NVIDIA B200, AI & Machine Learning Benchmarks & Specs

192GB · AI Score 78.0/100 · first-party measured on 12 AI workloads

78 AI Score ✓ Measured

Every number on this page is first-party: NVIDIA B200 was run on our pinned 12-workload AI suite on 2026-07-10, with under 0.5% run-to-run variance. On Llama 3.1 8B (Q4_K_M) NVIDIA B200 delivers about 274.41 tokens/sec. Stepping up to Qwen3 32B it holds roughly 78.56 tok/s. The full Llama 3.3 70B still runs, at about 44.54 tok/s. For image generation, SDXL runs at 23.06 it/s, and FLUX.1-dev at 6.13 it/s. All 12 workloads fit in 192GB. There is no model in our suite this card has to turn down. NVIDIA B200 isn't a retail purchase for most people. It's rented by the hour. You can run this exact card on RunPod.

AI & Machine Learning benchmark results

Text Generation tok/s 123

LFM2.5-1.2B916.79

gemma-3-270m889.52

Llama 3.2 1B881.78

Qwen2.5-Coder-0.5B828.48

Qwen2.5-0.5B818.27

SmolLM2-135M815.3

Qwen1.5-0.5B699.73

Qwen3 0.6B599.63

Qwen3-0.6B596.73

LFM2.5-8B-A1B575.15

Qwen3-1.7B551.93

Qwen3 1.7B551.76

Workload	Result	Telemetry	Data
LFM2.5-1.2B	916.79 tok/s	269 W33°CQ4_K_M	✓ Measured
gemma-3-270m	889.52 tok/s	254 W28°CQ4_K_M	✓ Measured
Llama 3.2 1B	881.78 tok/s	294 W34°CQ4_K_M	✓ Measured
Qwen2.5-Coder-0.5B	828.48 tok/s	255 W29°CQ4_K_M	✓ Measured
Qwen2.5-0.5B	818.27 tok/s	252 W29°CQ4_K_M	✓ Measured
SmolLM2-135M	815.3 tok/s	252 W29°CQ4_K_M	✓ Measured
Qwen1.5-0.5B	699.73 tok/s	267 W29°CQ4_K_M	✓ Measured
Qwen3 0.6B	599.63 tok/s	238 W31°CQ4_K_M	✓ Measured
Qwen3-0.6B	596.73 tok/s	266 W30°CQ4_K_M	✓ Measured
LFM2.5-8B-A1B	575.15 tok/s	287 W31°CQ4_K_M	✓ Measured
Qwen3-1.7B	551.93 tok/s	288 W32°CQ4_K_M	✓ Measured
Qwen3 1.7B	551.76 tok/s	278 W33°CQ4_K_M	✓ Measured
gemma-3-1b	519.19 tok/s	275 W30°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill 1.5B	456.7 tok/s	312 W33°CQ4_K_M	✓ Measured
Qwen2.5-Coder-1.5B	455.35 tok/s	293 W32°CQ4_K_M	✓ Measured
Qwen2.5-1.5B	454.69 tok/s	294 W30°CQ4_K_M	✓ Measured
Qwen2-1.5B	454.55 tok/s	288 W31°CQ4_K_M	✓ Measured
Llama 3.2 3B	422.09 tok/s	325 W35°CQ4_K_M	✓ Measured
Hermes-3-Llama-3.2-3B	420.6 tok/s	331 W32°CQ4_K_M	✓ Measured
Llama-3.2-3B-Instruct-uncensored	419.24 tok/s	320 W34°CQ4_K_M	✓ Measured
SmolLM3 3B	400.73 tok/s	334 W35°CQ4_K_M	✓ Measured
SmolLM3-3B	399.05 tok/s	322 W33°CQ4_K_M	✓ Measured
gemma-2-2b-it-abliterated	392.32 tok/s	322 W32°CQ4_K_M	✓ Measured
gemma-2-2b	391.94 tok/s	319 W32°CQ4_K_M	✓ Measured
Qwen2.5-Coder-3B	388.17 tok/s	324 W33°CQ4_K_M	✓ Measured
Qwen2.5-3B	388.14 tok/s	319 W32°CQ4_K_M	✓ Measured
Phi-4 Mini 3.8B	375.74 tok/s	329 W36°CQ4_K_M	✓ Measured
Phi-4-mini	375.06 tok/s	324 W34°CQ4_K_M	✓ Measured
AI21-Jamba-Reasoning-3B	360.54 tok/s	332 W32°CQ4_K_M	✓ Measured
gpt-oss-20b	351.42 tok/s	318 W32°CQ4_K_M	✓ Measured
phi-2	348.72 tok/s	325 W32°CQ4_K_M	✓ Measured
