AI hardware for running models on premises
Data center and workstation GPUs, compact AI systems and the NPU-equipped chips in AI PCs, with memory, bandwidth and compute as each maker publishes them. For running language models, memory is the first limit and memory bandwidth the second.
Models
56
Makers
7
Groups
4
Sources
Makers' own spec pages
Compact AI systems and workstations
| Model | Memory | Memory type | Bandwidth | Platform |
|---|---|---|---|---|
| NVIDIA DGX Station (GB300 Grace Blackwell Ultra) | 748 GB | 252 GB HBM3e (GPU) + 496 GB LPDDR5X (CPU), coherent via NVLink-C2C | 7,100 GB/s | 20,000 TOPS |
| Apple Mac Studio (M3 Ultra, 2025) | 512 GB | Unified memory | 819 GB/s | - |
| Apple Mac Studio (M5 Ultra, 2026) | 512 GB | Unified memory | 1,200 GB/s | - |
| Apple Mac Studio (M4 Max, 2025) | 128 GB | Unified memory | 546 GB/s | - |
| Apple Mac Studio (M5 Max, 2026) | 128 GB | Unified memory | 614 GB/s | - |
| Framework Framework Desktop (AMD Ryzen AI Max+ 395) | 128 GB | LPDDR5x-8000 unified, 256-bit (up to 96GB assignable to GPU) | 256 GB/s | 126 TOPS |
| HP HP Z2 Mini G1a (AMD Ryzen AI Max+ PRO 395) | 128 GB | LPDDR5X-8533 ECC unified (transfer rates up to 8000 MT/s; up to 96GB assignable to GPU) | - | - |
| NVIDIA DGX Spark (GB10 Grace Blackwell) | 128 GB | LPDDR5x coherent unified, 256-bit | 273 GB/s | 1,000 TOPS |
| NVIDIA Jetson AGX Thor Developer Kit (T5000) | 128 GB | LPDDR5X, 256-bit | 273 GB/s | 2,070 TOPS |
| Apple Mac mini (M4 Pro, 2024) | 64 GB | Unified memory | 273 GB/s | - |
| Apple Mac mini (M5 Pro, 2026) | 64 GB | Unified memory | 307 GB/s | - |
| NVIDIA Jetson AGX Orin 64GB Developer Kit | 64 GB | LPDDR5, 256-bit | 204.8 GB/s | 275 TOPS |
Workstation GPUs
| Model | Memory | Memory type | Bandwidth | Platform |
|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 96 GB | GDDR7 ECC | 1,792 GB/s | 3,511 TOPS |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | 96 GB | GDDR7 ECC | 1,597 GB/s | - |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | GDDR7 ECC | 1,792 GB/s | 4,000 TOPS |
| AMD Radeon PRO W7900 | 48 GB | GDDR6 ECC | 864 GB/s | - |
| NVIDIA RTX 6000 Ada Generation | 48 GB | GDDR6 ECC | 960 GB/s | 1,457 TOPS |
| NVIDIA RTX PRO 5000 Blackwell | 48 GB | GDDR7 ECC | 1,344 GB/s | 2,064 TOPS |
| AMD Radeon AI PRO R9700 | 32 GB | GDDR6 (ECC on Linux only) | 640 GB/s | - |
| AMD Radeon PRO W7800 | 32 GB | GDDR6 ECC | 576 GB/s | - |
| NVIDIA GeForce RTX 5090 | 32 GB | GDDR7 | 1,792 GB/s | 3,352 TOPS |
| NVIDIA RTX PRO 4500 Blackwell | 32 GB | GDDR7 ECC | 896 GB/s | 1,617 TOPS |
| NVIDIA RTX PRO 4000 Blackwell | 24 GB | GDDR7 ECC | 672 GB/s | 1,290 TOPS |
| NVIDIA GeForce RTX 5080 | 16 GB | GDDR7 | 960 GB/s | 1,801 TOPS |
Data center GPUs
| Model | Memory | Memory type | Bandwidth | FP16 |
|---|---|---|---|---|
| AMD Instinct MI350X | 288 GB | HBM3E | 8,000 GB/s | 2,300 TFLOPS |
| AMD Instinct MI355X | 288 GB | HBM3E | 8,000 GB/s | 2,500 TFLOPS |
| NVIDIA B300 (HGX, Blackwell Ultra SXM) | 288 GB | HBM3e | 8,000 GB/s | 2,250 TFLOPS |
| AMD Instinct MI325X | 256 GB | HBM3E | 6,000 GB/s | 1,300 TFLOPS |
| AMD Instinct MI300X | 192 GB | HBM3 | 5,300 GB/s | 1,300 TFLOPS |
| NVIDIA B200 (HGX, SXM) | 180 GB | HBM3e | 8,000 GB/s | 2,250 TFLOPS |
| NVIDIA H200 NVL | 141 GB | HBM3e | 4,800 GB/s | 835.5 TFLOPS |
| NVIDIA H200 SXM | 141 GB | HBM3e | 4,800 GB/s | 989.5 TFLOPS |
| Intel Gaudi 3 (HL-325L OAM) | 128 GB | HBM2E | 3,700 GB/s | 1,678 TFLOPS |
| Intel Gaudi 3 (HL-338 PCIe) | 128 GB | HBM2E | 3,700 GB/s | - |
| NVIDIA H100 NVL | 94 GB | HBM3 | 3,900 GB/s | 835.5 TFLOPS |
| NVIDIA H100 SXM | 80 GB | HBM3 | 3,350 GB/s | 989.5 TFLOPS |
| NVIDIA L40S | 48 GB | GDDR6 | 864 GB/s | 362.1 TFLOPS |
| NVIDIA L4 | 24 GB | GDDR6 | 300 GB/s | 121 TFLOPS |
AI PC processors (NPUs)
Before you shortlist
Published specifications are the maker's best case. Ask for a site survey or a trial on your own floor, parts and workflow before you commit, and read running ai on your own hardware first.
Every entry links to the page its figures came from. If a maker updates a specification, tell us through the contact form.