GPU ABetter value
NVIDIA RTX 3090
- VRAM
- 24 GB
- Bandwidth
- 936 GB/s
- Street price
- $850
- Vendor
- nvidia
GPU B
NVIDIA RTX 4090
- VRAM
- 24 GB
- Bandwidth
- 1008 GB/s
- Street price
- $1,999
- Vendor
- nvidia
The short answer
Both GPUs run all 33 models at nearly identical speeds (33 of 33 models are a tie, 0 models go to one side by a small margin). The specs show a bandwidth difference of just 8% — not enough to matter in practice. NVIDIA RTX 3090 costs $1,149 less (135% cheaper) — NVIDIA RTX 3090 is the better value.
33 equal (both run, <20% diff)
21 too large for either
NVIDIA RTX 3090 saves you $1,149 (135% cheaper) for the same real-world performance.
Model-by-model fit
Click any row for the full breakdown. Tie% shown in the Winner column when both GPUs run the model within 20% of each other.
| Model | NVIDIA RTX 3090 | NVIDIA RTX 4090 | Winner |
|---|---|---|---|
c4ai-command-r-v01 35B 35B · command | 39 tok/s · Q4_K_M | 42 tok/s · Q4_K_M | ≈tie B +8% |
Command-R+ 104B 104B · command | Too large | Too large | — |
DeepSeek R1 Distill Llama 8B 8B · deepseek | 48 tok/s · FP16 | 52 tok/s · FP16 | ≈tie B +8% |
DeepSeek R1 Distill Qwen 14B 14.8B · deepseek | 52 tok/s · Q8_0 | 56 tok/s · Q8_0 | ≈tie B +8% |
DeepSeek R1 Distill Llama 70B 70.6B · deepseek | Too large | Too large | — |
DeepSeek R1 671B 671B · deepseek | Too large | Too large | — |
DeepSeek-V3 685B 685B · deepseek | Too large | Too large | — |
DeepSeek-V3.2 685.4B 685.4B · deepseek | Too large | Too large | — |
gemma-2-9b 9.2B · gemma | 42 tok/s · FP16 | 45 tok/s · FP16 | ≈tie B +7% |
gemma-2-27b 27.2B · gemma | 35 tok/s · Q6_K | 38 tok/s · Q6_K | ≈tie B +9% |
Llama 3.1 8B Compact 8B · llama | 43 tok/s · FP16 | 47 tok/s · FP16 | ≈tie B +9% |
CodeLlama 34B 34B · llama | 41 tok/s · Q4_K_M | 44 tok/s · Q4_K_M | ≈tie B +7% |
CodeLlama 34B 34B · llama | 41 tok/s · Q4_K_M | 44 tok/s · Q4_K_M | ≈tie B +7% |
Llama 3.3 70B 70.6B · llama | Too large | Too large | — |
Llama 3.1 70B 70.6B · llama | Too large | Too large | — |
Llama 4 Scout 17B 109B · llama | Too large | Too large | — |
Llama-4-Maverick-17B-128E 400B · llama | Too large | Too large | — |
Llama 3.1 405B 405B · llama | Too large | Too large | — |
Mistral 7B v0.1 7.25B · mistral | 53 tok/s · FP16 | 57 tok/s · FP16 | ≈tie B +8% |
Codestral 22B 22.2B · mistral | 42 tok/s · Q6_K | 46 tok/s · Q6_K | ≈tie B +10% |
Mixtral 8x7B Instruct v0.1 47B · mixtral | Too large | Too large | — |
