GeForce RTX 5090 D Price & Used Prices: Which Local LLMs Can It Run?
Price alert
We'll email you when the eBay (US) price reaches your target.
The RTX 5090 D is the China-market 5090: same 32GB and bandwidth, capped AI compute, still about 74 tokens/s on Qwen3.8 27B at 4-bit.
eBay
GeForce RTX 5090 D 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 GeForce RTX 5090 D specifications
- Memory32 GB · GDDR7
- Bandwidth1,792 GB/s
- FP16 / BF16296.9 TFLOPS
- FP8 / INT8593.8 / 593.8 TOPS
- Bus width
- 512-bit
- Architecture
- Blackwell (GB202-250)
- Cores
- 21,760
- Tensor cores
- 680
- Ecosystem
- CUDA
- Interface
- PCIe 5.0 x16
- NVLink
- Not supported
- Power
- 575 W
- Power connector
- 16-pin 12V-2x6
- Type
- Consumer
- Launch date
- Jan 30, 2025
- MSRP
- $2,299
What LLMs can a single GeForce RTX 5090 D run?
One 5090 D fits every model up to 35B at 4-bit; at 8-bit it fits 27B-class at about 44 tokens/s, which the 24GB 5090 D V2 can't.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 E4B8B | 3.8 GB | Runs | 236.9 | 128K |
 Qwen3.5 9B9.7B | 4.1 GB | Runs | 223.2 | 256K |
| Gemma 4 12B12B | 4.7 GB | Runs | 200.4 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 167 | 256K |
| Qwen3.6 27B27.8B | 11.8 GB | Runs | 98.2 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 114.6 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 174.7 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 95 | 246K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 181.6 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 5 GB | Runs | 196.3 | 128K |
| Qwen3.5 9B9.7B | 5.7 GB | Runs | 177.1 | 256K |
| Gemma 4 12B12B | 7.1 GB | Runs | 148.5 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 149.5 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 72.3 | 245K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 73.5 | 250K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 161.5 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 65.5 | 168K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 163.9 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 8.2 GB | Runs | 134.7 | 128K |
| Qwen3.5 9B9.7B | 9.5 GB | Runs | 118.9 | 256K |
| Gemma 4 12B12B | 12.7 GB | Runs | 92.5 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 128.4 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 44.6 | 69K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 44 | 63K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 139.4 | 26K |
| Gemma 4 31B30.7B | 32.6 GB | Too big | — | — |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Needs RAM offload | 36.5 | 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
- A 5090 D is the 32GB consumer card you can get
It fits exactly the same models as the 5090
- Mainly single-user chat and coding
Generation speed depends on bandwidth, which is the same as the 5090's
- Running 27B-class models at 8-bit
It has 8GB more than the V2, and of the two versions only this one can hold them
Not a good fit
- Compute-heavy work
For image and video generation and long-prompt processing, it is rated at only 70% of the 5090
- Planning multi-GPU setups or power tuning
Around launch, forum users who analyzed the driver claimed it restricts multi-GPU and AI workloads and that the power limit can't be changed; NVIDIA has not confirmed this
Before you buy
The driver restriction claims
Around the January 2025 launch, a user on the Chiphell forum analyzed the driver and claimed that once the 5090 D detects AI inference, mining or multi-GPU operation, it locks power within 3 seconds, and the power limit can't be adjusted. NVIDIA has not commented officially, and our speed estimates don't account for these restrictions. If you plan to run vLLM across multiple cards, check with people already doing it first.
Compute is limited, not memory
Core count, VRAM and bandwidth are all the same as the 5090's; only the officially rated AI compute is lower, and NVIDIA hasn't explained exactly how it is limited. Text generation depends mainly on bandwidth, so the impact there is small. Processing long prompts and serving several requests at once will be somewhat slower, but we have no reliable test data on how much.
Power and connector
The card draws 575W and uses the same 12V-2x6 connector as the 5090, with the same risk of melting: use the power supply maker's native cable and seat it fully. For why, see the RTX 5090 page.
Compared with similar GPUs
GeForce RTX 5090 D vs GeForce RTX 5090, GeForce RTX 5090 D V2, and GeForce RTX 4090D 48GB: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | GeForce RTX 5090 D | GeForce RTX 5090 | GeForce RTX 5090 D V2 | GeForce RTX 4090D 48GB |
|---|---|---|---|---|
| Memory | 32 GB | 32 GBSame | 24 GB25% less | 48 GB50% more |
| Bandwidth | 1,792 GB/s | 1,792 GB/sSame | 1,344 GB/s25% lower | 1,008 GB/s44% lower |
| BF16 / FP16 compute | 297 TFLOPS | 419 TFLOPS41% higher | 297 TFLOPSSame | 294 TFLOPSSame |
| INT8 compute | 594 TOPS | 838 TOPS41% higher | 594 TOPSSame | 589 TOPSSame |
| Qwen3.8 27B speed (2-bit) | 114.6 t/s | 114.6 t/sSame | 89.5 t/s22% slower | 69.3 t/s40% slower |
| Largest dense model (4-bit) | 51B | 51BSame | 37B27% less | 79B55% more |
| Power | 575 W | 575 WSame | 575 WSame | 425 W26% lower |
| Interface | PCIe 5.0 x16 | PCIe 5.0 x16 | PCIe 5.0 x16 | PCIe 4.0 x16 |
| Released | 2025-01 | 2025-01 | 2025-08 | 2023-12 |
| 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