NVIDIA B300 vs NVIDIA H200, AI & Machine Learning Comparison
NVIDIA B300 wins 125 of 142 benchmarks, averaging 12.9% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in FLUX.1 Kontext dev, where NVIDIA B300 leads by 135% (10.329 vs 4.393 images/min); the closest fight is Qwen3 0.6B (0% apart).
Benchmark results head-to-head
| Benchmark | NVIDIA B300 | NVIDIA H200 | Difference |
|---|---|---|---|
| Qwen3 4B tok/s | 333.34 | 318.84 | +5% |
| Llama 3.1 8B tok/s | 287.23 | 268.31 | +7% |
| Qwen2.5-Coder 14B tok/s | 158.49 | 148.43 | +7% |
| Qwen3 32B tok/s | 83.68 | 76.58 | +9% |
| Llama 3.3 70B tok/s | 47.97 | 42.66 | +12% |
| Stable Diffusion XL images/min | 29.2 | 37.16 | -21% |
| Z-Image Turbo images/min | 39.675 | 23.175 | +71% |
| FLUX.1 dev images/min | 20.807 | 9.514 | +119% |
| FLUX.1 Kontext dev images/min | 10.329 | 4.393 | +135% |
| Qwen-Image-Edit images/min | 8.14 | 3.76 | +116% |
| LTX-Video (distilled) frames/s | 31.87 | 17.87 | +78% |
| Wan 2.2 5B (720p) frames/s | 2.94 | 1.41 | +109% |
| DeepSeek-R1 Distill Llama 8B tok/s | 283.77 | 265.25 | +7% |
| DeepSeek-R1 Distill 1.5B tok/s | 562.14 | 542.95 | +4% |
| DeepSeek-R1 Distill 14B tok/s | 156.33 | 146.82 | +6% |
| DeepSeek-R1 Distill 7B tok/s | 284.77 | 267.1 | +7% |
| Gemma 3 12B tok/s | 161.27 | 153.26 | +5% |
| Gemma 3 4B tok/s | 301.07 | 291.76 | +3% |
| Gemma 4 12B tok/s | 155.96 | 151.05 | +3% |
| Llama 3.2 1B tok/s | 889.19 | 875.53 | +2% |
| Llama 3.2 3B tok/s | 432.24 | 424.61 | +2% |
| Mistral 7B v0.3 tok/s | 298.95 | 278.5 | +7% |
| Mistral Small 24B tok/s | 119.3 | 107.84 | +11% |
| Phi-4 14B tok/s | 177.41 | 168.2 | +5% |
| Phi-4 Mini 3.8B tok/s | 398.16 | 392.38 | +1% |
| Qwen2.5-Coder 7B tok/s | 286.5 | 266.36 | +8% |
| Qwen3 0.6B tok/s | 718.58 | 718.67 | -0% |
| Qwen3 1.7B tok/s | 555.1 | 563.52 | -1% |
| Qwen3 14B tok/s | 165.44 | 154.51 | +7% |
| Qwen3 30B A3B tok/s | 284.66 | 292.11 | -3% |
| Qwen3 8B tok/s | 261.4 | 247.98 | +5% |
| SmolLM3 3B tok/s | 411.25 | 402.24 | +2% |
| Codestral 22B tok/s | 119.81 | 111.34 | +8% |
| DeepSeek-R1 Distill 32B tok/s | 82.9 | 75.8 | +9% |
| Devstral Small 24B tok/s | 121.18 | 109.05 | +11% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 85.15 | 76.7 | +11% |
| Dolphin Mistral 24B Venice tok/s | 120.93 | 109.14 | +11% |
| Dolphin X1 8B tok/s | 287.02 | 268.36 | +7% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 285.88 | 267.84 | +7% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 120.95 | 109.06 | +11% |
| Gemma 3 27B tok/s | 93.44 | 84.75 | +10% |
| Qwen2.5-Coder 32B tok/s | 82.9 | 75.77 | +9% |
| Qwen3-Coder 30B A3B tok/s | 284.88 | 297.9 | -4% |
| QwQ 32B tok/s | 82.91 | 75.77 | +9% |
| StarCoder2 15B tok/s | 144.05 | 136.29 | +6% |
