Head to head · real speed
The Arena: a race between graphics cards
Two to four graphics cards work through the same job, with the same model, each at its real speed. Nothing is sped up or staged: what you see taking time is what it takes.
1 The job
Same model, same quantisation and same job for all of them.
4 bits = the sweet spot for consumer cards · 8 bits = almost no loss, twice the memory · 16 bits = not quantised
2 The contenders
Two cards minimum, four maximum. Start from a ready-made line-up if you like.
Pick your cards and hit Fight!
The scoreboard
How much work comes out of each card
«People» = how many people working an eight-hour day would do the same work the card does in 24 h. «€ of card» spreads the price of the card across its daily capacity.
What is measured and what is estimated
The RTX 3090 Ti carries the MEASURED IN THE LAB label: its figures come from actually running Qwen3.8 27B · quantised to 4 bits (UD-Q4_K_XL, 17.55 GB) on that card (1,158 tokens/s reading, 39.7 tokens/s writing, 21 September 2026). Other cards carry MEASURED when there is a public benchmark with its source linked.
Changing the quantisation changes everything: the same model in 8 bits takes almost twice the memory and runs slower, and unquantised it does not fit in nearly any consumer card. That is why a model is always named here with its exact quantisation.
Everything else carries ESTIMATED: worked out with a physical model from each card's memory bandwidth and compute. They are good for comparing orders of magnitude, not for showing off decimal places. A race between an estimate and a measurement compares what is expected with what was actually seen.
How to read the race
Each lane is a graphics card doing the same job with the same model and the same quantisation. The time you see is the time it would really take: the ×4 and ×10 buttons only speed up the playback, not the maths. When it ends, the table below shows how much work each card gets through in 24 hours and how many people that equals.
It helps you see, before buying, whether the price gap between two cards shows up in the daily work. The RTX 3090 Ti is measured in our lab; the rest are estimated from each card's memory bandwidth and compute. How every figure is worked out: Method.