01 / SINGLE GPU
24 GBOne RTX 3090 FE
—tokens
per second
per second
— output tokens · — seconds
Open the source record ↗THE RESULTS / 01
One model.
Two ways to run it.
What changes when a model runs on one RTX 3090 FE or a connected pair? Explore one example from the lab, then open the evidence behind it.
THE SIMPLE VIEW
A closer look at speed.
MODELGemma 4 26B-A4B ITAWQ · W4A16
MEASUREOutput speedTokens per second · one request
TEST DATE03 Oct 2026Two separate runs
Loading source records…
02 / CONNECTED PAIR
48 GB TOTALTwo RTX 3090 FE
—tokens
per second
per second
— output tokens · — seconds
Open the source record ↗READING THE RESULT
—
—
THE CONTEXT
A number is
never the whole story.
What is a token?
A token is a piece of text used by a language model. Tokens per second describe output speed for this particular request; they are not a measure of answer quality.
What does 48 GB mean?
The pair has 24 GB on each card. Software must distribute the workload across the cards; connecting them does not turn them into one GPU. The lab has two independent pairs, each in its own virtual machine.
THE EXPERT VIEW
Follow the evidence.
The two run files are the source of the values above. Their settings are shown together so differences are visible.
Test conditions and settings
| Field | One GPU | NVLink pair |
|---|---|---|
| Loading… | ||
Different models, prompts, software settings and hardware can produce different results. The comparison does not isolate the effect of NVLink, and says nothing about quality or total system energy use.
KEEP EXPLORING