GeForce RTX 2080 Ti 22GB 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 2080 Ti 22GB is a 2018 11GB gaming card modded to double memory, running Qwen3.8 27B at 4-bit at about 27 tokens/s. No warranty; test all 22GB.
eBay
GeForce RTX 2080 Ti 22GB 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 2080 Ti 22GB specifications
- Memory22 GB · GDDR6
- Bandwidth616 GB/s
- FP16 / BF16107.6 TFLOPS
- INT8215.2 TOPS
- Bus width
- 352-bit
- Architecture
- Turing (TU102)
- Cores
- 4,352
- Tensor cores
- 544
- Ecosystem
- CUDA
- Interface
- PCIe 3.0 x16
- NVLink
- 100 GB/s
- Power
- 250 W
- Power connector
- 2× 8-pin
- Type
- Consumer
- Launch date
- Sep 20, 2018
What LLMs can a single GeForce RTX 2080 Ti 22GB run?
A single 2080 Ti 22GB is built for 27B to 31B dense at 4-bit, leaving about 90K of context. 30B-class MoE is better at 2-bit: full context, faster.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 E4B8B | 3.8 GB | Runs | 104.1 | 128K |
 Qwen3.5 9B9.7B | 4.1 GB | Runs | 96.5 | 256K |
| Gemma 4 12B12B | 4.7 GB | Runs | 84.5 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 118.4 | 256K |
| Qwen3.6 27B27.8B | 11.8 GB | Runs | 37.1 | 160K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 44 | 189K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 130.2 | 192K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 35.8 | 118K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 141.9 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 5 GB | Runs | 82.3 | 128K |
| Qwen3.5 9B9.7B | 5.7 GB | Runs | 72.7 | 256K |
| Gemma 4 12B12B | 7.1 GB | Runs | 59.1 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 95.3 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 26.6 | 85K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 27.1 | 90K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 110.7 | 76K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 24 | 40K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 113.9 | 21K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 8.2 GB | Runs | 52.8 | 128K |
| Qwen3.5 9B9.7B | 9.5 GB | Runs | 45.9 | 256K |
| Gemma 4 12B12B | 12.7 GB | Runs | 34.7 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Needs RAM offload | 25 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Too big | — | — |
| Qwen3.8 27B27.8B | 29 GB | Too big | — | — |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Needs RAM offload | 22.3 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Too big | — | — |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Needs RAM offload | 33 | 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 2080 Ti 22GB cards run?
Multiple 2080 Ti 22GB cards pay off from four (88GB), running gpt-oss 120B at 4-bit at about 85 tokens/s. NVLink works at 100 GB/s.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 4 cards | $2,204 | 88 GB | 89 | 1,000 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 8 cards | $4,408 | 176 GB | 70.1 | 2,000 W |
 Hy3295B (21B active) | 107.2 GB | 8 cards | $4,408 | 176 GB | 31.3 | 2,000 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 8 cards | $4,408 | 176 GB | 55.5 | 2,000 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 8 cards | $4,408 | 176 GB | 73.7 | 2,000 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 8 cards | $4,408 | 176 GB | 50.2 | 2,000 W |
At this quantization, even 8 × GeForce RTX 2080 Ti 22GB cannot hold any model over 50B.
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Running 27B to 31B dense models with llama.cpp or Ollama
About 24 to 27 tokens/s at 4-bit, and Qwen3.8 27B still has room for 90K of context
- Running MoE models like Gemma 4 26B A4B and GLM 4.7 Flash
About 95 to 111 tokens/s at 4-bit
- You want CUDA and a tier more VRAM than 16GB
This chip generation is still supported by CUDA 13, current drivers and vLLM
Not a good fit
- You want Qwen3.6 35B A3B with long context
At 4-bit only 21K of context is left, while 24GB cards can open 123K
- You need BF16 or official FlashAttention 2
This chip generation has neither
- You want a factory warranty
It's a modded card, so if it breaks, the seller is your only recourse
Before you buy
This isn't an official model
NVIDIA only ever made an 11GB 2080 Ti. The 22GB version is a third-party mod that replaces the eleven 1GB memory chips with 2GB chips, then changes resistors on the board and the BIOS so the GPU recognizes the new capacity. The base card varies too: overseas sellers state that Gigabyte, MSI, ASUS or Leadtek cards ship at random. Before buying, ask about the base card's brand, who did the mod, and the warranty period.
Official drivers work, but test all 22GB
According to VideoCardz, sellers say the standard NVIDIA driver recognizes the full 22GB without a modified driver, which makes it less hassle than the 4080 32GB modded card. But the memory has been resoldered, so when the card arrives, test all 22GB with a tool like memtest_vulkan. If you see errors, garbled output or the card dropping out, deal with it within the return window.
Blower cooling and power
Most of these 22GB cards use blower cooling, which suits several cards side by side in a case; for a single card next to your desk, you need to be OK with blower fan noise. The card draws 250W through two 8-pin connectors.
Software: no BF16, so use FP16
llama.cpp, Ollama and vLLM all work out of the box; what CUDA 13 dropped were older architectures. The thing to watch is that this generation has no BF16: most new models are released in BF16, and on this card vLLM automatically loads them in FP16 instead (the log tells you so). Official FlashAttention 2 doesn't support it either; there's only a separately maintained fork.
Compared with similar GPUs
GeForce RTX 2080 Ti 22GB vs GeForce RTX 3090, Tesla P40, and Tesla V100 32GB: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | GeForce RTX 2080 Ti 22GB | GeForce RTX 3090 | Tesla P40 | Tesla V100 32GB |
|---|---|---|---|---|
| Memory | 22 GB | 24 GB9% more | 24 GB9% more | 32 GB45% more |
| Bandwidth | 616 GB/s | 936 GB/s52% higher | 347 GB/s44% lower | 900 GB/s46% higher |
| BF16 / FP16 compute | 108 TFLOPS | 142 TFLOPS32% higher | — | 112 TFLOPS4% higher |
| INT8 compute | 215 TOPS | 285 TOPS32% higher | 47 TOPS78% lower | 56 TOPS74% lower |
| Qwen3.8 27B speed (2-bit) | 44 t/s | 64.8 t/s47% faster | 25.5 t/s42% slower | 62.6 t/s42% faster |
| Largest dense model (4-bit) | 34B | 37B9% more | 37B9% more | 51B50% more |
| Power | 250 W | 350 W40% higher | 250 WSame | 250 WSame |
| GPU interconnect | NVLink 100 GB/s | NVLink 112.5 GB/s | Not supported | Not supported |
| Interface | PCIe 3.0 x16 | PCIe 4.0 x16 | PCIe 3.0 x16 | PCIe 3.0 x16 |
| Released | 2018-09 | 2020-09 | 2016-09 | 2018-03 |
| 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 2080 Ti 22GB rental price per hour
On Vast.ai, the GeForce RTX 2080 Ti 22GB rents for $0.12 per hour on demand, about $29/month at 8 hours a day. Buying one (eBay median $551) pays off in about 5.1 years, 2.5 years or 10 months at 4, 8 or 24 hours a day.
| Option | Price | 4 h/day | 8 h/day | 24 h/day | View |
|---|---|---|---|---|---|
 Vast.aiRentUpdated 1 hr. ago | $0.12/hr | $15/mo | $29/mo | $87/mo | Rent |
 eBayUpdated 43 days ago | $551 | Payback ~5.1 yrsPower $5/mo | Payback ~2.5 yrsPower $11/mo | Payback ~10 moPower $33/mo | Buy |
Break-even uses the US residential average ($0.18/kWh) and 250 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. 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