| FLUX.1 Schnell images/min | 59.54 | 60.51 | -2% |
| Z-Image Turbo (1024px) images/min | 48.63 | 40.74 | +19% |
| BiRefNet images/min | 1064.73 | 1457.73 | -27% |
| Depth Anything V2 Large images/min | 1069.95 | 1040.81 | +3% |
| Depth Anything V2 Small images/min | 1491.6 | 1198.58 | +24% |
| SAM ViT-Base images/min | 1719.59 | 1469.34 | +17% |
| SAM ViT-Huge images/min | 399.37 | 325.19 | +23% |
| Swin2SR 4x Upscaler images/min | 58.29 | 25.85 | +125% |
| Qwen2.5 1.5B LoRA train tok/s | 22351.1 | 14931.7 | +50% |
| Qwen2.5 7B LoRA train tok/s | 14323.1 | 8855.4 | +62% |
| SmolLM2 1.7B LoRA train tok/s | 26114.3 | 17650.4 | +48% |
| TinyLlama 1.1B LoRA train tok/s | 26938.2 | 16222.4 | +66% |
| AI21-Jamba-Reasoning-3B tok/s | 380.88 | 370.66 | +3% |
| Olmo-3.1-32B-Think tok/s | 87.85 | 78.46 | +12% |
| Codestral 22B (Q3_K_M) tok/s | 107.98 | 88.03 | +23% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 121.87 | 109.2 | +12% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 292.55 | 267.82 | +9% |
| DeepSeek-Coder-V2-Lite tok/s | 340.53 | 312.1 | +9% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 270.35 | 250.61 | +8% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 48.11 | 42.76 | +13% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 144.74 | 119.66 | +21% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 83.19 | 75.71 | +10% |
| dolphin-2.9-llama3-8b tok/s | 290.47 | 267.96 | +8% |
| Dolphin X1 Trinity Nano 6B tok/s | 257.93 | 259.81 | -1% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 159.46 | 148.74 | +7% |
| gemma-2-2b-it-abliterated tok/s | 418.18 | 397.43 | +5% |
| gemma-2-2b tok/s | 417.84 | 398.48 | +5% |
| gemma-2-9b tok/s | 202.22 | 180.47 | +12% |
| Gemma 3 12B (Q3_K_M) tok/s | 149.61 | 127.04 | +18% |
| gemma-3-1b tok/s | 533.32 | 513.58 | +4% |
| gemma-3-270m tok/s | 926.67 | 967.35 | -4% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 178.49 | 171.45 | +4% |
| GLM-4.7-Flash tok/s | 196.52 | 188.57 | +4% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 270.09 | 250.88 | +8% |
| gpt-oss-20b tok/s | 355.56 | 354.23 | 0% |
| Hermes-3-Llama-3.2-3B tok/s | 446.78 | 430.79 | +4% |
| Hermes-4-70B tok/s | 48.1 | 42.77 | +12% |
| SmolLM3-3B tok/s | 420.3 | 404.27 | +4% |
| Qwen3-Coder-Next-abliterated tok/s | 199.12 | 197.14 | +1% |
| KAT-Coder-V2.5-Dev tok/s | 239.66 | 236.31 | +1% |
| L3-8B-Stheno-v3.2 tok/s | 293 | 267.92 | +9% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| LFM2.5-1.2B tok/s | 955.13 | 946.53 | +1% |
| LFM2.5-8B-A1B tok/s | 596.16 | 583.96 | +2% |
| Llama-2-7B tok/s | 315.9 | 290.81 | +9% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 448.94 | 429.78 | +4% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 48.1 | 42.75 | +13% |
