB VRAM

Coding

Qwen 3.5 122B A10B

Qwen 3.5 122B A10B is a specialized fit for Coding. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.

Decode 162.1 tok/s · 131K ctx · llama.cppEST.

91.9 GB / 141.0 GB VRAM

Agentic Coding

Full Model Compatibility

👁 Alibaba
Qwen 3.5 122B A10B

S98

122B91.9 GB162 tok/s131K ctx

moe

👁 Mistral
Devstral 2 123B Instruct

S98

123B95.4 GB58 tok/s152K ctx

dense

👁 Mistral
Mistral Small 4 119B

Just out of reach

Models you could run with an upgrade

High-quality models that need a bit more memory

👁 Alibaba
Qwen 3.5 397B A17B

397BTier 100Needs ~258.6 GB

Runs on AMD Instinct MI350X 288GB (~$8,000)

Also runs on 2× your GPU via PCIe — 85 tok/s

👁 Moonshot AI
Kimi K2.5

1000BTier 100Needs ~628.7 GB

👁 Moonshot AI
Kimi K2.6

1000BTier 100Needs ~628.7 GB

👁 DeepSeek
DeepSeek V4 Pro

1600BTier 100Needs ~877.9 GB

👁 DeepSeek
DeepSeek V4 Flash

284BTier 98Needs ~173.7 GB

Runs on AMD Instinct MI350X 288GB (~$8,000)

Also runs on 2× your GPU via PCIe — 136 tok/s

Image & Video Generation

Diffusion Model Compatibility

52 of 52 models can generate images or video on your NVIDIA H200 PCIe 141GB

Model	Max Resolution	Gen Time	Grade
SD TurboImage	512×512	0ms	S
Stable Diffusion 1.5Image	512×768	100ms	S
Realistic Vision v5.1Image	512×768	100ms	S
DreamShaper 8Image	512×768	100ms	S
LCM DreamShaper v7

Multi-GPU scaling

NVIDIA H200 PCIe 141GB — Up to 4× via PCIe

Scale out with multiple GPUs for larger models. PCIe interconnect with 22% scaling overhead.

Config	Effective memory	Models that fit	Est. bandwidth
1× NVIDIA	141 GB	353/374	4,800 GB/s
2× NVIDIA	282 GB	364/374	7,488 GB/s
4× NVIDIA	564 GB	373/374	14,976 GB/s

Model counts use default quantization at coding workload settings. Multi-GPU scaling factor: 0.78× per additional GPU.

Upgrade paths

Upgrade from NVIDIA H200 PCIe 141GB

See what you unlock with more powerful hardware

Upgrade options

👁 NVIDIA
4× NVIDIA H200 PCIe 141GBMulti-GPU

4 × 141 GB = 564 GB effectivevia PCIe

Unlocks 20 additional models that do not fit on the current setup.Unlocks Qwen 3.5 397B A17B, Kimi K2.5, Kimi K2.6+17 more · +54% faster avg

Unlocks 20 additional models that do not fit on the current setup.

Lifts average decode speed across fitting models by about 54%.

Scale-out only pays off if the host platform has enough PCIe lanes, slot spacing, power, and cooling.

The bigger the setup gets, the more the runtime matters. Multi-GPU and multi-user serving are where vLLM, SGLang, TGI, TensorRT-LLM, or tuned llama.cpp start to earn their complexity.

~$30,000 MSRP

👁 NVIDIA
NVIDIA B200 180GBNext step up

180 GB VRAM (+39)8000 GB/s (+3200)

Unlocks 6 additional models that do not fit on the current setup.Unlocks DeepSeek V4 Flash, DeepSeek Coder V2 236B, DeepSeek V2.5 236B+3 more · +22% faster avg

Unlocks 6 additional models that do not fit on the current setup.

Lifts average decode speed across fitting models by about 22%.

~$30,000 MSRP

👁 NVIDIA
NVIDIA GB200 192GBNVIDIA upgrade

192 GB VRAM (+51)8000 GB/s (+3200)

Unlocks 6 additional models that do not fit on the current setup.Unlocks DeepSeek V4 Flash, DeepSeek Coder V2 236B, DeepSeek V2.5 236B+3 more · +22% faster avg

Unlocks 6 additional models that do not fit on the current setup.

Lifts average decode speed across fitting models by about 22%.

~$60,000 MSRP

AMD Instinct MI325X 256GBBiggest leap

256 GB VRAM (+115)6000 GB/s (+1200)

Unlocks 10 additional models that do not fit on the current setup.Unlocks Qwen 3.5 397B A17B, DeepSeek V4 Flash, DeepSeek Coder V2 236B+7 more · +2% faster avg

Unlocks 10 additional models that do not fit on the current setup.

~$20,000 MSRP

AMD Instinct MI350X 288GBBest value

288 GB VRAM (+147)8000 GB/s (+3200)

Unlocks 11 additional models that do not fit on the current setup.Unlocks Qwen 3.5 397B A17B, DeepSeek V4 Flash, Qwen3-Coder 480B A35B Instruct+8 more · +14% faster avg

Unlocks 11 additional models that do not fit on the current setup.