← All GPUs
NVIDIA · Consumer · Ampere (GA102)
NVIDIA GeForce RTX 3090 Ti
Same 24 GB as the 3090 with faster memory and 450 W; FE needs a 16-pin (12VHPWR) cable or 3x 8-pin adapter.
VRAM
24 GB
GDDR6X
Memory bandwidth
1,008 GB/s
FP16 compute
80 TFLOPS
dense tensor
Price
$1,999
launch price; sold used now, used price not tracked yet
What fits at 4-bit
The 4-bit file plus 8,192 tokens of context and 1.5 GB of overhead, against the VRAM of one card. Speeds are measured medians where someone has published a run at that card count; a blank speed means no run, not a slow one. Machine pages carry estimates for the gaps.
| Model | Needs | 1 card (24 GB) |
|---|---|---|
| Qwen3.5 4B4B · Q4_K_M | 4.3 GB | fits |
| Llama 3.1 8B Instruct8B · Q4_K_M | 7.1 GB | 110 tok/smedian of 1 run |
| Qwen3 8B8.2B · Q4_K_M | 7.3 GB | 124 tok/smedian of 1 run |
| Gemma 4 E4B IT8B · Q4_K_M | 6.5 GB | fits |
| Granite 4.2 8B8.8B · Q4_K_M | 7.9 GB | fits |
| Qwen3.5 9B9B · Q4_K_M | 7.0 GB | fits |
| Gemma 4 12B IT11.95B · Q4_K_M | 8.2 GB | fits |
| Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M | 10.4 GB | fits |
| Phi-414.7B · Q4_K_M | 11.3 GB | fits |
| gpt-oss-20b21B, 3.6B active · MXFP4 (native) | 12.5 GB | 160 tok/smedian of 1 run |
| Devstral Small 2 24B (2512)24B · Q4_K_M | 16.1 GB | fits |
| Qwen3.6 27B27B · Q4_K_M | 17.7 GB | fits |
| Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M | 17.4 GB | fits |
| Qwen3.8 27B27B · Q4_K_M | 18.2 GB | fits |
| Granite 4.2 30B29.3B · Q4_K_M | 20.3 GB | fits |
| Gemma 4 31B IT30.7B · Q4_K_M | 18.9 GB | fits |
| Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M | 19.5 GB | fits |
| Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M | 21.9 GB | fits |
| Qwen3.6 35B-A3B35B, 3B active · Q4_K_M | 22.4 GB | fits |
Measured speeds
One card
| Model | Quant | Backend | Generation | Prompt | Context | Source |
|---|---|---|---|---|---|---|
| gpt-oss-20b | MXFP4 | llama.cpp (build not stated) | 160 tok/s | 4,805 tok/s4,096-token prompt | 4,096 | review2026-03 |
| Llama 3.1 8B Instruct | Q4_K_M | LocalScore (llamafile; version not stated on leaderboard) | 110 tok/s | 4,024 tok/s | 0 | LocalScore2026-10 |
| Qwen3 8B | Q4_K | llama.cpp (build not stated) | 124 tok/s | 4,468 tok/s4,096-token prompt | 4,096 | review2026-03 |
Speculative decoding, several users, or long context
Kept apart because each measures a different workload from one person chatting near the start of a conversation.
| Model | Quant | Backend | Generation | Prompt | Context | Measured on | Source |
|---|---|---|---|---|---|---|---|
| gpt-oss-20b | MXFP4 | llama.cpp (build not stated) | 138 tok/s | 3,440 tok/s16,384-token prompt | 16,384 | RTX 3090 Ti | review2026-03 |
| gpt-oss-20b | MXFP4 | llama.cpp (build not stated) | 120 tok/s | 2,521 tok/s32,768-token prompt | 32,768 | RTX 3090 Ti | review2026-03 |
| Qwen3 8B | Q4_K | llama.cpp (build not stated) | 94 tok/s | 2,834 tok/s16,384-token prompt | 16,384 | RTX 3090 Ti | review2026-03 |
| Qwen3 8B | Q4_K | llama.cpp (build not stated) | 72 tok/s | 1,868 tok/s32,768-token prompt | 32,768 | RTX 3090 Ti | review2026-03 |
Specifications
- Memory
- 24 GB GDDR6X
- FP16 basis
- Dense FP16/BF16 tensor with FP32 accumulate (Ada whitepaper Table 2); FP16 accumulate 160 dense, 320 sparse.
- Board power
- 450 W
- Slots
- 3
- Length
- 313 mm
- PCIe
- Gen 4 x16
- NVLink
- Yes
- Cooling
- Open-air
- Software
- CUDA, Vulkan
- Released
- 2022
- Availability
- used only
- Price basis
- Not sold new by NVIDIA.
- Maker page
- nvidia.com
Sources
- nvidia.com: 24 GB GDDR6X, 450 W, FE 313 mm 3-slot, NVLink Yes, power connectors
- images.nvidia.com: NVIDIA Ada whitepaper Table 2 (RTX 4090 vs 3090 Ti): bandwidth, TGP, FP16 tensor TFLOPS
- nvidia.com: launch price from $1,999