In the measurement queue
NVIDIA GeForce RTX 5090
32 GB · Blackwell sm_120
We have not measured this card yet.
It is in the queue. Until then we show its spec and an estimate from our formula, clearly labelled.
An estimate, not a measurement
≈35–43t/s
before, estimate · Qwen3.8-27B Q8_0 · after: soon
0 70 t/s
- Memory
- 32GB
- Bandwidth
- 1,792GB/s
- Architecture
- Blackwell sm_120
An estimate, not a measurement
What fits
The formula η · bandwidth / bytes per token, for stock llama.cpp. Once measured, real numbers replace it here.
The formula η · bandwidth / bytes per token, for stock llama.cpp. Once measured, real numbers replace it here.app/landing/src/lib/estimator.ts: decodeRange (η·BW/B, raport 04 §2.2.4), memoryNeed (raport 04 §2.2.4)app/landing/src/lib/data/gpus.jsonapp/landing/src/lib/data/models.json- Qwen3.8-27B Q8_032 GB
memory needed, estimate ≈ 91–92 % of the card's memory
before, estimate ≈ 35–43 t/s
after soon
- LFM2.5-24B-A2B Q5_K_M32 GB
memory needed, estimate ≈ 53–54 % of the card's memory
before, estimate ≈ 312–554 t/s
after soon