NVIDIA B300, AI & Machine Learning Benchmarks & Specs

288GB · AI Score 93.8/100 · first-party measured on 12 AI workloads

93.8 AI Score ✓ Measured

Every number on this page is first-party: NVIDIA B300 was run on our pinned 12-workload AI suite on 2026-07-12, with under 0.5% run-to-run variance. On Llama 3.1 8B (Q4_K_M) NVIDIA B300 delivers about 287.23 tokens/sec. Stepping up to Qwen3 32B it holds roughly 83.68 tok/s. The full Llama 3.3 70B still runs, at about 47.97 tok/s. For image generation, SDXL runs at 14.6 it/s, and FLUX.1-dev at 9.71 it/s. All 12 workloads fit in 288GB. There is no model in our suite this card has to turn down. NVIDIA B300 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

LFM2.5-1.2B955.13

gemma-3-270m926.67

Qwen1.5-0.5B894.84

Llama 3.2 1B889.19

SmolLM2-135M855.87

Qwen2.5-Coder-0.5B843.44

Qwen2.5-0.5B843.4

Qwen3-0.6B742.64

Qwen3 0.6B718.58

LFM2.5-8B-A1B596.16

Qwen2.5-1.5B584.96

Qwen2.5-Coder-1.5B584.5

Workload	Result	Telemetry	Data
LFM2.5-1.2B	955.13 tok/s	259 W38°CQ4_K_M	✓ Measured
gemma-3-270m	926.67 tok/s	242 W33°CQ4_K_M	✓ Measured
Qwen1.5-0.5B	894.84 tok/s	246 W34°CQ4_K_M	✓ Measured
Llama 3.2 1B	889.19 tok/s	263 W44°CQ4_K_M	✓ Measured
SmolLM2-135M	855.87 tok/s	244 W34°CQ4_K_M	✓ Measured
Qwen2.5-Coder-0.5B	843.44 tok/s	244 W33°CQ4_K_M	✓ Measured
Qwen2.5-0.5B	843.4 tok/s	252 W33°CQ4_K_M	✓ Measured
Qwen3-0.6B	742.64 tok/s	249 W34°CQ4_K_M	✓ Measured
Qwen3 0.6B	718.58 tok/s	223 W42°CQ4_K_M	✓ Measured
LFM2.5-8B-A1B	596.16 tok/s	277 W36°CQ4_K_M	✓ Measured
Qwen2.5-1.5B	584.96 tok/s	267 W35°CQ4_K_M	✓ Measured
Qwen2.5-Coder-1.5B	584.5 tok/s	265 W37°CQ4_K_M	✓ Measured
Qwen2-1.5B	583.79 tok/s	265 W36°CQ4_K_M	✓ Measured
Qwen3-1.7B	575.97 tok/s	274 W37°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill 1.5B	562.14 tok/s	274 W44°CQ4_K_M	✓ Measured
Qwen3 1.7B	555.1 tok/s	275 W44°CQ4_K_M	✓ Measured
gemma-3-1b	533.32 tok/s	271 W35°CQ4_K_M	✓ Measured
Llama-3.2-3B-Instruct-uncensored	448.94 tok/s	323 W39°CQ4_K_M	✓ Measured
Hermes-3-Llama-3.2-3B	446.78 tok/s	257 W37°CQ4_K_M	✓ Measured
Llama 3.2 3B	432.24 tok/s	268 W46°CQ4_K_M	✓ Measured
SmolLM3-3B	420.3 tok/s	321 W37°CQ4_K_M	✓ Measured
gemma-2-2b-it-abliterated	418.18 tok/s	305 W37°CQ4_K_M	✓ Measured
gemma-2-2b	417.84 tok/s	307 W36°CQ4_K_M	✓ Measured
SmolLM3 3B	411.25 tok/s	278 W46°CQ4_K_M	✓ Measured
Qwen2.5-Coder-3B	410.06 tok/s	317 W38°CQ4_K_M	✓ Measured
Phi-4-mini	409.58 tok/s	324 W39°CQ4_K_M	✓ Measured
Qwen2.5-3B	409.56 tok/s	300 W37°CQ4_K_M	✓ Measured
Phi-4 Mini 3.8B	398.16 tok/s	301 W47°CQ4_K_M	✓ Measured
AI21-Jamba-Reasoning-3B	380.88 tok/s	311 W37°CQ4_K_M	✓ Measured
phi-2	356.65 tok/s	316 W37°CQ4_K_M	✓ Measured
gpt-oss-20b	355.56 tok/s	279 W37°CQ4_K_M	✓ Measured
Phi-3.5-mini	352.25 tok/s	334 W37°CQ4_K_M	✓ Measured
