GPUs / NVIDIA

RTX 5090

Memory
GDDR7
Bus
512-bit
Published peak
1,792 GB/s
Capacity
32 GB

Measured

Read
1722GB/s
Write
1479GB/s
Copy
1394GB/s
Random
188.9GB/s
Latency
289ns
Sustain
1373GB/s

Medians of 34 verified runs from 13 operators. Read reaches 96 % of the published peak.

105.234 verified runs, median 105.5105.9

Local AI speed

ModelParameters4-bit, tokens/s8-bit, tokens/s
Llama 3.1 8B8.03B≤ 429≤ 214
Qwen 2.5 14B14.7B≤ 234≤ 117
gpt-oss-20b21B, 3.6B active≤ 957≤ 478
Mistral Small 24B24B≤ 143≤ 72
Qwen 2.5 32B32.5B≤ 106Does not fit
Llama 3.3 70B70.6BDoes not fitDoes not fit
gpt-oss-120b117B, 5.1B activeDoes not fitDoes not fit
Upper bounds: median read ÷ bytes read per token. Mixture-of-experts models read only their active experts. Fit checks weights against the 32 GB configuration; the KV cache needs room on top.

Top operators

#OperatorScoreReadDate
1OP-AAC8AC105.91762 GB/s2026-09-27
2OP-0E9B41105.71736 GB/s2026-09-27
3OP-B92C59105.71735 GB/s2026-09-27
4OP-39B43F105.71750 GB/s2026-09-27
5OP-F88900105.61731 GB/s2026-09-27
6OP-5AE543105.61728 GB/s2026-09-27
7OP-A6B99E105.61722 GB/s2026-09-27
8OP-9C3CED105.51720 GB/s2026-09-27
9OP-13D76C105.51720 GB/s2026-09-27
10OP-CAA909105.51720 GB/s2026-09-27

Full RTX 5090 leaderboard