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 NVIDIA A100 40GB is the PCIe variant of NVIDIA's landmark Ampere datacenter GPU, offering the same 312 TFLOPS FP16 compute as the SXM version but with 40 GB of HBM2e and lower bandwidth at 1,555 GB/s. The PCIe form factor makes it compatible with standard servers without SXM infrastructure, and it is available on cloud providers like Google Cloud (A2) and AWS. A single A100 40GB can run 30B models with Q4 quantization and smaller 13B models near FP16, making it a practical and widely accessible inference option. It lacks the 80 GB of the SXM flagship but is substantially more affordable and available.
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) | Needs offload | Llama 3.1 70B Q4 |
| Image Gen (SDXL) | Runs natively | SDXL 1.0 FP16 |
| Image Gen (Flux) | Runs natively | 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
Ampere is NVIDIA's second-generation RTX architecture, built on Samsung's 8nm process. It introduced 3rd-generation Tensor Cores with support for sparsity-accelerated INT8 operations and improved FP16 throughput over Turing.
AI Relevance
Sparsity-aware Tensor Cores can effectively double throughput for structured sparse workloads. However, the lack of FP8 support means quantized inference is less efficient than Ada Lovelace or Blackwell.
Conselho de compra
Excelente escolha para IA local
Roda 27 de 50 modelos principais bem — um ótimo coringa para inferência local.
40.0 GB
VRAM
$10,000
Preço sugerido
$250/GB
Custo por GB de VRAM
Melhores modelos para esta GPU
What will limit you first
This setup is broadly balanced for this model.
No major red flags
This recommendation has enough memory headroom and acceptable estimated speed for the selected workload.
Best upgrade itinerary
Unlocks 11 additional models that do not fit on the current setup.
Quer mais margem? MacBook Pro M1 Max 64GB (64.0 GB unified memory) é o próximo passo.
Cost vs cloud API
Assumes 4 hours/day of active inference at 166 tok/s, NVIDIA A100 40GB amortized over 36 months, US residential electricity ($0.15/kWh), blended cloud pricing at $10 per 1M tokens (GPT-4o / Claude Sonnet tier).
71.7M
Tokens/month at this pace
$227
Monthly local cost
$717
Same tokens on cloud API
$3.16
Local $/1M tokens
Break-even: pays for itself in 11.2 months vs cloud API at this workload. Price reference: $8.0k 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 fits natively with comfortable headroom. Known channels: huggingface, 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 fits natively with comfortable headroom. 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
SThis model is a direct match for rag. It belongs to a current frontier family for local AI. It fits natively with comfortable headroom. Known channels: huggingface, ollama, lm-studio.
Qwen 3.6 35B A3BQwen3-Coder 30B A3B InstructQwen 3.5 35B A3BQwen 3.5 27BQwen3-VL 30B A3B InstructQwen 3 32BQwen 3 30B A3BNemotron 3 Nano 30B
Magistral Small 2507Devstral Small 2 24B InstructQwen 3.6 27BNemotron Cascade 2 30B A3BDevstral Small 1.1
GPT-OSS 20B
EXAONE 4.0 32BQwen 3 14B
Phi-4-reasoning-plus 14BQwen 3.5 9B
Gemma 4 26B A4BGemma 4 31BQwen 3 8BQwen 3.5 4BMinistral 3 14BNemotron Nano 8BPhi-4 Mini Reasoning 4B
Jina Embeddings v3
BGE M3Qwen 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 InstructLeanstral 119B A6BDeepSeek Coder V2 236BDeepSeek R1 671BDeepSeek V3.1 671BQuase ao alcance
Modelos de alta qualidade que precisam de um pouco mais de memória
Image & Video Generation
45 of 52 models can generate images or video on your NVIDIA A100 40GB
| Model | Max Resolution | Gen Time | Grade |
|---|---|---|---|
| SD TurboImage | 512×512 | 100ms | S |
| Stable Diffusion 1.5Image | 512×768 | 200ms | S |
| Realistic Vision v5.1Image | 512×768 | 200ms | S |
| DreamShaper 8Image | 512×768 | 200ms | S |
| LCM DreamShaper v7Image | 512×768 | 100ms | S |
| PixArt-SigmaImage | 1024×1024 | ~1s | S |
| FramePack I2VVideo | 256×256 | ~3s/frame | S |
| SDXL TurboImage | 512×512 | 100ms | S |
| SDXL LightningImage | 1024×1024 | 400ms | S |
| Stable Diffusion XL 1.0Image | 1024×1024 | ~1s | S |
| Playground v2.5Image | 1024×1024 | ~1.4s | S |
| RealVisXL v5.0Image | 1024×1024 | ~1.1s | S |
| DreamShaper XLImage | 1024×1024 | ~1.1s | S |
| Juggernaut XL v9Image | 1024×1024 | ~1.1s | S |
| Animagine XL 3.1Image | 1024×1024 | ~1.1s | S |
| Pony Diffusion V6 XLImage | 1024×1024 | ~1.1s | S |
| Animagine XL 4.0Image | 1024×1024 | ~1.1s | S |
| Illustrious XLImage | 1024×1024 | ~1.1s | S |
| Wan Video 2.1 1.3BVideo | 480×832 | 700ms/frame | S |
| Stable Diffusion 3.5 MediumImage | 1024×1024 | ~1.7s | S |