Nemotron-3-Nano-30B-A3B	345.31 tok/s	280 W37°CQ4_K_M	✓ Measured
DeepSeek-Coder-V2-Lite	340.53 tok/s	279 W37°CQ4_K_M	✓ Measured
Qwen3-4B-Instruct-2507	339.78 tok/s	325 W38°CQ4_K_M	✓ Measured
Qwen3-4B-Thinking-2507	339.74 tok/s	325 W37°CQ4_K_M	✓ Measured
Qwen3-4B-Instruct-2507	339.72 tok/s	313 W38°CQ4_K_M	✓ Measured
Qwen3 4B	333.34 tok/s	3.1 GB peak295 W39°C1.13 tok/WQ4_K_M	✓ Measured
Llama-2-7B	315.9 tok/s	379 W40°CQ4_K_M	✓ Measured
Mistral-7B-Instruct-v0.2	306.79 tok/s	377 W40°CQ4_K_M	✓ Measured
Mistral-7B-Instruct-v0.3	306.46 tok/s	366 W40°CQ4_K_M	✓ Measured
Mistral-7B-Instruct-v0.1	306.44 tok/s	372 W40°CQ4_K_M	✓ Measured
Gemma 3 4B	301.07 tok/s	283 W46°CQ4_K_M	✓ Measured
Mistral 7B v0.3	298.95 tok/s	313 W48°CQ4_K_M	✓ Measured
Qwen2.5-7B	293.63 tok/s	370 W39°CQ4_K_M	✓ Measured
L3-8B-Stheno-v3.2	293 tok/s	381 W40°CQ4_K_M	✓ Measured
Meta-Llama-3.1-8B	293 tok/s	352 W40°CQ4_K_M	✓ Measured
DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored	292.55 tok/s	361 W41°CQ4_K_M	✓ Measured
Qwen2.5-Coder-7B-Instruct-abliterated	292.21 tok/s	370 W39°CQ4_K_M	✓ Measured
Qwen3-30B-A3B	291.82 tok/s	287 W36°CQ4_K_M	✓ Measured
dolphin-2.9-llama3-8b	290.47 tok/s	354 W40°CQ4_K_M	✓ Measured
Llama 3.1 8B	287.23 tok/s	5.3 GB peak338 W41°C0.85 tok/WQ4_K_M	✓ Measured
Dolphin X1 8B	287.02 tok/s	292 W40°CQ4_K_M	✓ Measured
Qwen2.5-Coder 7B	286.5 tok/s	299 W48°CQ4_K_M	✓ Measured
Dolphin 3.0 Llama 3.1 8B	285.88 tok/s	300 W40°CQ4_K_M	✓ Measured
Qwen3-Coder 30B A3B	284.88 tok/s	256 W36°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill 7B	284.77 tok/s	291 W42°CQ4_K_M	✓ Measured
Qwen3 30B A3B	284.66 tok/s	250 W39°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill Llama 8B	283.77 tok/s	306 W42°CQ4_K_M	✓ Measured
DeepSeek-R1-0528-Qwen3-8B	270.35 tok/s	375 W40°CQ4_K_M	✓ Measured
Qwen3-8B	270.35 tok/s	383 W41°CQ4_K_M	✓ Measured
Josiefied-Qwen3-8B-abliterated-v1	270.09 tok/s	362 W40°CQ4_K_M	✓ Measured
Qwen3 30B A3B (Q3_K_M)	268.71 tok/s	302 W37°CQ3_K_M	✓ Measured
Qwen3 8B	261.4 tok/s	275 W41°CQ4_K_M	✓ Measured
Dolphin X1 Trinity Nano 6B	257.93 tok/s	249 W33°CQ4_K_M	✓ Measured
Ornith-1.0-9B	240.67 tok/s	377 W40°CQ4_K_M	✓ Measured
KAT-Coder-V2.5-Dev	239.66 tok/s	287 W36°CQ4_K_M	✓ Measured
Qwen-AgentWorld-35B-A3B	233.04 tok/s	293 W36°CQ4_K_M	✓ Measured
Ornith-1.0-35B	232.36 tok/s	291 W36°CQ4_K_M	✓ Measured
gemma-2-9b	202.22 tok/s	386 W40°CQ4_K_M	✓ Measured
Qwen3-Next-80B-A3B-Thinking	200.55 tok/s	287 W36°CQ4_K_M	✓ Measured
Qwen3-Coder-Next	199.39 tok/s	288 W36°CQ4_K_M	✓ Measured
Qwen3-Coder-Next-abliterated	199.12 tok/s	280 W36°CQ4_K_M	✓ Measured
NemoMix-Unleashed-12B	197.06 tok/s	403 W41°CQ4_K_M	✓ Measured
Mistral-Nemo-Instruct-2407	196.92 tok/s	406 W41°CQ4_K_M	✓ Measured
GLM-4.7-Flash	196.52 tok/s	303 W35°CQ4_K_M	✓ Measured
Qwen3-Coder-Next	196.34 tok/s	292 W36°CQ4_K_M	✓ Measured
