Home › GPUs & Macs › Apple M3 Ultra (80-core GPU) vs NVIDIA RTX 5090
Apple M3 Ultra (80-core GPU) vs NVIDIA RTX 5090 for local LLMs
Across 5 tests measured on both, the NVIDIA RTX 5090 is faster in 4 and generates a median 1.15× the tokens/s of the Apple M3 Ultra (80-core GPU).
On Llama 2 7B (Q4_0, llama-bench) the Apple M3 Ultra (80-core GPU) generates 92.1 tokens/s and the NVIDIA RTX 5090 300 tokens/s (measured).
Same model, same quant, same test
| Model | Quant / test | M3 Ultra (80-core GPU) tok/s | RTX 5090 tok/s | Faster | Sources and run details |
|---|---|---|---|---|---|
| Llama 2 7B | Q4_0 llama-bench | 92.1 Metal | 300 CUDA | RTX 5090 +226% | llama.cpp #4167 M3 Ultra (80-core GPU) · llama.cpp #15013 RTX 5090 |
| gpt-oss 20B | MXFP4 llama-bench | 116 Metal | 283 CUDA | RTX 5090 +145% | llama.cpp #15396 M3 Ultra (80-core GPU) · llama.cpp #15396 RTX 5090 |
| Llama 3.1 8B Instruct | Q4_K_M LocalScore | 63.0 llamafile | 66.3 llamafile | RTX 5090 +5% | LocalScore M3 Ultra (80-core GPU) · LocalScore RTX 5090 |
| Llama 3.2 1B Instruct | Q4_K_M LocalScore | 178 llamafile | 120 llamafile | M3 Ultra (80-core GPU) +49% | LocalScore M3 Ultra (80-core GPU) · LocalScore RTX 5090 |
| Qwen2.5 14B Instruct | Q4_K_M LocalScore | 36.3 llamafile | 41.8 llamafile | RTX 5090 +15% | LocalScore M3 Ultra (80-core GPU) · LocalScore RTX 5090 |
Generation speed in tokens/s, best measured backend for each chip (shown under the number). Backends can differ between the two chips; check them before drawing conclusions. Measured results only, nothing estimated.
Specs
| Apple M3 Ultra (80-core GPU) | NVIDIA RTX 5090 | |
|---|---|---|
| Memory | Unified memory | 32 GB |
| Memory bandwidth | 819 GB/s | 1792 GB/s |
All results for the Apple M3 Ultra (80-core GPU) · All results for the NVIDIA RTX 5090 · More comparisons