GPUs / NVIDIA
RTX 3090
- Memory
- GDDR6X
- Bus
- 384-bit
- Published peak
- 936 GB/s
- Capacity
- 24 GB
Measured
- Read
- 831.1GB/s
- Write
- 809.8GB/s
- Copy
- 812.4GB/s
- Random
- 69.7GB/s
- Latency
- 270ns
- Sustain
- 812.4GB/s
Medians of 1 verified run from 1 operator. Read reaches 89 % of the published peak.
Local AI speed
| Model | Parameters | 4-bit, tokens/s | 8-bit, tokens/s |
|---|---|---|---|
| Llama 3.1 8B | 8.03B | ≤ 207 | ≤ 103 |
| Qwen 2.5 14B | 14.7B | ≤ 113 | ≤ 57 |
| gpt-oss-20b | 21B, 3.6B active | ≤ 462 | ≤ 231 |
| Mistral Small 24B | 24B | ≤ 69 | ≤ 35 |
| Qwen 2.5 32B | 32.5B | ≤ 51 | Does not fit |
| Llama 3.3 70B | 70.6B | Does not fit | Does not fit |
| gpt-oss-120b | 117B, 5.1B active | Does not fit | Does not fit |