NV NVIDIA Enterprise Discrete GPU

NVIDIA H200

Runs downloadable LLMs up to ~221B at Q4, ~782.1-932.1 tok/s on an 8B model.

NVIDIA H200
141GB
VRAM
4800
GB/s bandwidth
221B
Max model (Q4)
700W
TDP · PSU 1100W

What you can run

At 8k context, FP16 KV. Full-speed tok/s holds up to "max ctx"; beyond it the context spills to system RAM (CPU/PCIe offload) and drops to the degraded figure.

Model sizeFP16Q8Q4
3B 584-696 t/s
≤1865k ctx · then ~4.9-5.8 (slow)
1168-1392 t/s
≤1911k ctx · then ~9.7-11.6 (slow)
2085.7-2485.7 t/s
≤1931k ctx · then ~17.4-20.7 (slow)
7-8B 219-261 t/s
≤856k ctx · then ~1.8-2.2 (slow)
438-522 t/s
≤917k ctx · then ~3.7-4.4 (slow)
782.1-932.1 t/s
≤944k ctx · then ~6.5-7.8 (slow)
13B 134.8-160.6 t/s
≤624k ctx · then ~1.1-1.3 (slow)
269.5-321.2 t/s
≤703k ctx · then ~2.2-2.7 (slow)
481.3-573.6 t/s
≤738k ctx · then ~4-4.8 (slow)
32B 54.8-65.3 t/s
≤245k ctx · then ~0.5-0.5 (slow)
109.5-130.5 t/s
≤367k ctx · then ~0.9-1.1 (slow)
195.5-233 t/s
≤421k ctx · then ~1.6-1.9 (slow)
70B won't run 50.1-59.7 t/s
≤178k ctx · then ~0.4-0.5 (slow)
89.4-106.5 t/s
≤272k ctx · then ~0.7-0.9 (slow)
120B won't run 29.2-34.8 t/s
≤25k ctx · then ~0.2-0.3 (slow)
52.1-62.1 t/s
≤186k ctx · then ~0.4-0.5 (slow)
400B won't run won't run won't run

Estimate: tok/s ≈ bandwidth ÷ (params × bytes/param) × efficiency (NVIDIA). Cyan = fits ~129.7 GB usable. Beyond "ctx", KV spills to system RAM (CPU/PCIe offload) (rough degraded speed shown).

What else you need for a full build

Minimum supporting components for a desktop built around the NVIDIA H200. Guidance — exact needs depend on your case, other parts and how much you offload to system RAM.

Power supply
≥ 1100W · 80+ Gold
Headroom over the 700W card plus the rest of the build.
System RAM
≥ 256 GB · 512 GB recommended
At least the 141 GB of VRAM; 2× lets you CPU-offload models larger than VRAM.
Motherboard
1× free PCIe x16 slot (4.0/5.0)
A full-length x16 electrical slot; PCIe 4.0 is ample, 5.0 future-proofs.
CPU
Any modern multi-core
The GPU runs inference; more cores and memory channels only matter when offloading to RAM.
Storage
≥ 1 TB NVMe SSD
Weights are large — a 70B model at Q4 is ~40 GB; budget for a few.
PSU connector
12V-2×6 / 12VHPWR (or 3× 8-pin)
High-power NVIDIA cards use the 16-pin 12V-2×6 power connector.

Run popular models on the NVIDIA H200

Videos about the NVIDIA H200

H200 vs H100: Ultimate AI Inference GPU Comparison 2025

Nvidia Wouldn't Send Me This $30,000 GPU - H200 Holy $H!T

More hardware like the NVIDIA H200

About the NVIDIA H200

NVIDIA H200 is a discrete GPU from NVIDIA, 2024. For local genAI the numbers that matter are its 141 GB (how big a model + context fits) and 4800 GB/s bandwidth (generation speed). Plan for a ~1100W power supply.

Frequently asked questions

What size LLM can the NVIDIA H200 run?

With ~129.7 GB usable it runs up to ~221B at Q4 (~62.5B at FP16) at 8k context. Bigger spills into system RAM (CPU/PCIe offload) and slows sharply.

How many tokens/sec does the NVIDIA H200 generate?

Generation is bandwidth-bound. At 4800 GB/s an 8B model at Q4 runs ~782.1-932.1 tok/s (estimate).

Why does long context slow it down?

The KV cache for the context must also fit fast memory; once weights+KV exceed 129.7 GB it spills to system RAM (CPU/PCIe offload) and throughput collapses. Each row shows the max usable context.

What else do I need to run the NVIDIA H200?

Besides the card: a ≥1100W 80+ Gold power supply, at least 141GB of system RAM (2× the 141GB VRAM is ideal for CPU offload), a free PCIe x16 slot, any modern multi-core CPU, and a ≥1TB NVMe SSD for model weights.