| Meta-Llama-3.1-70B tok/s | 48.12 | 42.72 | +13% |
| Meta-Llama-3.1-8B tok/s | 293 | 262.93 | +11% |
| Phi-4-mini tok/s | 409.58 | 395.3 | +4% |
| Mistral-7B-Instruct-v0.1 tok/s | 306.44 | 282.64 | +8% |
| Mistral-7B-Instruct-v0.2 tok/s | 306.79 | 282.32 | +9% |
| Mistral-7B-Instruct-v0.3 tok/s | 306.46 | 282.39 | +9% |
| Mistral-Nemo-Instruct-2407 tok/s | 196.92 | 181.3 | +9% |
| Mistral Small 24B (Q3_K_M) tok/s | 108.37 | 85.1 | +27% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 197.06 | 181.53 | +9% |
| Nemotron-3-Nano-30B-A3B tok/s | 345.31 | 327.48 | +5% |
| Hermes-4-14B tok/s | 169.51 | 156.26 | +8% |
| Ornith-1.0-35B tok/s | 232.36 | 208.51 | +11% |
| Ornith-1.0-9B tok/s | 240.67 | 222.94 | +8% |
| phi-2 tok/s | 356.65 | 347.08 | +3% |
| Phi-3.5-mini tok/s | 352.25 | 348.14 | +1% |
| Phi-4 14B (Q3_K_M) tok/s | 165.06 | 139.97 | +18% |
| Qwen-AgentWorld-35B-A3B tok/s | 233.04 | 225.31 | +3% |
| Qwen3-0.6B tok/s | 742.64 | 724.67 | +2% |
| Qwen3-1.7B tok/s | 575.97 | 568.85 | +1% |
| Qwen3-14B tok/s | 169.52 | 156.3 | +8% |
| Qwen3-30B-A3B tok/s | 291.82 | 293.47 | -1% |
| Qwen3-4B-Instruct-2507 tok/s | 339.78 | 318.67 | +7% |
| Qwen3-8B tok/s | 270.35 | 251.04 | +8% |
| Qwen3-Coder-Next tok/s | 196.34 | 195.21 | +1% |
| Qwen3-Next-80B-A3B-Thinking tok/s | 194.98 | 194.82 | 0% |
| Qwen1.5-0.5B tok/s | 894.84 | 855.22 | +5% |
| Qwen2-1.5B tok/s | 583.79 | 546.47 | +7% |
| Qwen2.5-0.5B tok/s | 843.4 | 914.58 | -8% |
| Qwen2.5-1.5B tok/s | 584.96 | 549.36 | +6% |
| Uncensored tok/s | 159.48 | 148.7 | +7% |
| Qwen2.5-14B tok/s | 159.33 | 148.76 | +7% |
| Qwen2.5-32B tok/s | 83.41 | 75.76 | +10% |
| Qwen2.5-3B tok/s | 409.56 | 400.48 | +2% |
| Qwen2.5-72B tok/s | 48.28 | 43.4 | +11% |
| Qwen2.5-7B tok/s | 293.63 | 265.15 | +11% |
| Qwen2.5-Coder-0.5B tok/s | 843.44 | 920.36 | -8% |
| Qwen2.5-Coder-1.5B tok/s | 584.5 | 545.54 | +7% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 159.5 | 148.48 | +7% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 74.28 | 58.87 | +26% |
| Qwen2.5-Coder-3B tok/s | 410.06 | 401.11 | +2% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 292.21 | 270.92 | +8% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 268.71 | 246.35 | +9% |
| Qwen3-4B-Instruct-2507 tok/s | 339.72 | 318.27 | +7% |
| Qwen3-4B-Thinking-2507 tok/s | 339.74 | 318.94 | +7% |
| Qwen3-Coder-Next tok/s | 199.39 | 196.65 | +1% |
| Qwen3-Next-80B-A3B-Thinking tok/s | 200.55 | 201.12 | -0% |
| Qwen3-Next-80B-A3B tok/s | 191.32 | 192.52 | -1% |
| SmolLM2-135M tok/s | 855.87 | 905.38 | -5% |
| Cydonia-24B-v4.3 tok/s | 121.9 | 109.2 | +12% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA B300 is faster on 14 of 14; NVIDIA H200 on 0.
