NVIDIA
Operating mode
Use this to bias workload recommendations toward responsiveness, background autonomy, lighter serving, or multi-GPU scale-out.
Current mode
Balanced
Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.
The RTX 4090 is NVIDIA's flagship consumer GPU built on the Ada Lovelace architecture. With 24 GB of GDDR6X VRAM and 16,384 CUDA cores, it is among the most capable consumer cards for local AI inference. It can run 13B parameter models at full precision and 70B+ models with quantization, delivering class-leading decode speeds thanks to its massive tensor core count and 1 TB/s memory bandwidth.
Official product page ↗Beyond LLMs
What AI tasks this GPU can handle — from text generation to image and video creation.
| Capability | Status | Representative Model |
|---|---|---|
| LLM Chat (7B) | Runs natively | Llama 3.1 8B Q4 |
| LLM Coding (30B) | Runs natively | Qwen 3 30B Q4 |
| LLM Large (70B) | Won’t fit | Llama 3.1 70B Q4 |
| Image Gen (SDXL) | Runs natively | SDXL 1.0 FP16 |
| Image Gen (Flux) | Runs with offload | Flux.1 Dev FP16 |
| Image Gen (SD 3.5) | Runs natively | SD 3.5 Large FP16 |
| Video Short (25f) | Runs natively | LTX Video 2B |
| Video Long (100f) | Won't fit | Wan Video 14B |
Architecture
Ada Lovelace is NVIDIA's fourth-generation RTX architecture, manufactured on TSMC's custom 4N process. It introduces 4th-generation Tensor Cores with FP8 support, 3rd-generation ray tracing cores, and the Shader Execution Reordering (SER) engine for improved workload scheduling.
AI Relevance
FP8 Tensor Core operations provide a significant uplift for quantized LLM inference compared to Ampere's FP16-only Tensor Cores. DLSS 3 Frame Generation demonstrates the architecture's AI processing capabilities.
Ada Lovelace is NVIDIA's fourth-generation RTX architecture, manufactured on TSMC's custom 4N process. It introduces 3rd-generation ray tracing cores, 4th-generation Tensor Cores with FP8 support, and the Shader Execution Reordering (SER) engine for improved workload scheduling.
The RTX 4090 features the full AD102 GPU die with 128 Streaming Multiprocessors (SMs), each containing 128 CUDA cores for a total of 16,384. Its 512 Tensor Cores can perform FP8 matrix operations at up to 1,321 TOPS, making it exceptionally efficient for quantized LLM inference.
The memory subsystem uses a 384-bit bus connected to 24 GB of Micron GDDR6X running at 21 Gbps, delivering 1,008 GB/s of bandwidth. For AI inference, this bandwidth is the primary bottleneck — it directly determines how many tokens per second the GPU can generate during autoregressive decoding.
购买建议
本地 AI 的绝佳选择
能良好运行 50 个顶级模型中的 26 个 — 本地推理的全能之选。
24.0 GB
VRAM
$1,599
建议零售价
$67/GB
每 GB VRAM 成本
最适合此 GPU 的模型
What will limit you first
This setup is broadly balanced for this model.
Very little memory headroom
You can run the model, but there is not much room left for longer context, bigger batches, extra apps, or future model updates.
Best upgrade itinerary
Buy headroom, not only minimum fit
A slightly larger memory tier gives you safer context growth and makes the recommendation more future-proof.
Unlocks 1 additional models that do not fit on the current setup.
想要更多余量？ MacBook Pro M4 Max 36GB (36.0 GB unified memory) 是下一步升级选择。
Cost vs cloud API
Assumes 4 hours/day of active inference at 83 tok/s, RTX 4090 24GB amortized over 36 months, US residential electricity ($0.15/kWh), blended cloud pricing at $10 per 1M tokens (GPT-4o / Claude Sonnet tier).