Nemotron-3-Nano-30B-A3B	345.93 tok/s	306 W32°CQ4_K_M	✓ Measured
DeepSeek-Coder-V2-Lite	325.85 tok/s	295 W32°CQ4_K_M	✓ Measured
Phi-3.5-mini	325 tok/s	328 W32°CQ4_K_M	✓ Measured
Qwen3 4B	317.93 tok/s	3.1 GB peak288 W35°C1.11 tok/WQ4_K_M	✓ Measured
Qwen3-4B-Thinking-2507	317.03 tok/s	341 W33°CQ4_K_M	✓ Measured
Qwen3-4B-Instruct-2507	316.7 tok/s	349 W33°CQ4_K_M	✓ Measured
Qwen3-4B-Instruct-2507	316.66 tok/s	348 W33°CQ4_K_M	✓ Measured
Llama-2-7B	292 tok/s	391 W35°CQ4_K_M	✓ Measured
Gemma 3 4B	291.82 tok/s	331 W35°CQ4_K_M	✓ Measured
Mistral 7B v0.3	287.42 tok/s	324 W37°CQ4_K_M	✓ Measured
Mistral-7B-Instruct-v0.1	286.89 tok/s	390 W36°CQ4_K_M	✓ Measured
Mistral-7B-Instruct-v0.3	286.48 tok/s	412 W35°CQ4_K_M	✓ Measured
Mistral-7B-Instruct-v0.2	286.37 tok/s	375 W35°CQ4_K_M	✓ Measured
Qwen2.5-Coder 7B	277.98 tok/s	341 W38°CQ4_K_M	✓ Measured
Qwen3-Coder 30B A3B	277.68 tok/s	282 W37°CQ4_K_M	✓ Measured
Qwen2.5-7B	277.17 tok/s	387 W35°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill 7B	277.02 tok/s	299 W35°CQ4_K_M	✓ Measured
Qwen2.5-Coder-7B-Instruct-abliterated	276.57 tok/s	391 W35°CQ4_K_M	✓ Measured
dolphin-2.9-llama3-8b	274.48 tok/s	395 W35°CQ4_K_M	✓ Measured
L3-8B-Stheno-v3.2	274.48 tok/s	382 W35°CQ4_K_M	✓ Measured
Llama 3.1 8B	274.41 tok/s	5.1 GB peak373 W39°C0.74 tok/WQ4_K_M	✓ Measured
DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored	274.3 tok/s	398 W36°CQ4_K_M	✓ Measured
Meta-Llama-3.1-8B	274.17 tok/s	397 W35°CQ4_K_M	✓ Measured
Dolphin 3.0 Llama 3.1 8B	274.05 tok/s	363 W41°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill Llama 8B	273.9 tok/s	314 W35°CQ4_K_M	✓ Measured
Dolphin X1 8B	273.85 tok/s	327 W41°CQ4_K_M	✓ Measured
Qwen3-30B-A3B	271.69 tok/s	300 W31°CQ4_K_M	✓ Measured
Qwen3 30B A3B	270.84 tok/s	269 W32°CQ4_K_M	✓ Measured
Qwen3 8B	253.45 tok/s	284 W35°CQ4_K_M	✓ Measured
Josiefied-Qwen3-8B-abliterated-v1	253.28 tok/s	390 W35°CQ4_K_M	✓ Measured
Qwen3-8B	253.14 tok/s	401 W36°CQ4_K_M	✓ Measured
DeepSeek-R1-0528-Qwen3-8B	252.98 tok/s	382 W35°CQ4_K_M	✓ Measured
KAT-Coder-V2.5-Dev	230.68 tok/s	296 W31°CQ4_K_M	✓ Measured
Qwen3 30B A3B (Q3_K_M)	229.58 tok/s	313 W32°CQ3_K_M	✓ Measured
Ornith-1.0-9B	226.2 tok/s	384 W35°CQ4_K_M	✓ Measured
Qwen-AgentWorld-35B-A3B	221.78 tok/s	319 W32°CQ4_K_M	✓ Measured
Ornith-1.0-35B	221.13 tok/s	305 W31°CQ4_K_M	✓ Measured
Dolphin X1 Trinity Nano 6B	209.31 tok/s	272 W28°CQ4_K_M	✓ Measured
Qwen3-Coder-Next	189.76 tok/s	291 W31°CQ4_K_M	✓ Measured
Qwen3-Next-80B-A3B-Thinking	188.78 tok/s	288 W31°CQ4_K_M	✓ Measured
gemma-2-9b	187.45 tok/s	414 W35°CQ4_K_M	✓ Measured
Qwen3-Coder-Next-abliterated	186.57 tok/s	283 W31°CQ4_K_M	✓ Measured
Qwen3-Coder-Next	186.39 tok/s	302 W31°CQ4_K_M	✓ Measured
Qwen3-Next-80B-A3B-Thinking	186.03 tok/s	289 W31°CQ4_K_M	✓ Measured
NemoMix-Unleashed-12B	183.93 tok/s	426 W36°CQ4_K_M	✓ Measured
Mistral-Nemo-Instruct-2407	183.92 tok/s	427 W36°CQ4_K_M	✓ Measured
