Radeon PRO W7900 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 Radeon PRO W7900 (2023) is a 48GB ECC workstation card running 30B-class MoE at 8-bit at about 106 tokens/s. For multi-GPU, get the Dual Slot model.
eBay
Radeon PRO W7900 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 PRO W7900 specifications
- Memory48 GB · GDDR6 ECC
- Bandwidth864 GB/s
- FP16 / BF16123 TFLOPS
- INT8123 TOPS
- Bus width
- 384-bit
- Architecture
- RDNA 3 (Navi 31)
- Cores
- 6,144
- Tensor cores
- 192
- Ecosystem
- ROCm
- Interface
- PCIe 4.0 x16
- Power
- 295 W
- Power connector
- 2× 8-pin
- Slots
- 3
- Length
- 280 mm
- Type
- Workstation
- Launch date
- Apr 13, 2023
- MSRP
- $3,999
Best Radeon PRO W7900 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.
What LLMs can a single Radeon PRO W7900 run?
One W7900 fits everything up to 35B at 8-bit. Keep MoE at 8-bit, only 20% slower than 4-bit; dense drops to about 20 tokens/s there, so use 4-bit.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 12B12B | 4.7 GB | Runs | 113.1 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 135.7 | 256K |
 Qwen3.6 27B27.8B | 11.8 GB | Runs | 51 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 60.3 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 146.5 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 49.2 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 156.9 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 12B12B | 7.1 GB | Runs | 80.2 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 113.3 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 36.8 | 256K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 37.4 | 256K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 128.4 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 33.2 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 131.4 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 12B12B | 12.7 GB | Runs | 47.8 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 90.1 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 22.2 | 256K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 21.9 | 256K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 101.7 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Runs | 19.3 | 202K |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Runs | 105.6 | 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.
What LLMs can multiple Radeon PRO W7900 cards run?
Four W7900s (192GB) run DeepSeek V4 Flash at 4-bit at about 64 tokens/s, barely slower at 8-bit; two handle Qwen3.8 Flash Next at 2-bit.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 2 cards | $7,198 | 96 GB | 106.9 | 590 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 4 cards | $14,396 | 192 GB | 86.8 | 1,180 W |
 Hy3295B (21B active) | 107.2 GB | 4 cards | $14,396 | 192 GB | 41.4 | 1,180 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 4 cards | $14,396 | 192 GB | 70.4 | 1,180 W |
| GLM 5.3744B (40B active) | 253.9 GB | 8 cards | $28,792 | 384 GB | 36.4 | 2,360 W |
| Hy4 preview770B (49B active) | 307.2 GBest. | 8 cards | $28,792 | 384 GB | 27.5 | 2,360 W |
 Muse Spark Open816B (42B active) | 311.2 GB | 8 cards | $28,792 | 384 GB | 32.5 | 2,360 W |
 Kimi K2.7 Code1,000B (32B active) | 339.5 GB | 8 cards | $28,792 | 384 GB | 43.9 | 2,360 W |
| Kimi K2.61,000B (32B active) | 340.5 GB | 8 cards | $28,792 | 384 GB | 43.8 | 2,360 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 4 cards | $14,396 | 192 GB | 90.7 | 1,180 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 4 cards | $14,396 | 192 GB | 64.3 | 1,180 W |
| Hy3295B (21B active) | 182.2 GB | 8 cards | $28,792 | 384 GB | 31 | 2,360 W |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | 8 cards | $28,792 | 384 GB | 45.6 | 2,360 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 4 cards | $14,396 | 192 GB | 62.4 | 1,180 W |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | 8 cards | $28,792 | 384 GB | 66.7 | 2,360 W |
| Hy3295B (21B active) | 317.7 GB | 8 cards | $28,792 | 384 GB | 21.4 | 2,360 W |
| GLM 5.3 Flash320B (18B active) | 341 GB | 8 cards | $28,792 | 384 GB | 29.5 | 2,360 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Running MoE models at 8-bit
Qwen3.6 35B A3B at 8-bit runs at about 106 tokens/s, only 20% slower than at 4-bit
- Long context with 30B-class models
Gemma 4 31B and Qwen3.8 27B at 4-bit can both use the full 256K, at about 33 to 37 tokens/s
- Your workstation calls for a factory card
Blower cooling, ECC VRAM, and a dual-slot version that makes four cards easy to fit
Not a good fit
- You're after single-card speed
The 4090 48GB modded card, also 48GB, runs dense models a bit over 10% faster and processes long prompts much faster still
- Your case is tight on space
The standard version takes three slots, and the dual-slot version has to be sought out separately
- You need vLLM's AWQ or GPTQ quantization
They don't work on AMD cards
Before you buy
Triple-slot and dual-slot versions
The standard W7900 is a triple-slot blower card, 280 mm long; the later W7900 Dual Slot slims it to two slots at the same power, so one workstation can hold four side by side. When buying used, check which version it is first, because several triple-slot cards take up a lot more slots.
Power supply
Two 8-pin connectors and 295W, so a 750W power supply is enough for one card (AMD's recommendation). The 4090 48GB modded card, also 48GB, draws 450W through a 12VHPWR connector; by comparison, the W7900 asks far less of your power supply and case cooling.
Software
It uses the same chip as the RX 7900 XTX with the same ROCm support, and AMD's Windows PyTorch support table even lists the W7900 and the dual-slot version separately. There's one limitation with vLLM: AWQ and GPTQ quantization don't work on AMD cards. 48GB does give you room to run original BF16 weights directly: a 14B-class model's BF16 weights are about 28GB, which still leaves room for context. For bigger models, switch to GGUF and run them with llama.cpp or Ollama.
Compared with similar GPUs
Radeon PRO W7900 vs RTX A6000, GeForce RTX 4090 48GB, and Radeon RX 7900 XTX: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | Radeon PRO W7900 |  RTX A6000 | GeForce RTX 4090 48GB | Radeon RX 7900 XTX |
|---|---|---|---|---|
| Memory | 48 GB | 48 GBSame | 48 GBSame | 24 GB50% less |
| Bandwidth | 864 GB/s | 768 GB/s11% lower | 1,008 GB/s17% higher | 960 GB/s11% higher |
| BF16 / FP16 compute | 123 TFLOPS | 155 TFLOPS26% higher | 330 TFLOPS169% higher | 123 TFLOPSSame |
| INT8 compute | 123 TOPS | 310 TOPS152% higher | 661 TOPS437% higher | 123 TOPSSame |
| Qwen3.8 27B speed (2-bit) | 60.3 t/s | 54.1 t/s10% slower | 69.3 t/s15% faster | 66.3 t/s10% faster |
| Largest dense model (4-bit) | 79B | 79BSame | 79BSame | 37B53% less |
| Power | 295 W | 300 WSame | 450 W53% higher | 355 W20% higher |
| GPU interconnect | — | NVLink 112.5 GB/s | Not supported | — |
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 |
| Released | 2023-04 | 2020-10 | 2022-10 | 2022-12 |
| 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.
Amazon.jp
Xianyu