Specs CalculatorROI CalculatorWorkstation BuilderModel TreemapGap TrackerStatus TrackerTools
Compare GPUs, edge devices, TPUs, and more. Find the right hardware for your AI workloads with real specs and benchmarks.
Get notified when new GPUs, edge devices, and AI accelerators are added to the directory.
| Manufacturer | Tags | Compare | ||||
|---|---|---|---|---|---|---|
| ACEMAGIC M1A Pro (i9-13900HK + ARC A770) | ACEMAGIC | 16 GB | — | $799 | Edge AIBudget Friendly | |
| Acer Veriton GN100 AI Mini | Acer | 128 GB | 29.71 | $3,999 | Edge AIEnterprise | |
| AMD Instinct MI300X | AMD | 192 GB | 1307.4 | $22,000 | Best for LLMsEnterprise+3 | |
| AMD Instinct MI325X | AMD | 256 GB | 1307.4 | $30,000 | Best for LLMsEnterprise+3 | |
| AMD Instinct MI355X | AMD | 288 GB | — | — | Best for LLMsEnterprise+3 | |
| AMD Radeon RX 7600 8GB | AMD | 8 GB | 36.7 | $269 | Budget FriendlyEnergy Efficient | |
| AMD Radeon RX 7700 XT | AMD | 12 GB | 55.3 | $449 | Budget FriendlyBest for Computer Vision | |
| AMD Radeon RX 7800 XT | AMD | 16 GB | 74.6 | $499 | Best for Computer VisionBudget Friendly | |
Ad PayPerQPay-per-query access to top LLMs without a subscription. Use any model on demand.Try PayPerQ | ||||||
| AMD Radeon RX 7900 XT | AMD | 20 GB | 103 | $899 | Premium / High-EndBest for Computer Vision | |
| AMD Radeon RX 7900 XTX | AMD | 24 GB | 122.8 | $999 | Premium / High-EndBest for Computer Vision | |
| AMD Radeon RX 9070 | AMD | 16 GB | — | $549 | Best for Computer VisionBudget Friendly | |
| AMD Radeon RX 9070 XT | AMD | 16 GB | — | $599 | Best for Computer VisionPremium / High-End | |
| AMD Ryzen AI 9 HX 370 | AMD | — | — | — | Mobile / On-DeviceEnergy Efficient+2 | |
| Apple M3 Ultra (32-core CPU, 80-core GPU) | Apple | 512 GB | — | $5,999 | Best for LLMsPremium / High-End+1 | |
| Apple M4 | Apple | 32 GB | — | $1,399 | Mobile / On-DeviceEnergy Efficient+1 | |
The right AI hardware is whatever runs the models you actually use at a speed you can live with, for a total cost that beats renting.
The answer depends on three things: how large the models you want to run are, whether you need to train or just infer, and how many hours per day you spend at the machine. A 16GB RTX or a 16GB Apple Silicon Mac covers most 7B to 13B models, while 70B models need at least 48GB of VRAM or a high-memory Mac.
This directory normalizes specs across NVIDIA, AMD, Apple, and edge vendors, scores compatibility per model, and links to live retail prices and cloud rental rates so you can match a card or device to a real workload instead of a marketing claim.
Side by Side
A practical comparison of unified memory Macs and discrete RTX GPUs for running AI models locally. Specs, throughput, and a plain-English buying guide.
Not Ready to Buy?
Live cloud GPU rental prices across RunPod and Vast.ai. Find the cheapest H100, A100, or RTX 4090 right now, with a 30-day price trend per card.
For 7B to 13B models, a 16GB consumer GPU like an RTX 4080 or a 16GB Apple Silicon Mac is enough. For 30B to 70B models, you need 24GB to 48GB of VRAM, which usually means an RTX 4090, RTX 5090, or a 64GB unified memory Mac. Anything above 70B needs a workstation card like the RTX 6000 Ada or two GPUs in parallel.
Apple Silicon wins on power draw, quiet operation, and unified memory, which lets a 64GB or 128GB Mac run very large models in one shot. NVIDIA RTX wins on raw speed, software support, and price per token. Pick Apple if you value a quiet laptop and want to run big quantized models offline. Pick NVIDIA if you need the fastest inference or any kind of training.
A 70B model at FP16 needs about 140GB of VRAM, which only data-center cards reach. At 4-bit quantization, the same model fits in 40GB to 48GB, which means an RTX 4090 with offload, an RTX 6000 Ada, or a 64GB unified memory Mac. Each detail page on this site shows the exact VRAM each model needs at every quantization level.
For occasional or short workloads, renting on RunPod or Vast.ai is much cheaper than buying. The break-even point is roughly 4 to 8 hours of use per day, depending on the card. A $2,000 RTX 4090 pays back in around 6 months at heavy use, but never if you only run it for a few hours a week. Our ROI calculator runs the math for your specific usage.
Specs are normalized from vendor data sheets, then verified against community testing. Retail prices link directly to Amazon or the manufacturer when in stock. GPU rental prices are pulled live from RunPod and Vast.ai every hour. New launches are added within days of release.
For every model on this site, we compute the VRAM it needs at each quantization level, then check whether each piece of hardware can hold it. The compatibility table also estimates tokens per second using the card’s memory bandwidth and FP16 or INT8 throughput, so you see speed and fit together. Numbers are estimates, not guarantees.
Edge devices like the Jetson Orin and Coral TPU are built for low-power inference at the device, not the cloud. AI PCs from Intel, AMD, and Qualcomm bundle a neural accelerator with a CPU and GPU for laptop-grade local inference. Use the category filter at the top of the directory to compare them side by side.