GLM-4.7-Flash	183.82 tok/s	309 W31°CQ4_K_M	✓ Measured
Qwen3-Next-80B-A3B	179.71 tok/s	296 W32°CQ4_K_M	✓ Measured
GLM-4.7-Flash-REAP-23B-A3B	167.76 tok/s	307 W31°CQ4_K_M	✓ Measured
Phi-4 14B	163.85 tok/s	365 W37°CQ4_K_M	✓ Measured
Qwen3 14B	158.34 tok/s	316 W37°CQ4_K_M	✓ Measured
Hermes-4-14B	158.22 tok/s	440 W36°CQ4_K_M	✓ Measured
Qwen3-14B	158.19 tok/s	445 W36°CQ4_K_M	✓ Measured
Gemma 3 12B	151.41 tok/s	291 W35°CQ4_K_M	✓ Measured
Qwen2.5-Coder 14B	150.96 tok/s	9.1 GB peak399 W40°C0.38 tok/WQ4_K_M	✓ Measured
DeepSeek-R1 Distill 14B	150.69 tok/s	331 W37°CQ4_K_M	✓ Measured
EVA-Qwen2.5-14B-v0.2	150.67 tok/s	442 W36°CQ4_K_M	✓ Measured
Qwen2.5-Coder-14B-Instruct-abliterated	150.67 tok/s	443 W36°CQ4_K_M	✓ Measured
Qwen2.5-14B	150.57 tok/s	432 W36°CQ4_K_M	✓ Measured
Uncensored	150.51 tok/s	433 W36°CQ4_K_M	✓ Measured
Gemma 4 12B	146.36 tok/s	336 W35°CQ4_K_M	✓ Measured
StarCoder2 15B	138.64 tok/s	345 W42°CQ4_K_M	✓ Measured
Phi-4 14B (Q3_K_M)	135.21 tok/s	451 W36°CQ3_K_M	✓ Measured
Gemma 3 12B (Q3_K_M)	125.31 tok/s	425 W35°CQ3_K_M	✓ Measured
DeepSeek-R1 Distill 14B (Q3_K_M)	120.47 tok/s	433 W35°CQ3_K_M	✓ Measured
Mistral Small 24B	114.04 tok/s	331 W38°CQ4_K_M	✓ Measured
Dolphin 3.0 R1 Mistral 24B	113.95 tok/s	333 W44°CQ4_K_M	✓ Measured
Dolphin Mistral 24B Venice	113.93 tok/s	365 W44°CQ4_K_M	✓ Measured
Devstral Small 24B	113.91 tok/s	374 W44°CQ4_K_M	✓ Measured
Cydonia-24B-v4.3	113.81 tok/s	476 W38°CQ4_K_M	✓ Measured
Dolphin-Mistral-24B-Venice-Edition	113.8 tok/s	473 W38°CQ4_K_M	✓ Measured
Codestral 22B	112.16 tok/s	366 W43°CQ4_K_M	✓ Measured
Codestral 22B (Q3_K_M)	89.9 tok/s	475 W37°CQ3_K_M	✓ Measured
Mistral Small 24B (Q3_K_M)	87.72 tok/s	483 W37°CQ3_K_M	✓ Measured
Gemma 3 27B	86.08 tok/s	380 W43°CQ4_K_M	✓ Measured
Olmo-3.1-32B-Think	79.75 tok/s	503 W38°CQ4_K_M	✓ Measured
Dolphin 2.9.1 Yi 1.5 34B	79.35 tok/s	373 W44°CQ4_K_M	✓ Measured
Qwen3 32B	78.56 tok/s	19 GB peak352 W42°C0.22 tok/WQ4_K_M	✓ Measured
DeepSeek-R1 Distill 32B	78.55 tok/s	392 W44°CQ4_K_M	✓ Measured
Qwen2.5-32B	78.55 tok/s	499 W38°CQ4_K_M	✓ Measured
Qwen2.5-Coder 32B	78.53 tok/s	363 W44°CQ4_K_M	✓ Measured
DeepSeek-R1-Distill-Qwen-32B-abliterated	78.48 tok/s	508 W38°CQ4_K_M	✓ Measured
QwQ 32B	78.47 tok/s	349 W44°CQ4_K_M	✓ Measured
Qwen2.5-Coder 32B (Q3_K_M)	60.24 tok/s	495 W37°CQ3_K_M	✓ Measured
Qwen2.5-72B	45.34 tok/s	556 W42°CQ4_K_M	✓ Measured
Llama 3.3 70B	44.54 tok/s	40.2 GB peak387 W47°C0.12 tok/WQ4_K_M	✓ Measured
DeepSeek-R1-Distill-Llama-70B	44.49 tok/s	604 W41°CQ4_K_M	✓ Measured
Hermes-4-70B	44.49 tok/s	583 W41°CQ4_K_M	✓ Measured
Llama-3.3-70B-Instruct-abliterated	44.48 tok/s	575 W41°CQ4_K_M	✓ Measured
Meta-Llama-3.1-70B	44.48 tok/s	594 W41°CQ4_K_M	✓ Measured
Laguna-XS-2.1	✕ Won't fit needs ~26 GB	VRAM-gated at this precision	✓ Measured
Nanbeige4.2-3B	✕ Won't fit needs ~4 GB	VRAM-gated at this precision	✓ Measured