36.0M
Tokens/month at this pace
$50.7
Monthly local cost
$360
Same tokens on cloud API
$1.41
Local $/1M tokens
Break-even: pays for itself in 4.5 months vs cloud API at this workload. Price reference: $1.6k MSRP.
Chat
SThis model is a direct match for chat. It belongs to a current frontier family for local AI. It fits natively with comfortable headroom. Known channels: huggingface, ollama, lm-studio.
Coding
SThis model is a direct match for coding. It belongs to a current frontier family for local AI. It should run, but memory headroom will be limited. Known channels: huggingface, ollama, lm-studio.
Agentic Coding
SThis model is still usable for agentic-coding, but it is not the most specialized pick. It belongs to a current frontier family for local AI. It should run, but memory headroom will be limited. Known channels: huggingface, lm-studio.
Reasoning
SThis model is a direct match for reasoning. It belongs to a current frontier family for local AI. It fits natively with comfortable headroom. Known channels: huggingface, ollama, lm-studio.
RAG
AThis model is a direct match for rag. It sits in the middle of the current model mix. It fits natively with comfortable headroom. Known channels: huggingface, ollama.
Qwen3-Coder 30B A3B InstructQwen3-VL 30B A3B Instruct
GPT-OSS 20BQwen 3 14B
Phi-4-reasoning-plus 14BQwen 3 30B A3BQwen 3.5 27BQwen 3.5 9B
Magistral Small 2507Devstral Small 2 24B InstructQwen 3.6 27BNemotron Cascade 2 30B A3BQwen 3 8BDevstral Small 1.1Ministral 3 14BNemotron 3 Nano 30B
Gemma 4 26B A4BQwen 3.5 4BNemotron Nano 8BQwen 3.5 35B A3BPhi-4 Mini Reasoning 4BQwen 3 32BQwen 3.6 35B A3B
Jina Embeddings v3
BGE M3
EXAONE 4.0 32BQwen 3.5 397B A17BDevstral 2 123B Instruct
Kimi K2.5Kimi K2.6
DeepSeek V4 ProQwen 3.5 122B A10BDeepSeek V4 FlashMistral Small 4 119B
Command A 111BQwen 2.5 VL 72BGPT-OSS 120BQwen3-Coder-Next
GLM-5.1Pixtral Large 124BGLM-5DeepSeek V3.2Qwen 3 235B A22BQwen3-Coder 480B A35B InstructGemma 4 31BLeanstral 119B A6BDeepSeek Coder V2 236BDeepSeek R1 671BDeepSeek V3.1 671B触手可及
高质量模型，只需稍多一点内存
Image & Video Generation
41 of 52 models can generate images or video on your RTX 4090 24GB
| Model | Max Resolution | Gen Time | Grade |
|---|---|---|---|
| SD TurboImage | 512×512 | 500ms | S |
| Stable Diffusion 1.5Image | 512×768 | ~1s | S |
| Realistic Vision v5.1Image | 512×768 | ~1s | S |
| DreamShaper 8Image | 512×768 | ~1s | S |
| LCM DreamShaper v7Image | 512×768 | 300ms | S |
| PixArt-SigmaImage | 1024×1024 | ~4s | S |
| FramePack I2VVideo | 256×256 | ~7.3s/frame | S |
| SDXL TurboImage | 512×512 | 500ms | S |
| SDXL LightningImage | 1024×1024 | ~1.5s | S |
| Stable Diffusion XL 1.0Image | 1024×1024 | ~4s | S |
| Playground v2.5Image | 1024×1024 | ~6s | S |
| RealVisXL v5.0Image | 1024×1024 | ~4.5s | S |
| DreamShaper XLImage | 1024×1024 | ~4.5s | S |
| Juggernaut XL v9Image | 1024×1024 | ~4.5s | S |