Qwen3-Next-80B-A3B-Thinking	194.98 tok/s	292 W36°CQ4_K_M	✓ Measured
Qwen3-Next-80B-A3B	191.32 tok/s	284 W36°CQ4_K_M	✓ Measured
GLM-4.7-Flash-REAP-23B-A3B	178.49 tok/s	300 W36°CQ4_K_M	✓ Measured
Phi-4 14B	177.41 tok/s	334 W44°CQ4_K_M	✓ Measured
Qwen3-14B	169.52 tok/s	422 W41°CQ4_K_M	✓ Measured
Hermes-4-14B	169.51 tok/s	422 W41°CQ4_K_M	✓ Measured
Qwen3 14B	165.44 tok/s	315 W43°CQ4_K_M	✓ Measured
Phi-4 14B (Q3_K_M)	165.06 tok/s	439 W42°CQ3_K_M	✓ Measured
Gemma 3 12B	161.27 tok/s	315 W42°CQ4_K_M	✓ Measured
Qwen2.5-Coder-14B-Instruct-abliterated	159.5 tok/s	423 W40°CQ4_K_M	✓ Measured
Uncensored	159.48 tok/s	416 W41°CQ4_K_M	✓ Measured
EVA-Qwen2.5-14B-v0.2	159.46 tok/s	416 W41°CQ4_K_M	✓ Measured
Qwen2.5-14B	159.33 tok/s	424 W40°CQ4_K_M	✓ Measured
Qwen2.5-Coder 14B	158.49 tok/s	9 GB peak367 W42°C0.43 tok/WQ4_K_M	✓ Measured
DeepSeek-R1 Distill 14B	156.33 tok/s	328 W43°CQ4_K_M	✓ Measured
Gemma 4 12B	155.96 tok/s	314 W42°CQ4_K_M	✓ Measured
Gemma 3 12B (Q3_K_M)	149.61 tok/s	412 W40°CQ3_K_M	✓ Measured
DeepSeek-R1 Distill 14B (Q3_K_M)	144.74 tok/s	424 W41°CQ3_K_M	✓ Measured
StarCoder2 15B	144.05 tok/s	330 W41°CQ4_K_M	✓ Measured
Cydonia-24B-v4.3	121.9 tok/s	477 W43°CQ4_K_M	✓ Measured
Dolphin-Mistral-24B-Venice-Edition	121.87 tok/s	473 W43°CQ4_K_M	✓ Measured
Devstral Small 24B	121.18 tok/s	323 W42°CQ4_K_M	✓ Measured
Dolphin 3.0 R1 Mistral 24B	120.95 tok/s	348 W43°CQ4_K_M	✓ Measured
Dolphin Mistral 24B Venice	120.93 tok/s	331 W43°CQ4_K_M	✓ Measured
Codestral 22B	119.81 tok/s	328 W42°CQ4_K_M	✓ Measured
Mistral Small 24B	119.3 tok/s	345 W45°CQ4_K_M	✓ Measured
Mistral Small 24B (Q3_K_M)	108.37 tok/s	460 W43°CQ3_K_M	✓ Measured
Codestral 22B (Q3_K_M)	107.98 tok/s	465 W43°CQ3_K_M	✓ Measured
Gemma 3 27B	93.44 tok/s	326 W42°CQ4_K_M	✓ Measured
Olmo-3.1-32B-Think	87.85 tok/s	476 W43°CQ4_K_M	✓ Measured
Dolphin 2.9.1 Yi 1.5 34B	85.15 tok/s	312 W43°CQ4_K_M	✓ Measured
Qwen3 32B	83.68 tok/s	19.1 GB peak346 W44°C0.24 tok/WQ4_K_M	✓ Measured
Qwen2.5-32B	83.41 tok/s	490 W43°CQ4_K_M	✓ Measured
DeepSeek-R1-Distill-Qwen-32B-abliterated	83.19 tok/s	468 W43°CQ4_K_M	✓ Measured
QwQ 32B	82.91 tok/s	328 W43°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill 32B	82.9 tok/s	342 W43°CQ4_K_M	✓ Measured
Qwen2.5-Coder 32B	82.9 tok/s	316 W43°CQ4_K_M	✓ Measured
Qwen2.5-Coder 32B (Q3_K_M)	74.28 tok/s	482 W43°CQ3_K_M	✓ Measured
Qwen2.5-72B	48.28 tok/s	553 W45°CQ4_K_M	✓ Measured
Meta-Llama-3.1-70B	48.12 tok/s	550 W45°CQ4_K_M	✓ Measured
DeepSeek-R1-Distill-Llama-70B	48.11 tok/s	551 W46°CQ4_K_M	✓ Measured
Hermes-4-70B	48.1 tok/s	529 W46°CQ4_K_M	✓ Measured
Llama-3.3-70B-Instruct-abliterated	48.1 tok/s	545 W46°CQ4_K_M	✓ Measured
Llama 3.3 70B	47.97 tok/s	40.2 GB peak378 W46°C0.13 tok/WQ4_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 5