Image Generation images/min 6

FLUX.1 Schnell81.41

Z-Image Turbo (1024px)62.47

Stable Diffusion XL46.12

Z-Image Turbo32.325

FLUX.1 dev13.136

Krea 2 Turbo9.5

Workload	Result	Telemetry	Data
FLUX.1 Schnell	81.41 images/min	770 W43°C	✓ Measured
Z-Image Turbo (1024px)	62.47 images/min	850 W45°C	✓ Measured
Stable Diffusion XL	46.12 images/min	16.3 GB peak719 W44°C1.3 s/img	✓ Measured
Z-Image Turbo	32.33 images/min	26.1 GB peak898 W50°C1.9 s/img	✓ Measured
FLUX.1 dev	13.14 images/min	37 GB peak883 W51°C4.6 s/img	✓ Measured
Krea 2 Turbo	9.5 images/min	801 W59°C	✓ Measured

Fine-Tuning train tok/s 4

SmolLM2 1.7B LoRA18064.7

TinyLlama 1.1B LoRA17993.5

Qwen2.5 1.5B LoRA15322

Qwen2.5 7B LoRA13722

Workload	Result	Telemetry	Data
SmolLM2 1.7B LoRA	18064.7 train tok/s	435 W42°C	✓ Measured
TinyLlama 1.1B LoRA	17993.5 train tok/s	363 W40°C	✓ Measured
Qwen2.5 1.5B LoRA	15322 train tok/s	388 W40°C	✓ Measured
Qwen2.5 7B LoRA	13722 train tok/s	692 W50°C	✓ Measured

LLM Serving serve tok/s 4

TinyLlama 1.1B served11562.2

SmolLM2 1.7B served9672.9

Qwen2.5 1.5B served9186.4

Qwen2.5 7B served5799.1

Workload	Result	Telemetry	Data
TinyLlama 1.1B served	11562.2 serve tok/s	314 W33°C	✓ Measured
SmolLM2 1.7B served	9672.9 serve tok/s	356 W34°C	✓ Measured
Qwen2.5 1.5B served	9186.4 serve tok/s	332 W33°C	✓ Measured
Qwen2.5 7B served	5799.1 serve tok/s	559 W37°C	✓ Measured

Image Editing images/min 2

Workload	Result	Telemetry	Data
FLUX.1 Kontext dev	5.72 images/min	35.8 GB peak884 W53°C10.5 s/img	✓ Measured
Qwen-Image-Edit	4.86 images/min	60.5 GB peak881 W52°C12.3 s/img	✓ Measured

Video Generation frames/s 2

Workload	Result	Telemetry	Data
LTX-Video (distilled)	26.8 frames/s	60.5 GB peak730 W46°C3.6 s/clip	✓ Measured
Wan 2.2 5B (720p)	1.88 frames/s	38 GB peak893 W51°C26 s/clip	✓ Measured

Depth Estimation images/min 2

Workload	Result	Telemetry	Data
Depth Anything V2 Small	1137.49 images/min	190 W28°C	✓ Measured
Depth Anything V2 Large	1039.47 images/min	192 W29°C	✓ Measured