| Animagine XL 3.1Image | 1024×1024 | ~4.5s | S |
| Pony Diffusion V6 XLImage | 1024×1024 | ~4.5s | S |
| Animagine XL 4.0Image | 1024×1024 | ~4.5s | S |
| Illustrious XLImage | 1024×1024 | ~4.5s | S |
| Wan Video 2.1 1.3BVideo | 256×256 | ~2.9s/frame | S |
| Stable Diffusion 3.5 MediumImage | 1024×1024 | ~7s | S |
| Flux.2 Klein 4BImage | 1024×1024 | ~1.2s | S |
| LTX Video 2BVideo | 768×512 | ~3.5s/frame | S |
| KolorsImage | 1024×1024 | ~8s | S |
| Stable CascadeImage | 1024×1024 | ~10s | S |
| AuraFlow v0.3Image | 1536×1536 | ~18s | S |
| Stable Diffusion 3.5 LargeImage | 1024×1024 | ~22s | S |
| Stable Diffusion 3.5 Large TurboImage | 1024×1024 | ~4s | S |
| CogVideoX 2BVideo | 720×480 | ~3.5s/frame | A |
| HunyuanVideoVideo | 256×256 | ~7.3s/frame | A |
| ChromaImage | 256×256 | ~7.3s | A |
| Z-Image TurboImage | 1536×1536 | ~4.1s | B |
| Flux.1 DevImage | 256×256 | ~18s | B |
| Flux.1 SchnellImage | 256×256 | ~3.5s | B |
| LTX Video 13BVideo | 256×256 | ~7.3s/frame | B |
| Flux.1 Kontext DevImage | 256×256 | ~20s | B |
| AnimateDiff v1.5.3Video | 512×768 | ~1.8s/frame | B |
| Cosmos Diffusion 7BVideo | 256×256 | ~11.1s/frame | B |
| CogVideoX 5BVideo | 256×256 | ~10.5s/frame | B |
| Wan2.2 TI2V 5BVideo | 256×256 | ~10.5s/frame | B |
| Flux.2 Klein 9BImage | 256×256 | ~3.7s | D |
| Flux.1 Fill DevImage | 256×256 | ~17s | D |
| Mochi 1 PreviewVideo | 256×256 | ~6.6s/frame | F |
| HunyuanVideo 1.5Video | 256×256 | ~6.1s/frame | F |
| Helios 14BVideo | 256×256 | ~7.6s/frame | F |
| SkyReels V2 14BVideo | 256×256 | ~7.6s/frame | F |
| Wan Video 2.1 14BVideo | 256×256 | ~7.6s/frame | F |
| Wan Video 2.2 14BVideo | 256×256 | ~7.6s/frame | F |
| Qwen ImageImage | 256×256 | ~6.7s | F |
| Qwen Image EditImage | 256×256 | ~6.7s | F |
| Flux.2 DevImage | 256×256 | ~3m 9s | F |
| MAGI-1Video | 256×256 | ~9.4s/frame | F |
| HunyuanImage 3.0Image | 256×256 | ~11.9s | F |
Image models estimated at 1024×1024 (28 steps, FP16). Video models estimated at 768×512 (25 frames, 30 steps, FP16). Actual performance varies with runtime and system load.
Multi-GPU scaling
Scale out with multiple GPUs for larger models. PCIe interconnect with 30% scaling overhead.
| Config | Effective memory | Models that fit | Est. bandwidth |
|---|---|---|---|
| 1× RTX | 24 GB | 319/374 | 1,008 GB/s |
| 2× RTX | 48 GB | 338/374 | 1,411 GB/s |
Model counts use default quantization at coding workload settings. Multi-GPU scaling factor: 0.7× per additional GPU.
Upgrade paths
See what you unlock with more powerful hardware
升级选项
2× RTX 4090 24GBMulti-GPUUnlocks 19 additional models that do not fit on the current setup.
Scale-out only pays off if the host platform has enough PCIe lanes, slot spacing, power, and cooling.
The bigger the setup gets, the more the runtime matters. Multi-GPU and multi-user serving are where vLLM, SGLang, TGI, TensorRT-LLM, or tuned llama.cpp start to earn their complexity.
