GeForce RTX 4090 48GB Price & Used Prices: Which Local LLMs Can It Run?
The RTX 4090 48GB is a modded card, the 4090 chip on a new board with double memory, running up to 35B at 8-bit. NVIDIA makes no such model.
eBay
GeForce RTX 4090 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.
Full GeForce RTX 4090 48GB specifications
- Memory48 GB · GDDR6X
- Bandwidth1,008 GB/s
- FP16 / BF16330.3 TFLOPS
- FP8 / INT8660.6 / 660.6 TOPS
- Bus width
- 384-bit
- Architecture
- Ada Lovelace (AD102)
- Cores
- 16,384
- Tensor cores
- 512
- Ecosystem
- CUDA
- Interface
- PCIe 4.0 x16
- NVLink
- Not supported
- Power
- 450 W
- Power connector
- 16-pin 12VHPWR
- Slots
- 2
- Type
- Consumer
- Launch date
- Oct 12, 2022
What LLMs can a single GeForce RTX 4090 48GB run?
A single 4090 48GB is made for 8-bit: Qwen3.8 27B at about 26 tokens/s, 30B-class MoE about 110. At 4-bit it squeezes in 70B-class with short context.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 4.7 GB | Runs | 128.5 | 256K | |
| 10.5 GB | Runs | 143.1 | 256K | |
| 11.8 GB | Runs | 58.8 | 256K | |
| 9.8 GB | Runs | 69.3 | 256K | |
| 11.9 GB | Runs | 153.4 | 198K | |
| 11.8 GB | Runs | 56.7 | 256K | |
| 12.3 GB | Runs | 163 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 7.1 GB | Runs | 91.9 | 256K | |
| 16.9 GB | Runs | 121.4 | 256K | |
| 16.8 GB | Runs | 42.6 | 256K | |
| 16.5 GB | Runs | 43.3 | 256K | |
| 18.3 GB | Runs | 136.1 | 198K | |
| 18.3 GB | Runs | 38.4 | 256K | |
| 22.1 GB | Runs | 139 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 12.7 GB | Runs | 55.1 | 256K | |
| 26.9 GB | Runs | 98.1 | 256K | |
| 28.6 GB | Runs | 25.8 | 256K | |
| 29 GB | Runs | 25.5 | 256K | |
| 31.8 GB | Runs | 109.9 | 198K | |
| 32.6 GB | Runs | 22.4 | 202K | |
| 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 4090 48GB cards run?
Two 4090 48GB cards (96GB) run Qwen3.8 Flash Next at 2-bit at about 115 tokens/s; four run DeepSeek V4 Flash at 4-bit. No NVLink.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| 78.9 GB | 2 cards | No price | 96 GB | 115.1 | 900 W | |
| 96.8 GB | 4 cards | No price | 192 GB | 94.8 | 1,800 W | |
| 107.2 GB | 4 cards | No price | 192 GB | 46.8 | 1,800 W | |
| 108.7 GB | 4 cards | No price | 192 GB | 77.8 | 1,800 W | |
| 253.9 GB | 8 cards | No price | 384 GB | 41.3 | 3,600 W | |
| 307.2 GBest. | 8 cards | No price | 384 GB | 31.4 | 3,600 W | |
| 311.2 GB | 8 cards | No price | 384 GB | 37 | 3,600 W | |
| 339.5 GB | 8 cards | No price | 384 GB | 49.5 | 3,600 W | |
| 340.5 GB | 8 cards | No price | 384 GB | 49.4 | 3,600 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| 111.3 GB | 4 cards | No price | 192 GB | 98.8 | 1,800 W | |
| 155.1 GB | 4 cards | No price | 192 GB | 71.4 | 1,800 W | |
| 182.2 GB | 8 cards | No price | 384 GB | 35.4 | 3,600 W | |
| 199.7 GB | 8 cards | No price | 384 GB | 51.4 | 3,600 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| 161.9 GB | 4 cards | No price | 192 GB | 69.4 | 1,800 W | |
| 188.2 GB | 8 cards | No price | 384 GB | 74 | 3,600 W | |
| 317.7 GB | 8 cards | No price | 384 GB | 24.5 | 3,600 W | |
| 341 GB | 8 cards | No price | 384 GB | 33.6 | 3,600 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Running a 70B-class dense model on one card
Llama 3.3 70B at 4-bit needs 46.3 GB at 8K of context, and it fits
- You want 8-bit precision for 27B to 35B models
Qwen3.8 27B at 8-bit runs at about 26 tokens/s, and Qwen3.6 35B A3B at 8-bit at about 114