Segmentation images/min 2

Workload	Result	Telemetry	Data
SAM ViT-Base	1662.34 images/min	204 W29°C	✓ Measured
SAM ViT-Huge	388.61 images/min	229 W32°C	✓ Measured

Background Removal images/min 1

Workload	Result	Telemetry	Data
BiRefNet	1082.03 images/min	252 W31°C	✓ Measured

Upscaling images/min 1

Workload	Result	Telemetry	Data
Swin2SR 4x Upscaler	11.97 images/min	282 W34°C	✓ Measured

How we measured this. Every result comes from our own pinned, reproducible AI suite, 12 workloads: the Qwen3-4B to Llama-70B LLM ladder (llama.cpp, Q4_K_M), SDXL / Z-Image / FLUX-dev generation, FLUX-Kontext / Qwen-Edit editing, and LTX / Wan video, run first-party on rented hardware with under 0.5% run-to-run variance. Peak VRAM, power draw, temperature and tokens-per-watt are captured per workload. “Won’t fit” rows are real data: where a model exceeds the card’s VRAM at the tested precision we record a hard gate rather than silently dropping to a smaller quant. Measured 2026-07-10 · harness 2.0.0.

NVIDIA B200 specifications

Architecture	Blackwell
VRAM	192GB HBM3e
Memory bus	8192-bit
Memory bandwidth	8 TB/s
TDP	1000 W
Process	TSMC 4NP
Interface	SXM6
Release date	2025-02-01
Launch MSRP	$40,000

Verdict, nothing in our suite slows it down

NVIDIA B200 scores 78.0/100, #2 of 102. It ran all 12 workloads. Every figure here is our own measurement.

Relative performance: where the NVIDIA B200 lands

100% = this card, AI & Machine Learning headline metric (AI Score). #2 of 21 desktop cards in this vertical.

GPU	%	AI Score
NVIDIA B300	120%	93.8
NVIDIA B200	100%	78
NVIDIA B100	86%	67
NVIDIA H100 NVL	86%	67
NVIDIA GH200 Grace Hopper	84%	65.6
NVIDIA H200	83%	65

One tier up: NVIDIA B300 (+20%) · compare head-to-head
One tier down: NVIDIA B100 (-14%)

← All AI & Machine Learning GPU rankings

What this card can build

Whole-job timings, composed from our measured per-model results on this card.

Workflow	Time	Energy	Basis
50-image depth pass	6 s	0.15 Wh	measured
500-image masking run	85 s	4.9 Wh	measured
24-frame storyboard	89 s	12.85 Wh	all 2 stages measured
60-second AI short film	1.9 min	16.03 Wh	all 3 stages measured
6-panel comic page	2.3 min	23.05 Wh	all 3 stages measured
Character sheet, 12 poses	2.9 min	32.04 Wh	all 2 stages measured
10 short social clips	5.6 min	69.91 Wh	all 3 stages measured
Full codebase review	6.6 min	44.06 Wh	measured
40-product photo shoot	8.4 min	113.44 Wh	all 2 stages measured
20 long-form articles	10.5 min	67.6 Wh	measured
40-product shoot, start to finish	12.1 min	129.28 Wh	all 4 stages measured
200-product catalogue cutout	17.2 min	79.19 Wh	all 2 stages measured
100-photo restoration batch	17.8 min	257.65 Wh	measured
100-photo restore and enlarge	26.4 min	296.86 Wh	all 2 stages measured

Rent or buy?

This card is $40,000 to buy. The cheapest listed rate on RunPod is $5.980/hour, but that is the floor: we budget $7.176/hour, a 20% premium, because idle time, storage and unavailable cheap instances all land on the same bill. At that rate buying wins after 5,574 GPU-hours. Below it you are paying for idle silicon.

How you would use it	GPU-hours a year	Rental cost a year	Time to break even
2 hours a day, hobby	730	$5,238	7.6 years
8 hours a day, working on it	2,920	$20,954	1.9 years
24/7, always-on agent	8,760	$62,862	7.6 months

At hobby usage this card is very unlikely to pay for itself before it is superseded. Rent it. Rental figures include a 20% premium over the cheapest listed rate. Ignores electricity, resale and the fact that a rented card can be a newer one tomorrow.

Rental price

$5.980/hr+1.5% since 2026-08-14low $5.890 · high $5.980

Cheapest of the RunPod and Vast on-demand rates we see, sampled daily. Spot and interruptible pricing runs lower.