~$1,599 MSRP
Unlocks 1 additional models that do not fit on the current setup.
~$2,499 MSRP
RTX 5000 Ada 32GBNVIDIA 升级Unlocks 6 additional models that do not fit on the current setup.
~$4,000 MSRP
Unlocks 17 additional models that do not fit on the current setup.
~$1,099 MSRP
Unlocks 45 additional models that do not fit on the current setup.
Lifts average decode speed across fitting models by about 97%.
~$8,000 MSRP
RTX 4090 24GB (24 GB VRAM) can run these top models: Qwen3-Coder 30B A3B Instruct (score: 97/100), Qwen3-VL 30B A3B Instruct (score: 96/100), GPT-OSS 20B (score: 95/100). See the full compatibility list above.
RTX 4090 24GB has 24 GB of VRAM available for AI model inference. This determines which models and quantization levels you can run locally.
Yes, RTX 4090 24GB is excellent for running LLMs locally with top compatibility scores above 80/100.
For coding on RTX 4090 24GB, we recommend Devstral Small 2 24B Instruct. It achieves 40.0 tokens per second with 40K context window. This model is a direct match for coding. It belongs to a current frontier family for local AI. It should run, but memory headroom will be limited. Known channels: huggingface, ollama, lm-studio.
There are 5 upgrade path(s) from RTX 4090 24GB: RTX 4090 24GB, MacBook Pro M4 Max 36GB. Upgrading would unlock larger models and faster inference speeds.
Yes, RTX 4090 24GB with 24 GB of usable memory can run Flux.1 Dev at FP16 natively. Flux is a 12B parameter diffusion transformer that produces high-quality images. You can also run the Schnell variant for faster generation.
RTX 4090 24GB (24 GB VRAM) can handle various AI generation tasks beyond LLMs. For image generation, SDXL and Stable Diffusion 3.5 run well. Flux.1 Dev also runs natively for state-of-the-art image quality. For video, LTX Video 2.3 can generate short clips. Check the AI Capability Matrix above for detailed compatibility.
RTX 4090 24GB is excellent for AI image generation. With 24 GB of usable memory, it runs all major diffusion models including Flux.1, SDXL, and Stable Diffusion 3.5 at full precision. You can generate high-resolution images quickly and even handle video generation models.
Yes, RTX 4090 24GB with 24 GB of usable memory can run Qwen 3.5 27B at Q4_K_M (~16.5 GB) with ~7 GB headroom for context and runtime. Quality at Q4 is very close to full precision for most tasks. Run it with: ollama run qwen3.5:27b
With 24 GB on RTX 4090 24GB, Q4_K_M is the sweet spot for 27B-35B models, Q6_K for 14B models, and Q8_0 for 8B-9B models. The general rule: use the highest quantization that fits with at least 2-3 GB headroom for KV cache and runtime.
RTX 4090 24GB already has strong memory bandwidth, so the next limit is often memory capacity and context headroom rather than raw decode speed. For local LLMs, fit first and bandwidth second is the right mental model.
RTX 4090 24GB supports up to 2× GPU scaling via PCIe. With 2× GPUs, you get 48 GB effective memory with a 0.7× scaling factor per GPU. This enables running models like Qwen 3.5 397B A17B and Devstral 2 123B Instruct that don't fit on a single card.
RTX 4090 24GB uses PCIe for multi-GPU communication, which has approximately 30% scaling overhead. For best multi-GPU performance, consider NVLink-equipped variants.
Usually yes. If you want to run 2-4× RTX 4090 24GB for local AI, the bottleneck often becomes the platform, not the card. Workstation and server boards give you more CPU PCIe lanes, better x16 slot wiring, more spacing between cards, stronger power delivery, and usually more RAM capacity. Consumer x8/x8 layouts can work, but they are a common weak point in multi-GPU builds.
Compare with similar
Related guides
