GeForce RTX 4090D 48GB Price, VRAM & Local LLMs
The RTX 4090D 48GB is a dual-slot blower card with 48GB of VRAM, built in China from the cut-down 4090D. It runs 30B-class MoE models at 8-bit at about 114 tokens/s. To confirm what you bought, check that the driver reports it as an RTX 4090 D.
eBay
GeForce RTX 4090D 48GB 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.
GeForce RTX 4090D 48GB VRAM & full specs
- Memory48 GB · GDDR6X
- Bandwidth1,008 GB/s
- FP16 / BF16294.3 TFLOPS
- FP8 / INT8588.5 / 588.5 TOPS
- Bus width
- 384-bit
- Architecture
- Ada Lovelace (AD102)
- Cores
- 14,592
- Tensor cores
- 456
- Ecosystem
- CUDA
- Interface
- PCIe 4.0 x16
- NVLink
- Not supported
- Power
- 425 W
- Power connector
- 16-pin 12VHPWR
- Slots
- 2
- Type
- Consumer
- Launch date
- Dec 28, 2023
What LLMs can a single GeForce RTX 4090D 48GB run?
A single 4090D 48GB runs every model up to 35B at 8-bit, Qwen3.8 27B at about 25 tokens/s. 70B-class at 4-bit fits only with very short context.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 12B12B | 4.7 GB | Runs | 128.5 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 143.1 | 256K |
 Qwen3.6 27B27.8B | 11.8 GB | Runs | 58.8 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 69.3 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 153.4 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 56.7 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 163 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 12B12B | 7.1 GB | Runs | 91.9 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 121.4 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 42.6 | 256K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 43.3 | 256K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 136.1 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 38.4 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 139 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 12B12B | 12.7 GB | Runs | 55.1 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 98.1 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 25.8 | 256K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 25.5 | 256K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 109.9 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Runs | 22.4 | 202K |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Runs | 113.7 | 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 GeForce RTX 4090D 48GB cards run?
Four 4090D 48GB cards (192GB) run DeepSeek V4 Flash at 8-bit at about 69 tokens/s; eight run the trillion-parameter Kimi K2.6 at 2-bit. No NVLink.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 2 cards | No price | 96 GB | 115.1 | 850 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 4 cards | No price | 192 GB | 94.8 | 1,700 W |
 Hy3295B (21B active) | 107.2 GB | 4 cards | No price | 192 GB | 46.8 | 1,700 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 4 cards | No price | 192 GB | 77.8 | 1,700 W |
| GLM 5.3744B (40B active) | 253.9 GB | 8 cards | No price | 384 GB | 41.3 | 3,400 W |
| Hy4 preview770B (49B active) | 307.2 GBest. | 8 cards | No price | 384 GB | 31.4 | 3,400 W |
 Muse Spark Open816B (42B active) | 311.2 GB | 8 cards | No price | 384 GB | 37 | 3,400 W |
 Kimi K2.7 Code1,000B (32B active) | 339.5 GB | 8 cards | No price | 384 GB | 49.5 | 3,400 W |
| Kimi K2.61,000B (32B active) | 340.5 GB | 8 cards | No price | 384 GB | 49.4 | 3,400 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 4 cards | No price | 192 GB | 98.8 | 1,700 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 4 cards | No price | 192 GB | 71.4 | 1,700 W |
| Hy3295B (21B active) | 182.2 GB | 8 cards | No price | 384 GB | 35.4 | 3,400 W |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | 8 cards | No price | 384 GB | 51.4 | 3,400 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 4 cards | No price | 192 GB | 69.4 | 1,700 W |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | 8 cards | No price | 384 GB | 74 | 3,400 W |
| Hy3295B (21B active) | 317.7 GB | 8 cards | No price | 384 GB | 24.5 | 3,400 W |
| GLM 5.3 Flash320B (18B active) | 341 GB | 8 cards | No price | 384 GB | 33.6 | 3,400 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Generation speed is all you care about
It runs the same model as fast as the modified 4090 48GB, about 139 tokens/s for Qwen3.6 35B A3B at 4-bit
- You want one card to run 70B-class models at 4-bit, or 30B-class models at 8-bit with long context
- Building a multi-GPU rack
It is a dual-slot blower card, and just two of them hold Qwen3.8 Flash Next at 2-bit
Not a good fit
- Regularly processing long documents or batched requests
It has about 10% less compute than a 4090
- You need a factory warranty
Once the card is converted to 48GB, the board partner's warranty on the original card is gone
Before you buy
What the 4090D is
The 4090D is a version NVIDIA released for the Chinese market in late 2023 to comply with US export rules: about 10% fewer cores, power cut from 450W to 425W, and the same memory specs as the 4090. The 48GB conversion works the same way as on the 4090: the GPU is moved to a new board with memory on both sides, then flashed with a modified BIOS.
