GeForce RTX 4080 SUPER 32GB Price & Used Prices: Which Local LLMs Can It Run?
The RTX 4080 SUPER 32GB is a modded 16GB card that runs 27B-class at 8-bit, but bandwidth is unchanged: about 32 tokens/s at 4-bit. No warranty.
eBay
GeForce RTX 4080 SUPER 32GB 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 4080 SUPER 32GB specifications
- Memory32 GB · GDDR6X
- Bandwidth736 GB/s
- FP16 / BF16208.9 TFLOPS
- FP8 / INT8417.8 / 417.8 TOPS
- Bus width
- 256-bit
- Architecture
- Ada Lovelace (AD103)
- Cores
- 10,240
- Tensor cores
- 320
- Ecosystem
- CUDA
- Interface
- PCIe 4.0 x16
- NVLink
- Not supported
- Power
- 320 W
- Power connector
- 16-pin 12VHPWR
- Type
- Consumer
- Launch date
- Jan 31, 2024
What LLMs can a single GeForce RTX 4080 SUPER 32GB run?
A single 4080 SUPER 32GB lives at 4-bit: Qwen3.8 27B at about 32 tokens/s, 30B-class MoE above 100. 8-bit fits only 27B-class, in the teens.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 E4B8B | 3.8 GB | Runs | 121 | 128K |
 Qwen3.5 9B9.7B | 4.1 GB | Runs | 112.4 | 256K |
| Gemma 4 12B12B | 4.7 GB | Runs | 98.7 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 127.7 | 256K |
| Qwen3.6 27B27.8B | 11.8 GB | Runs | 43.9 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 52 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 139 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 42.3 | 246K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 150.1 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 5 GB | Runs | 96.2 | 128K |
| Qwen3.5 9B9.7B | 5.7 GB | Runs | 85.2 | 256K |
| Gemma 4 12B12B | 7.1 GB | Runs | 69.5 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 104.8 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 31.6 | 245K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 32.2 | 250K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 120.1 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 28.5 | 168K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 123.3 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 8.2 GB | Runs | 62.3 | 128K |
| Qwen3.5 9B9.7B | 9.5 GB | Runs | 54.2 | 256K |
| Gemma 4 12B12B | 12.7 GB | Runs | 41.1 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 81.9 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 19 | 69K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 18.8 | 63K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 93.3 | 26K |
| Gemma 4 31B30.7B | 32.6 GB | Too big | — | — |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Needs RAM offload | 33.8 | 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 4080 SUPER 32GB cards run?
Two 4080 SUPER 32GB cards fit gpt-oss 120B at 2-bit; four (128GB) run Qwen3.8 Flash Next at 4-bit at about 83 tokens/s. No NVLink.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 4 cards | No price | 128 GB | 98.4 | 1,280 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 4 cards | No price | 128 GB | 78.7 | 1,280 W |
 Hy3295B (21B active) | 107.2 GB | 4 cards | No price | 128 GB | 36.3 | 1,280 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 4 cards | No price | 128 GB | 63.1 | 1,280 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 | 128 GB | 82.5 | 1,280 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 8 cards | No price | 256 GB | 57.3 | 2,560 W |
| Hy3295B (21B active) | 182.2 GB | 8 cards | No price | 256 GB | 27 | 2,560 W |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | 8 cards | No price | 256 GB | 40.1 | 2,560 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 8 cards | No price | 256 GB | 55.6 | 2,560 W |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | 8 cards | No price | 256 GB | 59.6 | 2,560 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Long context for 27B to 31B dense models
At 4-bit, Qwen3.8 27B can open up to 250K, and Gemma 4 31B up to 168K
- You want 32GB plus FP8
Qwen3.8 27B at 8-bit (29 GB) fits too, at about 19 tokens/s
- Running MoE models like GLM 4.7 Flash or Qwen3.6 35B A3B
About 120 to 123 tokens/s at 4-bit
- You want a multi-GPU build without 575W cards
Each card draws 320W, or 1,280W for four
Not a good fit
- Generation speed matters to you
On the same dense model it's about 20% slower than even an RTX 3090, and less than half as fast as an RTX 5090
- You need a factory warranty and long-term stability
The warranty sticker is the modder's, and neither NVIDIA nor the board partners honor it
Before you buy
Check whether the base card is a 4080 SUPER or a 4080
In 2026, 32GB cards modded from the regular RTX 4080 have also appeared on used marketplaces. Generation speed differs by only 2% to 3%, but the listing should say which it is. After installing, check the model name in nvidia-smi or the device ID in lspci: the 4080 SUPER is 10DE:2702 and the 4080 is 10DE:2704.
BIOS and drivers
Doubling the memory means resoldering the chips, changing the memory configuration on the board, and flashing a modified BIOS so the GPU recognizes 32GB; a video cited by VideoCardz says this relies on internal files leaked from NVIDIA. Reports differ on whether the official driver works: one user runs llama.cpp fine on official driver 595.84, while VideoCardz says these cards may need a patched driver, and that the patch has to be reapplied every time you want to game. Before ordering, ask the seller which driver version they tested with and whether they tried both Windows and Linux.
Blower cooling and warranty
The batches VideoCardz reported on all had their coolers swapped for blowers and carry a warranty sticker on the back, but that warranty is the modder's. The card draws 320W through a 16-pin 12VHPWR connector; if you plan to use a single card near you, listen to the fan at full load first.
Software: FP8 yes, FP4 no
This generation supports FP8, so FP8 models get a real speedup in vLLM, but it can't run the FP4 (NVFP4) that arrived with the RTX 50 series. One user loading an NVFP4 model in SGLang got an immediate error; switching to AWQ worked, but at only about half the speed of llama.cpp (with MTP on), so they went back to llama.cpp.
Compared with similar GPUs
GeForce RTX 4080 SUPER 32GB vs GeForce RTX 5090, GeForce RTX 4090, and Radeon AI PRO R9700: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | GeForce RTX 4080 SUPER 32GB | GeForce RTX 5090 | GeForce RTX 4090 | Radeon AI PRO R9700 |
|---|---|---|---|---|
| Memory | 32 GB | 32 GBSame | 24 GB25% less | 32 GBSame |
| Bandwidth | 736 GB/s | 1,792 GB/s143% higher | 1,008 GB/s37% higher | 640 GB/s13% lower |
| BF16 / FP16 compute | 209 TFLOPS | 419 TFLOPS101% higher | 330 TFLOPS58% higher | 191 TFLOPS9% lower |
| INT8 compute | 418 TOPS | 838 TOPS101% higher | 661 TOPS58% higher | 383 TOPS8% lower |
| Qwen3.8 27B speed (2-bit) | 52 t/s | 114.6 t/s120% faster | 69.3 t/s33% faster | 45.6 t/s12% slower |
| Largest dense model (4-bit) | 51B | 51BSame | 37B27% less | 51BSame |
| Power | 320 W | 575 W80% higher | 450 W41% higher | 300 W6% lower |
| Interface | PCIe 4.0 x16 | PCIe 5.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x16 |
| Released | 2024-01 | 2025-01 | 2022-10 | 2025-07 |
| 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