Radeon RX 7900 XT Price & Used Prices: Which Local LLMs Can It Run?
Price alert
We'll email you when the eBay (US) price reaches your target.
AMD's RX 7900 XT (2022) is a gaming card whose 20GB just fits 27B-class at 4-bit, about 35 tokens/s. Don't confuse it with the 16GB RX 7900 GRE.
eBay
Radeon RX 7900 XT price history
eBay
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Full Radeon RX 7900 XT specifications
- Memory20 GB · GDDR6
- Bandwidth800 GB/s
- FP16 / BF16103 TFLOPS
- INT8103 TOPS
- Bus width
- 320-bit
- Architecture
- RDNA 3 (Navi 31)
- Cores
- 5,376
- Tensor cores
- 168
- Ecosystem
- ROCm
- Interface
- PCIe 4.0 x16
- Power
- 315 W
- Power connector
- 2× 8-pin
- Length
- 276 mm
- Type
- Consumer
- Launch date
- Dec 13, 2022
- MSRP
- $899
Best Radeon RX 7900 XT deals on eBay
Hand-picked deals from trusted eBay sellers. Shop with confidence.
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
UsedSapphire Pulse Radeon RX 7900 XT 20GB RAM - Graphics Card - Excellent ConditionTop Rated$770Buy on eBay
UsedXFX Speedster Merc310 Radeon RX 7900 XT 20GB BLACK Graphics Card RX-79TMERTop Rated$800Buy on eBay
NewSAPPHIRE PULSE AMD RADEON RX 7900 XT GAMING OC 20GB GDDR6 GPU (11323-02)$850+ shippingBuy on eBay
UsedSapphire 11323-02-20G Pulse AMD Radeon RX 7900 XT Gaming Graphics Card with 20GB$966Buy on eBay
UsedASRock Radeon RX 7900 XT 20GB Phantom Gaming GPU | 1yr Warranty, Fast Ship!Top Rated$990Buy on eBay
New99%NEW Gigabyte Radeon RX 7900 XT GAMING OC 20GB Graphic Card GPU AMD Radeon$1,049Buy on eBay
What LLMs can a single Radeon RX 7900 XT run?
One 7900 XT at 4-bit suits 27B-class best, with about 58K of context; Gemma 4 31B leaves only about 15K. For long context, run 30B-class MoE at 2-bit.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 3.8 GB | Runs | 129.5 | 128K | |
| 4.1 GB | Runs | 120.5 | 256K | |
| 4.7 GB | Runs | 105.9 | 256K | |
| 10.5 GB | Runs | 131.9 | 256K | |
| 11.8 GB | Runs | 47.5 | 128K | |
| 9.8 GB | Runs | 56.1 | 157K | |
| 11.9 GB | Runs | 143 | 153K | |
| 11.8 GB | Runs | 45.8 | 92K | |
| 12.3 GB | Runs | 153.7 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 5 GB | Runs | 103.3 | 128K | |
| 5.7 GB | Runs | 91.7 | 256K | |
| 7.1 GB | Runs | 74.9 | 256K | |
| 16.9 GB | Runs | 109.2 | 156K | |
| 16.8 GB | Runs | 34.2 | 53K | |
| 16.5 GB | Runs | 34.8 | 58K | |
| 18.3 GB | Runs | 124.4 | 37K | |
| 18.3 GB | Runs | 30.8 | 15K | |
| 22.1 GB | Needs RAM offload | 51.3 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 8.2 GB | Runs | 67.1 | 128K | |
| 9.5 GB | Runs | 58.5 | 256K | |
| 12.7 GB | Runs | 44.5 | 256K | |
| 26.9 GB | Needs RAM offload | 25.8 | 256K | |
| 28.6 GB | Too big | — | — | |
| 29 GB | Too big | — | — | |
| 31.8 GB | Needs RAM offload | 22.6 | 198K | |
| 32.6 GB | Too big | — | — | |
| 36.9 GB | Needs RAM offload | 34.2 | 256K |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Needs RAM offloadMoE only: expert weights sit in system RAM, so speed is estimated from 70 GB/s RAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Running Qwen3.8 27B and Qwen3.6 27B at 4-bit
About 34 to 35 tokens/s, with more than 50K of context
- Running MoE models like Gemma 4 26B A4B and GLM 4.7 Flash
About 109 to 124 tokens/s at 4-bit
- You want official ROCm support
It uses the same chip as the 7900 XTX, with exactly the same Linux, Windows and vLLM support
Not a good fit
- Running dense models above 30B, like Gemma 4 31B
It fits at 4-bit, but with only about 15K of context
- Fitting Qwen3.6 35B A3B entirely in VRAM
At 4-bit it needs 22.1 GB, more than 20GB, and with its experts in system RAM it runs at about 51 tokens/s
Before you buy
Plan for context separately
Whether this card is enough usually depends not on the model itself but on context: once a 27B model is loaded at 4-bit, just over 3GB is left for context. If you need longer context, use llama.cpp's -ctk / -ctv options to store the context cache as q8_0, which roughly halves its footprint, or switch to a 2-bit model.
Don't confuse it with the 7900 GRE
The RX 7900 GRE has a similar name and the same chip, but only 16GB, which can't hold a 27B model at 4-bit; before buying, confirm it is the 20GB 7900 XT. The reference 7900 XT is 276 mm long, takes 2.5 slots and draws 315W, and AMD recommends a 750W power supply.
Software
ROCm officially supports it on both Linux and Windows, and Ollama and llama.cpp recognize it directly, just as with the XTX. 20GB is most awkward in vLLM: AMD cards can't use AWQ or GPTQ, and the original BF16 version of a 27B model needs more than 50GB, so in practice this card mostly runs GGUF.
Compared with similar GPUs
Radeon RX 7900 XT vs Radeon RX 7900 XTX, GeForce RTX 3090, and GeForce RTX 5070 Ti: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | Radeon RX 7900 XT | Radeon RX 7900 XTX | ||
|---|---|---|---|---|
| Memory | 20 GB | 24 GB20% more | 24 GB20% more | 16 GB20% less |
| Bandwidth | 800 GB/s | 960 GB/s20% higher | 936 GB/s17% higher | 896 GB/s12% higher |
| BF16 / FP16 compute | 103 TFLOPS | 123 TFLOPS19% higher | 142 TFLOPS38% higher | 176 TFLOPS71% higher |
| INT8 compute | 103 TOPS | 123 TOPS19% higher | 285 TOPS176% higher | 352 TOPS241% higher |
| Qwen3.8 27B speed (2-bit) | 56.1 t/s | 66.3 t/s18% faster | 64.8 t/s16% faster | 62.3 t/s11% faster |
| Largest dense model (4-bit) | 30B | 37B23% more | 37B23% more | 23B23% less |
| Power | 315 W | 355 W13% higher | 350 W11% higher | 300 W5% lower |
| GPU interconnect | — | — | NVLink 112.5 GB/s | Not supported |
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x16 |
| Released | 2022-12 | 2022-12 | 2020-09 | 2025-02 |
| eBay median |
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.