← All GPUs
NVIDIA · Consumer · Blackwell (GB202)
NVIDIA GeForce RTX 5090
575 W on one 12V-2x6 (600 W) cable; NVIDIA recommends a 1000 W PSU. FE is 2-slot.
VRAM
32 GB
GDDR7
Memory bandwidth
1,792 GB/s
FP16 compute
209.5 TFLOPS
dense tensor
Price
$1,999
launch price; no current price
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 1 and 2 cards. 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 (32 GB) | 2 cards (64 GB) |
|---|---|---|---|
| Qwen3.5 4B4B · Q4_K_M | 4.3 GB | fits | fits |
| Llama 3.1 8B Instruct8B · Q4_K_M | 7.1 GB | 200 tok/smedian of 3 runs | fits |
| Qwen3 8B8.2B · Q4_K_M | 7.3 GB | 200 tok/smedian of 1 run | fits |
| Gemma 4 E4B IT8B · Q4_K_M | 6.5 GB | fits | fits |
| Granite 4.2 8B8.8B · Q4_K_M | 7.9 GB | fits | fits |
| Qwen3.5 9B9B · Q4_K_M | 7.0 GB | fits | fits |
| Gemma 4 12B IT11.95B · Q4_K_M | 8.2 GB | fits | fits |
| Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M | 10.4 GB | fits | fits |
| Phi-414.7B · Q4_K_M | 11.3 GB | fits | fits |
| gpt-oss-20b21B, 3.6B active · MXFP4 (native) | 12.5 GB | 290 tok/smedian of 4 runs | fits |
| Devstral Small 2 24B (2512)24B · Q4_K_M | 16.1 GB | fits | fits |
| Qwen3.6 27B27B · Q4_K_M | 17.7 GB | fits | fits |
| Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M | 17.4 GB | 122 tok/smedian of 2 runs | fits |
| Qwen3.8 27B27B · Q4_K_M | 18.2 GB | fits | fits |
| Granite 4.2 30B29.3B · Q4_K_M | 20.3 GB | fits | fits |
| Gemma 4 31B IT30.7B · Q4_K_M | 18.9 GB | 61 tok/smedian of 1 run | fits |
| Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M | 19.5 GB | fits | fits |
| Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M | 21.9 GB | fits | fits |
| Qwen3.6 35B-A3B35B, 3B active · Q4_K_M | 22.4 GB | 117 tok/smedian of 1 run | fits |
| Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M | 25.3 GB | fits | fits |
| Llama 3.3 70B Instruct70B · Q4_K_M | 43.6 GB | too large | fits |
| Qwen3 Coder Next80B, 3B active · Q4_K_M | 46.9 GB | too large | fits |
| gpt-oss-120b117B, 5.1B active · MXFP4 (native) | 60.2 GB | too large | fits |
| Llama 4 Scout 17B-16E Instruct109B, 17B active · Q4_K_M | 62.4 GB | too large | fits |
Machines built on it
- 1x RTX 5090~$3,234 + unpriced parts
- 2x RTX 5090~$4,109 + unpriced parts
- RTX 5090 + EPYC 9355P with 768 GB DDR5 (MoE offload)~$3,763 + unpriced parts
Measured speeds
One card
| Model | Quant | Backend | Generation | Prompt | Context | Source |
|---|---|---|---|---|---|---|
| Gemma 4 26B-A4B IT | Q4_K | llama.cpp (build not stated) | 180 tok/s | 8,799 tok/s4,096-token prompt | 4,096 | review2026-03 |
| Gemma 4 26B-A4B IT | INT4 experts / HQQ4 attention / k4v4 KV | Krasis (MoE runtime; version not stated in table) | 64 tok/s | 5,599 tok/s | 0 | independent benchmark2026-10 |
| Gemma 4 31B IT | Q4_K | llama.cpp (build not stated) | 61 tok/s | 3,395 tok/s4,096-token prompt | 4,096 | review2026-03 |
| gpt-oss-20b | MXFP4 MoE | llama.cpp b6782 CUDA,BLAS | 299 tok/s | 13,002 tok/s2,048-token prompt | 0 | llama.cpp benchmark discussion2025-10 |
| gpt-oss-20b | Q4_K | llama.cpp (build not stated) | 298 tok/s | 9,444 tok/s4,096-token prompt | 4,096 | review2026-03 |
| gpt-oss-20b | MXFP4 MoE | llama.cpp b6210 CUDA | 283 tok/s | 9,848 tok/s2,048-token prompt | 0 | llama.cpp benchmark discussion2025-08 |
| gpt-oss-20b | MXFP4 | Ollama (version not stated) | 205 tok/s | 8,519 tok/s | 0 | review2025-10 |
| Llama 3.1 8B Instruct | Q4_K_M | llama.cpp b11741 | 278 tok/s | – | 0 | review2026-10 |
| Llama 3.1 8B Instruct | Q4_K_M | Ollama (version not stated) | 200 tok/s | 30,982 tok/s | 0 | review2025-10 |
| Llama 3.1 8B Instruct | Q4_K_M | LocalScore (llamafile; version not stated on leaderboard) | 66 tok/s | 6,297 tok/s | 0 | LocalScore2026-10 |
