Runyard / GPUs / RX 7900 XTX

RX 7900 XTX for local AI

The RX 7900 XTX pairs 24 GB of VRAM with 960 GB/s of memory bandwidth. Those two numbers, in that order, determine everything about what it can run and how quickly.

AMD · Launched 2022-12 · 355W · $999 MSRP · Updated 2026-09-01

Of the 53 open-weight models we track, 53 fit entirely in this card's memory at some usable quantisation. The largest is Mixtral 8x22B Instruct at 39B parameters, which works because only 39B are active per token.

Models this card runs

Highest-fidelity quantisation that fits in 24 GB, with an 8K context window:

ModelParametersBest quantVRAM usedEst. speed
Mixtral 8x22B Instruct39B / 39B activeQ3_K_M21.1 GB~45 tok/s
Qwen3.6 35B A3B35BQ4_K_M23.6 GB~39 tok/s
Qwen3.5-35B-A3B35BQ4_K_M23.6 GB~39 tok/s
Laguna XS 2.133BQ4_K_M22.4 GB~41 tok/s
Qwen2.5 Coder 32B Instruct32BQ4_K_M21.8 GB~43 tok/s
Qwen3 VL 32B Instruct32BQ4_K_M21.8 GB~43 tok/s
Qwen3 32B32BQ4_K_M21.8 GB~43 tok/s
Gemma 4 31B31BQ4_K_M21.2 GB~44 tok/s
GLM 4.7 Flash30BQ4_K_M20.6 GB~46 tok/s
Qwen3 30B A3B30BQ4_K_M20.6 GB~46 tok/s
Muse Glimmer 30B30BQ4_K_M20.6 GB~46 tok/s
Qwen3 30B A3B Instruct 250730B / 3B activeQ5_K_M23.3 GB~327 tok/s
Qwen3 Coder 30B A3B Instruct30BQ4_K_M20.6 GB~46 tok/s
Qwen3 VL 30B A3B Instruct30BQ4_K_M20.6 GB~46 tok/s
Nemotron 3.5 Lightning30B / 3B activeQ5_K_M23.2 GB~359 tok/s
Nemotron 3 Nano Omni30BQ4_K_M20.6 GB~46 tok/s
Nemotron 3 Nano 30B A3B30BQ4_K_M20.6 GB~46 tok/s
Cohere North Mini Code30B / 3B activeQ5_K_M23.2 GB~359 tok/s
Qwen3.8 27B27BQ5_K_M22.8 GB~40 tok/s
Qwen3.6 27B27BQ5_K_M22.8 GB~40 tok/s
Qwen3.5-27B27BQ5_K_M22.8 GB~40 tok/s
Gemma 2 27B27BQ5_K_M22.8 GB~40 tok/s
Gemma 4 26B A4B25BQ5_K_M21.4 GB~43 tok/s
Mistral Small 3.1 24B24BQ6_K23.2 GB~39 tok/s
Voxtral Small 24B 250724BQ6_K23.2 GB~39 tok/s

Why bandwidth matters more than you think

Generating a token requires reading the model's active weights out of memory once. That makes decoding a bandwidth problem, not a compute problem — which is why a card's headline TFLOPS number tells you almost nothing about how fast it will feel.

At 960 GB/s, the RX 7900 XTX can stream roughly 1,365 billion Q4 parameters per second. Divide that by a model's active parameter count and you have its ceiling in tokens per second. A 7B model lands near 195 tok/s; a 30B dense model near 46 tok/s. No amount of compute changes that ratio.

The buying rule this implies. Choose capacity for the model you want to run, then bandwidth for how it will feel. Two cards with identical VRAM can differ threefold in speed, and the spec sheet buries the number that decides it.

Power and running cost

At 355W under sustained inference load, four hours a day works out to 1.4 kWh — about $7 a month at $0.17/kWh, or ₹341 at Indian residential rates. Size the PSU with headroom: 355W on the card means a 650W supply is the sensible floor once you add a CPU and transients.

How it compares at this price

GPUMSRPVRAMBandwidthModels it runs
RX 7900 XTX$99924 GB960 GB/s53
RTX 5080$99916 GB960 GB/s38
RTX 4080 SUPER$99916 GB736 GB/s38
RTX 4070 Ti SUPER$79916 GB672 GB/s38

Read that table by column, not by row. The VRAM column tells you which models are even on the menu; the bandwidth column tells you how they will feel. Cards that look interchangeable on price routinely differ by a factor of two in one column and not the other, and which of those matters depends entirely on whether your target model fits.

Frequently asked questions

How many parameters can the RX 7900 XTX run?

At Q4_K_M, 24 GB holds roughly 40B parameters. Mixture-of-experts models change the arithmetic for speed but not for capacity — all experts still have to be resident.

Is the RX 7900 XTX good for local LLMs?

For models up to about 40B parameters at Q4, yes. Its 960 GB/s of bandwidth is what sets the speed ceiling, and its 24 GB of capacity is what sets the size ceiling. Beyond that you are offloading to system RAM and losing most of the benefit.

How much power does the RX 7900 XTX draw during inference?

Its rated board power is 355W, and sustained generation will sit near that. Inference load is steadier than gaming load, so plan for the full figure continuously rather than in bursts.

Do two GPUs double what I can run?

For capacity, largely yes — llama.cpp and vLLM both split layers across cards, so two 24 GB cards hold roughly 48 GB of model. For speed, no: layers run sequentially across the split, so you gain very little throughput on a single request.

Is more VRAM or more bandwidth better on the RX 7900 XTX?

They answer different questions. VRAM decides whether a model runs at all; bandwidth decides how fast it generates once it does. Buy enough capacity for your target model first, because no amount of bandwidth rescues a model that does not fit.

Related