| Workflow | NVIDIA B300 | NVIDIA H200 | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 5 s | 6 s | NVIDIA B300 1.33x faster | $0.009 vs $0.006 |
| 500-image masking run | 78 s | 1.7 min | NVIDIA B300 1.29x faster | $0.150 vs $0.100 |
| 24-frame storyboard | 1.5 min | 3.9 min | NVIDIA B300 2.60x faster | $0.174 vs $0.234 |
| 6-panel comic page | 2 min | 3.2 min | NVIDIA B300 1.59x faster | $0.234 vs $0.192 |
| 60-second AI short film | 2.2 min | 3.9 min | NVIDIA B300 1.79x faster | $0.251 vs $0.233 |
| Character sheet, 12 poses | 2.2 min | 3.9 min | NVIDIA B300 1.75x faster | $0.257 vs $0.234 |
| 200-product catalogue cutout | 3.7 min | 8 min | NVIDIA B300 2.15x faster | $0.430 vs $0.479 |
| 10 short social clips | 4.7 min | 10 min | NVIDIA B300 2.11x faster | $0.548 vs $0.598 |
| 40-product photo shoot | 6.1 min | 11.2 min | NVIDIA B300 1.84x faster | $0.701 vs $0.668 |
| Full codebase review | 6.3 min | 6.7 min | NVIDIA B300 1.07x faster | $0.730 vs $0.403 |
| 40-product shoot, start to finish | 6.9 min | 12.9 min | NVIDIA B300 1.87x faster | $0.797 vs $0.770 |
| 20 long-form articles | 9.7 min | 10.9 min | NVIDIA B300 1.12x faster | $1.125 vs $0.655 |
| 100-photo restoration batch | 10.2 min | 23.3 min | NVIDIA B300 2.29x faster | $1.178 vs $1.396 |
| 100-photo restore and enlarge | 11.9 min | 27.2 min | NVIDIA B300 2.29x faster | $1.378 vs $1.629 |
Renting by the hour, NVIDIA H200 finishes 9 of 14 cheaper. The quicker card is not automatically the cheaper way to get the work done.
Cost to rent
| Card | Per hour |
|---|---|
| NVIDIA B300 | $6.940 |
| NVIDIA H200 | $3.590 |
NVIDIA H200 is 1.93x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA B300 | NVIDIA H200 | |
|---|---|---|
| VRAM | 288GB | 141GB |
| Architecture | Blackwell Ultra | Hopper |
| Memory bandwidth | 8 TB/s | 4800 GB/s |
| Boost clock | n/a | 1,980 MHz |
| TDP | 1400 W | 700 W |
| Launch MSRP | $40,000 | $31,000 |
| Release | 2025-11-01 | 2024-03-18 |
FAQ
- Which is better for ai & machine learning: NVIDIA B300 or NVIDIA H200?
- NVIDIA B300 performs better for ai & machine learning, winning 125 of 142 benchmarks in our suite with an average 12.9% advantage.
- What are the main hardware differences between NVIDIA B300 and NVIDIA H200?
- NVIDIA B300 has 288GB VRAM and a 1400W TDP, while NVIDIA H200 has 141GB VRAM and a 700W TDP.
- Where is the biggest performance difference between NVIDIA B300 and NVIDIA H200?
- FLUX.1 Kontext dev: NVIDIA B300 leads by roughly 135% (10.329 vs 4.393 images/min) in our testing.
NVIDIA B300 full review · NVIDIA H200 full review · All AI & Machine Learning rankings