Mistral Large 2 123B 123B · mistral | Too large | Too large | — |
Phi-4-mini 3.8B 3.8B · phi | 98 tok/s · FP16 | 106 tok/s · FP16 | ≈tie B +8% |
Phi-4 14B 14B · phi | 49 tok/s · Q8_0 | 53 tok/s · Q8_0 | ≈tie B +8% |
Qwen 2.5 1.5B 1.5B · qwen | 228 tok/s · FP16 | 246 tok/s · FP16 | ≈tie B +8% |
Qwen 2.5 3B 3.1B · qwen | 119 tok/s · FP16 | 128 tok/s · FP16 | ≈tie B +8% |
Qwen3.5-4B 4.7B · qwen | 80 tok/s · FP16 | 87 tok/s · FP16 | ≈tie B +9% |
Qwen 2.5 7B 7.6B · qwen | 51 tok/s · FP16 | 55 tok/s · FP16 | ≈tie B +8% |
Qwen 2.5 7B 7.6B · qwen | 51 tok/s · FP16 | 55 tok/s · FP16 | ≈tie B +8% |
Qwen 3 8B 8B · qwen | 43 tok/s · FP16 | 47 tok/s · FP16 | ≈tie B +9% |
Qwen3.5-9B 9.7B · qwen | 40 tok/s · FP16 | 43 tok/s · FP16 | ≈tie B +8% |
Qwen 3 32B 32B · qwen | 38 tok/s · Q4_K_M | 41 tok/s · Q4_K_M | ≈tie B +8% |
Qwen3.5-35B-A3B 36B · qwen | 38 tok/s · Q4_K_M | 41 tok/s · Q4_K_M | ≈tie B +8% |
Qwen 2.5 72B 72.7B · qwen | Too large | Too large | — |
Qwen 2.5 72B 72.7B · qwen | Too large | Too large | — |
Llama 3.2 1B 1.24B · llama | 279 tok/s · FP16 | 300 tok/s · FP16 | ≈tie B +8% |
Llama 4 Scout 17B 109B · llama | Too large | Too large | — |
DeepSeek R1 671B 671B · deepseek | Too large | Too large | — |
Gemma 3 27B 27B · gemma | 37 tok/s · Q5_K_M | 40 tok/s · Q5_K_M | ≈tie B +8% |
Qwen 3 8B 8B · qwen | 43 tok/s · FP16 | 47 tok/s · FP16 | ≈tie B +9% |
Qwen 3 32B 32B · qwen | 38 tok/s · Q4_K_M | 41 tok/s · Q4_K_M | ≈tie B +8% |
Llama 3.1 8B Compact 8B · llama | 43 tok/s · FP16 | 47 tok/s · FP16 | ≈tie B +9% |
Mixtral 8x7B Instruct v0.1 47B · mixtral | Too large | Too large | — |
Mistral Small 3.2 24B 24B · mistral | 42 tok/s · Q6_K | 45 tok/s · Q6_K | ≈tie B +7% |
Command A 111B 111B · command | Too large | Too large | — |
DeepSeek R1 0528 685B · deepseek | Too large | Too large | — |
DeepSeek-V3-0324 684.5B · deepseek | Too large | Too large | — |
DeepSeek-R1-0528-Qwen3-8B 8.2B · qwen | 50 tok/s · FP16 | 53 tok/s · FP16 | ≈tie B +6% |
Qwen3-235B-A22B-Instruct-2507 235B · qwen | Too large | Too large | — |
Qwen3-30B-A3B-Instruct-2507 30B · qwen | 48 tok/s · Q4_K_M | 52 tok/s · Q4_K_M | ≈tie B +8% |
Qwen3-4B-Instruct-2507 4B · qwen | 102 tok/s · FP16 | 110 tok/s · FP16 | ≈tie B +8% |
gemma-4-E4B-it 8B · gemma | 51 tok/s · FP16 | 55 tok/s · FP16 | ≈tie B +8% |
gemma-4-26B-A4B-it 26.5B · gemma | 38 tok/s · Q6_K | 41 tok/s · Q6_K | ≈tie B +8% |
gemma-4-31B-it 32.7B · gemma | 44 tok/s · Q4_K_M | 48 tok/s · Q4_K_M | ≈tie B +9% |
Want a different pairing? Browse all comparisons →