| Llama 3.1 8B Instruct | Q8_0 | Ollama (version not stated) | 145 tok/s | 31,442 tok/s | 0 | review2025-10 |
| Llama 3.1 8B Instruct | FP8 | SGLang (lmsysorg/sglang:spark image) | 138 tok/s | 21,956 tok/s2,048-token prompt | 0 | review2025-10 |
| Llama 3.3 70B Instruct | IQ3_XXS | llama.cpp b11741 | 53 tok/s | – | 0 | review2026-10 |
| Qwen3 8B | Q4_K | llama.cpp b8189 | 200 tok/s | 11,933 tok/s4,096-token prompt | 4,096 | review2026-03 |
| Qwen3.6 35B-A3B | INT4 experts / HQQ4 attention / k4v4 KV | Krasis (MoE runtime; version not stated in table) | 117 tok/s | 10,670 tok/s | 0 | independent benchmark2026-10 |
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 |
|---|---|---|---|---|---|---|---|
| Gemma 4 26B-A4B IT | Q4_K | llama.cpp (build not stated) | 167 tok/s | 7,733 tok/s16,384-token prompt | 16,384 | RTX 5090 | review2026-03 |
| Gemma 4 26B-A4B IT | Q4_K | llama.cpp (build not stated) | 159 tok/s | 6,293 tok/s32,768-token prompt | 32,768 | RTX 5090 | review2026-03 |
| Gemma 4 31B IT | Q4_K | llama.cpp (build not stated) | 59 tok/s | 2,794 tok/s16,384-token prompt | 16,384 | RTX 5090 | review2026-03 |
| Gemma 4 31B IT | Q4_K | llama.cpp (build not stated) | 55 tok/s | 2,229 tok/s32,768-token prompt | 32,768 | RTX 5090 | review2026-03 |
| gpt-oss-20b | Q4_K | llama.cpp (build not stated) | 249 tok/s | 7,168 tok/s16,384-token prompt | 16,384 | RTX 5090 | review2026-03 |
| gpt-oss-20b | Q4_K | llama.cpp (build not stated) | 215 tok/s | 5,183 tok/s32,768-token prompt | 32,768 | RTX 5090 | review2026-03 |
| Qwen3 8B | Q4_K | llama.cpp b8189 | 162 tok/s | 8,538 tok/s16,384-token prompt | 16,384 | RTX 5090 | review2026-03 |
| Qwen3 8B | Q4_K | llama.cpp b8189 | 130 tok/s | 6,034 tok/s32,768-token prompt | 32,768 | RTX 5090 | review2026-03 |
| Qwen3.8 27B | NVFP4 (native, NVFP4 KV) | vLLM 0.27.1 + heavy local patches | 101 tok/s | – | 0 | RTX 5090 | independent benchmark2026-08 |
| Qwen3.8 27B | AutoRound INT4 (fp8 KV) | vLLM v0.27.1 (club-3090 dual-superfast compose) | 192 tok/s | 5,650 tok/s10,000-token prompt | 0 | 2× RTX 5090 | independent benchmark2026-08 |
| Qwen3.8 27B | AutoRound INT4 (fp8 KV) | vLLM v0.27.1 (club-3090 dual-superfast compose) | 94 tok/s | 7,400 tok/s10,000-token prompt | 0 | 2× RTX 5090 | independent benchmark2026-08 |
Specifications
- Memory
- 32 GB GDDR7
- FP16 basis
- Dense FP16/BF16 tensor with FP32 accumulate; FP16 accumulate 419 dense (838 sparse). NVIDIA markets 3,352 AI TOPS (FP4 sparse).
- Board power
- 575 W
- Slots
- 2
- Length
- 304 mm
- PCIe
- Gen 5 x16
- NVLink
- No
- Cooling
- Open-air
- Software
- CUDA, Vulkan
- Released
- 2025
- Availability
- sold new
- Price basis
- Founders Edition removed from NVIDIA's store in Sept 2026 (NVIDIA says it is out of stock, not discontinued). The last NVIDIA store listing was reported at $4,929.99 in Aug 2026, against the $1,999 launch MSRP. Partner cards are still sold.
- Maker page
- nvidia.com
Sources
- nvidia.com: 32 GB GDDR7 512-bit, 575 W TGP, 1000 W system PSU, FE 304 mm 2-slot, PCIe Gen 5, NVLink No
- images.nvidia.com: NVIDIA RTX Blackwell whitepaper appendix tables: memory, bandwidth, TGP, PCIe gen, dense/sparse FP16 tensor TFLOPS
- nvidia.com: NVIDIA GeForce compare page: RTX 50 series memory bandwidth row; no RTX 50 SUPER listed
- nvidianews.nvidia.com: NVIDIA launch pricing: RTX 5090 $1,999, RTX 5080 $999, RTX 5070 Ti $749
- tomshardware.com: RTX 5090/5080 Founders Edition removed from NVIDIA's store in Sept 2026; NVIDIA says still in production
- kami-labs.fr: Aug 15 2026 report: NVIDIA's own store listed the RTX 5090 at $4,929.99