- You want 48GB with FP8 and the full CUDA ecosystem in a dual-slot card that's easy to fit into a regular case or a multi-GPU server
Not a good fit
- You want quiet
In a review, the single-fan blower peaked at 65 dB during benchmarks
- You need a warranty or plan to run it 24/7 long-term
The chip may come from a pulled card, and the BIOS was reportedly modified with leaked internal NVIDIA tools
- You just want the same model to run faster
Bandwidth is the same as the original 24GB card's, so speed is the same too
Before you buy
How it's made
On the original 4090, memory sits only on the front of the board. Modders move the GPU onto a new circuit board that can carry memory chips on both sides (with a layout close to the RTX 3090's), fit twelve 2GB chips on each side, then flash a BIOS modified for the new capacity. hardware-corner believes the modders used leaked internal NVIDIA tools that leave the BIOS signature intact, which is why the official driver recognizes 48GB.
Rear memory temperature
Half the memory is on the back of the board, cooled only by the backplate. In a review by the Russian channel МК, the GPU ran at about 70°C while the memory hit 86°C, the hottest part of the card. When it arrives, run it at full load for ten minutes and check memory temperature in HWiNFO; once it's in the case, ideally have some airflow over the backplate side.
Noise, power and slots
The reviewed card is a dual-slot, single-fan blower, which suits multi-GPU setups with cards side by side. The trade-off is noise, peaking at about 65 dB during benchmarks, which is very noticeable under a desk. Power comes through a 16-pin 12VHPWR connector; plan your power supply for 450W.
Sourcing and warranty
Sourcing varies by seller: the reviewed card's GPU showed signs of prior use while its memory chips were new and from a single batch, whereas a personal mod reported by PC Gamer used old memory on hand. There's no NVIDIA or board-partner warranty, so you rely on the seller; ask about the warranty period and whether you can return the card if it fails your tests. All 24 memory chips have been resoldered, so a memtest_vulkan run must cover the full 48GB; testing only part of it won't catch bad chips.
Compared with similar GPUs
GeForce RTX 4090 48GB vs GeForce RTX 4090D 48GB, RTX 6000 Ada, and RTX PRO 5000 48GB: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | ||||
|---|---|---|---|---|
| Memory | 48 GB | 48 GBSame | 48 GBSame | 48 GBSame |
| Bandwidth | 1,008 GB/s | 1,008 GB/sSame | 960 GB/s5% lower | 1,344 GB/s33% higher |
| BF16 / FP16 compute | 330 TFLOPS | 294 TFLOPS11% lower | 364 TFLOPS10% higher | 258 TFLOPS22% lower |
| INT8 compute | 661 TOPS | 589 TOPS11% lower | 729 TOPS10% higher | 516 TOPS22% lower |
| Qwen3.8 27B speed (2-bit) | 69.3 t/s | 69.3 t/sSame | 66.3 t/s4% slower | 89.5 t/s29% faster |
| Largest dense model (4-bit) | 79B | 79BSame | 79BSame | 79BSame |
| Power | 450 W | 425 W6% lower | 300 W33% lower | 300 W33% lower |
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x16 |
| Released | 2022-10 | 2023-12 | 2022-09 | 2025-03 |
| 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 4090 48GB rental price per hour
On Vast.ai, the GeForce RTX 4090 48GB rents for $0.8 per hour on demand, about $196/month at 8 hours a day.
| Option | Price | 4 h/day | 8 h/day | 24 h/day | View |
|---|---|---|---|---|---|
| $0.8/hr | $98/mo | $196/mo | $587/mo | Rent |
Rental links are referral links: we may receive a commission or credit if you sign up through them.