FLUX.1 Schnell59.54

Z-Image Turbo (1024px)48.63

Z-Image Turbo39.675

Stable Diffusion XL29.2

FLUX.1 dev20.807

Workload	Result	Telemetry	Data
FLUX.1 Schnell	59.54 images/min	619 W45°C	✓ Measured
Z-Image Turbo (1024px)	48.63 images/min	776 W46°C	✓ Measured
Z-Image Turbo	39.68 images/min	26.1 GB peak932 W50°C1.5 s/img	✓ Measured
Stable Diffusion XL	29.2 images/min	16.5 GB peak488 W41°C2.1 s/img	✓ Measured
FLUX.1 dev	20.81 images/min	37 GB peak989 W54°C2.9 s/img	✓ Measured

Fine-Tuning train tok/s 4

TinyLlama 1.1B LoRA26938.2

SmolLM2 1.7B LoRA26114.3

Qwen2.5 1.5B LoRA22351.1

Qwen2.5 7B LoRA14323.1

Workload	Result	Telemetry	Data
TinyLlama 1.1B LoRA	26938.2 train tok/s	323 W37°C	✓ Measured
SmolLM2 1.7B LoRA	26114.3 train tok/s	404 W40°C	✓ Measured
Qwen2.5 1.5B LoRA	22351.1 train tok/s	371 W37°C	✓ Measured
Qwen2.5 7B LoRA	14323.1 train tok/s	723 W45°C	✓ Measured

Image Editing images/min 2

Workload	Result	Telemetry	Data
FLUX.1 Kontext dev	10.33 images/min	35.8 GB peak1030 W56°C5.8 s/img	✓ Measured
Qwen-Image-Edit	8.14 images/min	60.5 GB peak1020 W55°C7.4 s/img	✓ Measured

Video Generation frames/s 2

Workload	Result	Telemetry	Data
LTX-Video (distilled)	31.87 frames/s	60.5 GB peak716 W51°C3.2 s/clip	✓ Measured
Wan 2.2 5B (720p)	2.94 frames/s	37 GB peak1013 W57°C16.7 s/clip	✓ Measured

Depth Estimation images/min 2

Workload	Result	Telemetry	Data
Depth Anything V2 Small	1491.6 images/min	235 W35°C	✓ Measured
Depth Anything V2 Large	1069.95 images/min	235 W35°C	✓ Measured

Segmentation images/min 2

Workload	Result	Telemetry	Data
SAM ViT-Base	1719.59 images/min	235 W35°C	✓ Measured
SAM ViT-Huge	399.37 images/min	278 W40°C	✓ Measured

Background Removal images/min 1

Workload	Result	Telemetry	Data
BiRefNet	1064.73 images/min	240 W35°C	✓ Measured

Upscaling images/min 1

Workload	Result	Telemetry	Data
Swin2SR 4x Upscaler	58.29 images/min	437 W42°C	✓ Measured

Extended workloads

Everything measured on this card beyond the standard 12-workload suite, grouped by the kind of work. Each group carries one unit, so numbers inside a group compare and numbers across groups do not.