Tell a 4090 chip from a 4090D chip
The two kinds of modified cards are hard to tell apart by looks, and listing titles often mix them up. The name the driver reports is the most reliable sign: a 4090D chip shows up as “RTX 4090 D”, with device ID 10DE:2685 in lspci, while a 4090 is 10DE:2684. Both generate at the same speed; the difference is in prompt processing and batch work. In the stock-card comparison reported by wccftech, the 4090D was about 6% slower on average in games and AI applications.
The driver recognizes it, but test the full VRAM
The “RTX 4090 D” is on the official NVIDIA driver's supported list, and reports say the official driver recognizes all 48GB. When the card arrives, run a job that nearly fills the 48GB for a while, such as Gemma 4 31B at 8-bit with 200K of context, and confirm there are no errors, no garbled output and no card dropping off the bus.
Cooling and power
It is also a dual-slot, single-fan blower card with memory on both sides of the board, so for noise and rear memory temperatures, see the reviews of the 4090 48GB. Rated power is 425W, 25W less than the 4090, and the connector is the same 16-pin 12VHPWR.
Compared with similar GPUs
GeForce RTX 4090D 48GB vs GeForce RTX 4090 48GB, Radeon PRO W7900, and L40S: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | GeForce RTX 4090D 48GB | GeForce RTX 4090 48GB | Radeon PRO W7900 | L40S |
|---|---|---|---|---|
| Memory | 48 GB | 48 GBSame | 48 GBSame | 48 GBSame |
| Bandwidth | 1,008 GB/s | 1,008 GB/sSame | 864 GB/s14% lower | 864 GB/s14% lower |
| BF16 / FP16 compute | 294 TFLOPS | 330 TFLOPS12% higher | 123 TFLOPS58% lower | 362 TFLOPS23% higher |
| INT8 compute | 589 TOPS | 661 TOPS12% higher | 123 TOPS79% lower | 733 TOPS25% higher |
| Qwen3.8 27B speed (2-bit) | 69.3 t/s | 69.3 t/sSame | 60.3 t/s13% slower | 60.3 t/s13% slower |
| Largest dense model (4-bit) | 79B | 79BSame | 79BSame | 79BSame |
| Power | 425 W | 450 W6% higher | 295 W31% lower | 350 W18% lower |
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 |
| Released | 2023-12 | 2022-10 | 2023-04 | 2023-08 |
| 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.
GeForce RTX 4090D 48GB rental price per hour
On Vast.ai, the GeForce RTX 4090D 48GB rents for $0.92 per hour on demand, about $225/month at 8 hours a day. Buying one (Xianyu median asking price $3,889) pays off in about 3.1 years, 1.6 years or 6 months at 4, 8 or 24 hours a day.
| Option | Price | 4 h/day | 8 h/day | 24 h/day | View |
|---|---|---|---|---|---|
 Vast.aiRentUpdated 11 min. ago | $0.92/hr | $112/mo | $225/mo | $674/mo | Rent |
 XianyuUpdated 24 min. ago | $3,889 | Payback ~3.1 yrsPower $9/mo | Payback ~1.6 yrsPower $19/mo | Payback ~6 moPower $56/mo | Buy |
Break-even uses the US residential average ($0.18/kWh) and 425 W rated power, excluding resale value and the rest of the build.
Rental links are referral links: we may receive a commission or credit if you sign up through them.
Amazon.jp