| Flux.2 Klein 4BImage | 1024×1024 | 300ms | S |
| LTX Video 2BVideo | 1280×720 | 800ms/frame | S |
| KolorsImage | 1024×1024 | ~1.9s | S |
| Stable CascadeImage | 1024×1024 | ~2.4s | S |
| AuraFlow v0.3Image | 1536×1536 | ~4.3s | S |
| Stable Diffusion 3.5 LargeImage | 1024×1024 | ~5.2s | S |
| Stable Diffusion 3.5 Large TurboImage | 1024×1024 | ~1s | S |
| CogVideoX 2BVideo | 720×480 | 800ms/frame | S |
| HunyuanVideoVideo | 256×256 | ~3s/frame | S |
| ChromaImage | 1024×1024 | ~1s | S |
| Z-Image TurboImage | 1536×1536 | ~1s | S |
| Flux.1 DevImage | 1024×1024 | ~4.3s | S |
| Flux.1 SchnellImage | 1024×1024 | 800ms | S |
| LTX Video 13BVideo | 256×256 | ~3s/frame | S |
| Flux.1 Kontext DevImage | 1024×1024 | ~4.8s | S |
| AnimateDiff v1.5.3Video | 512×768 | 400ms/frame | S |
| Cosmos Diffusion 7BVideo | 1024×576 | ~1.4s/frame | S |
| CogVideoX 5BVideo | 720×480 | ~1.2s/frame | S |
| Wan2.2 TI2V 5BVideo | 832×480 | ~1.2s/frame | S |
| Flux.2 Klein 9BImage | 1024×1024 | 500ms | S |
| Flux.1 Fill DevImage | 1024×1024 | ~4s | S |
| Mochi 1 PreviewVideo | 848×480 | ~1.6s/frame | B |
| HunyuanVideo 1.5Video | 256×256 | ~1.5s/frame | B |
| Helios 14BVideo | 256×256 | ~3.1s/frame | D |
| SkyReels V2 14BVideo | 256×256 | ~3.1s/frame | D |
| Wan Video 2.1 14BVideo | 256×256 | ~3.1s/frame | F |
| Wan Video 2.2 14BVideo | 256×256 | ~3.1s/frame | F |
| Qwen ImageImage | 256×256 | ~1.6s | F |
| Qwen Image EditImage | 256×256 | ~1.6s | F |
| Flux.2 DevImage | 256×256 | ~45s | F |
| MAGI-1Video | 256×256 | ~2.2s/frame | F |
| HunyuanImage 3.0Image | 256×256 | ~2.8s | 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 22% scaling overhead.
| Config | Effective memory | Models that fit | Est. bandwidth |
|---|---|---|---|
| 1× NVIDIA | 40 GB | 325/374 | 1,555 GB/s |
| 2× NVIDIA | 80 GB | 350/374 | 2,426 GB/s |
| 4× NVIDIA | 160 GB | 359/374 | 4,852 GB/s |
Model counts use default quantization at coding workload settings. Multi-GPU scaling factor: 0.78× per additional GPU.
Upgrade paths
See what you unlock with more powerful hardware
Opções de upgrade
4× NVIDIA A100 40GBMulti-GPUUnlocks 34 additional models that do not fit on the current setup.
Lifts average decode speed across fitting models by about 48%.
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.
~$10,000 MSRP
Unlocks 11 additional models that do not fit on the current setup.
~$2,499 MSRP
RTX PRO 5000 Blackwell 48GBUpgrade NVIDIAUnlocks 13 additional models that do not fit on the current setup.
~$4,999 MSRP
Unlocks 26 additional models that do not fit on the current setup.
~$2,499 MSRP
Unlocks 39 additional models that do not fit on the current setup.
Lifts average decode speed across fitting models by about 70%.
~$8,000 MSRP
NVIDIA A100 40GB (40 GB VRAM) can run these top models: Qwen 3.6 35B A3B (score: 99/100), Qwen3-Coder 30B A3B Instruct (score: 98/100), Qwen 3.5 35B A3B (score: 98/100). See the full compatibility list above.
NVIDIA A100 40GB has 40 GB of VRAM available for AI model inference. This determines which models and quantization levels you can run locally.
Yes, NVIDIA A100 40GB is excellent for running LLMs locally with top compatibility scores above 80/100.
For coding on NVIDIA A100 40GB, we recommend Qwen 3.6 27B. It achieves 53.4 tokens per second with 262K context window. This model is a direct match for coding. It belongs to a current frontier family for local AI. It fits natively with comfortable headroom. Known channels: huggingface, lm-studio.
There are 5 upgrade path(s) from NVIDIA A100 40GB: NVIDIA A100 40GB, MacBook Pro M1 Max 64GB. Upgrading would unlock larger models and faster inference speeds.
Yes, NVIDIA A100 40GB with 40 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.
NVIDIA A100 40GB (40 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.
NVIDIA A100 40GB is excellent for AI image generation. With 40 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, NVIDIA A100 40GB with 40 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 40 GB on NVIDIA A100 40GB, 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.
NVIDIA A100 40GB 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.
NVIDIA A100 40GB supports up to 4× GPU scaling via PCIe. With 4× GPUs, you get 160 GB effective memory with a 0.78× 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.
NVIDIA A100 40GB uses PCIe for multi-GPU communication, which has approximately 22% scaling overhead. For best multi-GPU performance, consider NVLink-equipped variants.
Usually yes. If you want to run 2-4× NVIDIA A100 40GB 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