Text Generation (tok/s)

GPT-OSS 20B346.76

DeepSeek-R1 Distill 8B289.05

Phi-4 14B178.74

Qwen3 14B168.03

Mistral Small 24B120.55

Gemma 3 27B93.31

Workload	Result	Telemetry
GPT-OSS 20B	346.76 tok/s	11.9 GB peak286 W40°C1.22 tok/WQ4_K_M
DeepSeek-R1 Distill 8B	289.05 tok/s	5.3 GB peak348 W42°C0.83 tok/WQ4_K_M
Phi-4 14B	178.74 tok/s	9.2 GB peak392 W44°C0.46 tok/WQ4_K_M
Qwen3 14B	168.03 tok/s	9 GB peak366 W43°C0.46 tok/WQ4_K_M
Mistral Small 24B	120.55 tok/s	14 GB peak352 W44°C0.34 tok/WQ4_K_M
Gemma 3 27B	93.31 tok/s	16.9 GB peak353 W43°C0.27 tok/WQ4_K_M

Image Generation (it/s)

Stable Diffusion 1.533.11

Qwen-Image9.08

FLUX.1 schnell7.75

Workload	Result	Telemetry
Stable Diffusion 1.5	33.11 it/s	26.7 GB peak361 W38°C0.9 s/img
Qwen-Image	9.08 it/s	60.3 GB peak945 W50°C3.3 s/img
FLUX.1 schnell	7.75 it/s	37 GB peak890 W48°C0.5 s/img

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 2026-07-12 · harness 2.1.0-b300-cuda13.

NVIDIA B300 specifications

Architecture	Blackwell Ultra
CUDA cores	20,480
VRAM	288GB HBM3e
Memory bus	8192-bit
Memory bandwidth	8 TB/s
TDP	1400 W
Process	TSMC 4NP
Interface	SXM6
Release date	2025-11-01
Launch MSRP	$40,000

Verdict, nothing in our suite slows it down

NVIDIA B300 scores 93.8/100, #1 of 102. It ran all 12 workloads. Every figure here is our own measurement.

Relative performance: where the NVIDIA B300 lands

100% = this card, AI & Machine Learning headline metric (AI Score). #1 of 21 desktop cards in this vertical.

GPU	%	AI Score
NVIDIA B300	100%	93.8
NVIDIA B200	83%	78
NVIDIA B100	71%	67
NVIDIA H100 NVL	71%	67
NVIDIA GH200 Grace Hopper	70%	65.6

One tier down: NVIDIA B200 (-17%) · compare head-to-head

← 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	5 s	0.18 Wh	measured
500-image masking run	78 s	5.8 Wh	measured
24-frame storyboard	1.5 min	11 Wh	all 2 stages measured
6-panel comic page	2 min	15.52 Wh	all 3 stages measured
60-second AI short film	2.2 min	14.29 Wh	all 3 stages measured
Character sheet, 12 poses	2.2 min	20.73 Wh	all 2 stages measured
200-product catalogue cutout	3.7 min	25.74 Wh	all 2 stages measured
10 short social clips	4.7 min	51.37 Wh	all 3 stages measured
40-product photo shoot	6.1 min	77.6 Wh	all 2 stages measured
Full codebase review	6.3 min	38.61 Wh	measured
40-product shoot, start to finish	6.9 min	82.75 Wh	all 4 stages measured
20 long-form articles	9.7 min	61.26 Wh	measured
100-photo restoration batch	10.2 min	166.12 Wh	measured
100-photo restore and enlarge	11.9 min	178.61 Wh	all 2 stages measured

Rent or buy?

This card is $40,000 to buy. The cheapest listed rate on RunPod is $6.940/hour, but that is the floor: we budget $8.328/hour, a 20% premium, because idle time, storage and unavailable cheap instances all land on the same bill. At that rate buying wins after 4,803 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	$6,079	6.6 years
8 hours a day, working on it	2,920	$24,318	1.6 years
24/7, always-on agent	8,760	$72,953	6.6 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

$6.940/hr+0.0% since 2026-08-14low $6.940 · high $6.940

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