Sold by
Percepta Solutions Ltd
Registered UK business seller on eBay: perceptasolutions
Graphics card / GeForce RTX 4090 48GB
An RTX 4090 (AD102) graphics card with 48GB of GDDR6X, twice the memory of a standard RTX 4090, and blower cooling that exhausts through the rear bracket. Used and load-tested. Free tracked UK delivery.
Condition
Used, load-tested
GPU
AD102-301, 16,384 CUDA cores
Delivery
Free, tracked, UK only
GDDR6X on a 384-bit bus
CUDA cores, AD102 GPU
Memory bandwidth
Board power, blower cooled
01/ 08
What it is for
The 48GB of GDDR6X holds a language model, its context and its working memory on one card. In our test, Qwen3.8 27B ran with its full 262,144-token context window using 38.3GB.
Twice the memory of a standard RTX 4090, in a dual-slot blower card.
The blower pushes hot air out through the rear bracket, so cards can sit side by side. We ran two of these cards together in a dual-processor Lenovo ThinkStation PX.
02/ 08
Please note
An independently produced 48GB board based on the NVIDIA AD102 GPU. It is not made by NVIDIA.
03/ 08
Supplier specification and our readings
03.1Graphics card
| GPU | NVIDIA AD102-301-A1 (GeForce RTX 4090), Ada Lovelace architecture, 16,384 CUDA cores |
|---|---|
| Memory | 48GB GDDR6X (Micron, 24 x 2GB chips), 384-bit interface, 21Gbps |
| Memory bandwidth | 1,008.4GB/s |
| Clocks | 2,235MHz base, 2,520MHz boost |
| Interface | PCIe 4.0 x16 |
| Display outputs | 3 x DisplayPort 1.4a, 1 x HDMI 2.1 |
| Cooling | Blower (radial) fan, exhausting through the vented rear bracket |
| Power | 450W board power limit, adjustable down to 150W; one 16-pin power connector |
| Size | 266.7 x 111 x 38.5mm; full-height, dual-slot |
| PCB | 14 layers, matt black |
| APIs | CUDA (compute capability 8.9), DirectX 12 Ultimate, OpenGL 4.6, Vulkan |
| VBIOS | 95.02.3C.C0.7B on both cards |
| Supplier specification | BP49K-EA46B |
Ratings are from the supplier specification. Bandwidth, memory, VBIOS and power limits are as read from our cards with GPU-Z and nvidia-smi; the 48GB shows as 49,152MB in GPU-Z and 49,140MiB in nvidia-smi.
04/ 08
Percepta tests, October 2026
04.1Qwen3.8 27B on one card
| Reply speed, one request | 39.4 tok/s; 57.1 tok/s with MTP speculative decoding on |
|---|---|
| Prompt reading | 2,164 tok/s at its peak (8,816-token prompt); 1,255 tok/s with a 236,558-token prompt |
| Time to first token | 0.15s after a short prompt; 3 min 08s after a 236,558-token prompt |
| After a long prompt | 35.5 tok/s after 37,567 tokens; 23.8 tok/s after 236,558 tokens |
| Many requests at once | 202 tok/s in total across 16 requests; 158 tok/s across 8 requests with MTP on |
| GPU memory in use | 38.3GB with the full 262,144-token window and a 236,558-token prompt, including about 2.1GB for the Windows desktop |
| Sustained load | 155 tok/s in total from 8 requests for 15 minutes with no errors; GPU peak 63°C, blower at 100% |
Measured on 6 October 2026: Qwen3.8 27B (Unsloth UD-Q5_K_XL GGUF) in LM Studio 0.4.25, NVIDIA driver 610.47, 450W limit, MTP off unless stated, on a Ryzen 7 7800X3D PC with 32GB DDR5 and Windows 11 that ran the card at PCIe 4.0 x8. Results depend on the host system, driver, software, model, quantisation and settings.
04.2Both cards under load
| Host | Lenovo ThinkStation PX, 2 x Xeon Silver 4510, Ubuntu 24.04, NVIDIA driver 595.71; both cards at PCIe 4.0 x16 |
|---|---|
| Test | Both cards together for 10 minutes at the 450W limit, with a compute and memory load |
| Result | 395W to 412W per card, GPU core up to 77°C, memory up to 82°C; no thermal throttling and no driver errors |
26 October 2026, with the fan curve set to a 70% minimum.
05/ 08
In the box
06/ 08
Used
Condition grade
eBay: Used
Used
Used by Percepta in October 2026, for performance testing and in our own GPU server.
Both cards ran together at the full 450W limit for 10 minutes on 26 October 2026, with no thermal throttling and no driver errors.
Both cards report 48GB and VBIOS 95.02.3C.C0.7B.
Our performance testing ran the Windows desktop through the display outputs. In the server the cards ran compute work without a display.
Photographs
The card photographs are the supplier's product photographs of this model. Each card supplied is the same model, used, as described above. The GPU-Z image is from our test card.
07/ 08
Free and tracked
Free delivery to UK addresses.
08/ 08
Before you buy
Is this an RTX 4090 D? No. It has the full RTX 4090 configuration: 16,384 CUDA cores and device ID 10DE 2684 in GPU-Z. The RTX 4090 D has 14,592 CUDA cores.
Will two cards give me 96GB for one model? Each card has its own 48GB. Software such as llama.cpp, LM Studio or vLLM can split a model across two cards over PCIe, but the memory is not pooled into one 96GB device.
How loud is it? Like other blower cards, it is clearly audible at full load: the blower reached 100% in our 15-minute test. Lowering the power limit reduces heat and noise.
Sold by Percepta Solutions Ltd
Registered UK business seller on eBay: perceptasolutions
Seller
Registered UK business seller
Company No.
SC895219
VAT No.
GB526453489
A VAT invoice will be supplied with VAT separately displayed.
No manufacturer warranty is included with this sale. This does not affect your statutory rights.
GeForce RTX 4090 48GB GDDR6X | AD102-301 | Blower | Modified 48GB board | Used | UK delivery