Skip to content

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.

Compact AI systems and workstations

ModelMemoryMemory typeBandwidthPlatform
NVIDIA DGX Station (GB300 Grace Blackwell Ultra)748 GB252 GB HBM3e (GPU) + 496 GB LPDDR5X (CPU), coherent via NVLink-C2C7,100 GB/s20,000 TOPS
Apple Mac Studio (M3 Ultra, 2025)512 GBUnified memory819 GB/s-
Apple Mac Studio (M5 Ultra, 2026)512 GBUnified memory1,200 GB/s-
Apple Mac Studio (M4 Max, 2025)128 GBUnified memory546 GB/s-
Apple Mac Studio (M5 Max, 2026)128 GBUnified memory614 GB/s-
Framework Framework Desktop (AMD Ryzen AI Max+ 395)128 GBLPDDR5x-8000 unified, 256-bit (up to 96GB assignable to GPU)256 GB/s126 TOPS
HP HP Z2 Mini G1a (AMD Ryzen AI Max+ PRO 395)128 GBLPDDR5X-8533 ECC unified (transfer rates up to 8000 MT/s; up to 96GB assignable to GPU)--
NVIDIA DGX Spark (GB10 Grace Blackwell)128 GBLPDDR5x coherent unified, 256-bit273 GB/s1,000 TOPS
NVIDIA Jetson AGX Thor Developer Kit (T5000)128 GBLPDDR5X, 256-bit273 GB/s2,070 TOPS
Apple Mac mini (M4 Pro, 2024)64 GBUnified memory273 GB/s-
Apple Mac mini (M5 Pro, 2026)64 GBUnified memory307 GB/s-
NVIDIA Jetson AGX Orin 64GB Developer Kit64 GBLPDDR5, 256-bit204.8 GB/s275 TOPS

Workstation GPUs

ModelMemoryMemory typeBandwidthPlatform
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition96 GBGDDR7 ECC1,792 GB/s3,511 TOPS
NVIDIA RTX PRO 6000 Blackwell Server Edition96 GBGDDR7 ECC1,597 GB/s-
NVIDIA RTX PRO 6000 Blackwell Workstation Edition96 GBGDDR7 ECC1,792 GB/s4,000 TOPS
AMD Radeon PRO W790048 GBGDDR6 ECC864 GB/s-
NVIDIA RTX 6000 Ada Generation48 GBGDDR6 ECC960 GB/s1,457 TOPS
NVIDIA RTX PRO 5000 Blackwell48 GBGDDR7 ECC1,344 GB/s2,064 TOPS
AMD Radeon AI PRO R970032 GBGDDR6 (ECC on Linux only)640 GB/s-
AMD Radeon PRO W780032 GBGDDR6 ECC576 GB/s-
NVIDIA GeForce RTX 509032 GBGDDR71,792 GB/s3,352 TOPS
NVIDIA RTX PRO 4500 Blackwell32 GBGDDR7 ECC896 GB/s1,617 TOPS
NVIDIA RTX PRO 4000 Blackwell24 GBGDDR7 ECC672 GB/s1,290 TOPS
NVIDIA GeForce RTX 508016 GBGDDR7960 GB/s1,801 TOPS

Data center GPUs

ModelMemoryMemory typeBandwidthFP16
AMD Instinct MI350X288 GBHBM3E8,000 GB/s2,300 TFLOPS
AMD Instinct MI355X288 GBHBM3E8,000 GB/s2,500 TFLOPS
NVIDIA B300 (HGX, Blackwell Ultra SXM)288 GBHBM3e8,000 GB/s2,250 TFLOPS
AMD Instinct MI325X256 GBHBM3E6,000 GB/s1,300 TFLOPS
AMD Instinct MI300X192 GBHBM35,300 GB/s1,300 TFLOPS
NVIDIA B200 (HGX, SXM)180 GBHBM3e8,000 GB/s2,250 TFLOPS
NVIDIA H200 NVL141 GBHBM3e4,800 GB/s835.5 TFLOPS
NVIDIA H200 SXM141 GBHBM3e4,800 GB/s989.5 TFLOPS
Intel Gaudi 3 (HL-325L OAM)128 GBHBM2E3,700 GB/s1,678 TFLOPS
Intel Gaudi 3 (HL-338 PCIe)128 GBHBM2E3,700 GB/s-
NVIDIA H100 NVL94 GBHBM33,900 GB/s835.5 TFLOPS
NVIDIA H100 SXM80 GBHBM33,350 GB/s989.5 TFLOPS
NVIDIA L40S48 GBGDDR6864 GB/s362.1 TFLOPS
NVIDIA L424 GBGDDR6300 GB/s121 TFLOPS

AI PC processors (NPUs)

ModelMemoryMemory typeBandwidthNPU
AMD Ryzen AI 7 350256 GBLPDDR5X-8000 or DDR5-5600-50 TOPS
AMD Ryzen AI 9 HX 375256 GBLPDDR5X-8000 or DDR5-5600-55 TOPS
Intel Core Ultra 9 285K (Arrow Lake-S desktop, Core Ultra 200S)256 GBDDR5-6400-13 TOPS
AMD Ryzen AI Max+ 395 (Strix Halo)128 GBLPDDR5X-8000, 256-bit256 GB/s50 TOPS
Apple M4 Max128 GB-546 GB/s-
Apple M5 Max128 GB-614 GB/s-
Intel Core Ultra 7 265U (Arrow Lake U, Core Ultra 200U)128 GBLPDDR5/X-8400 or DDR5-6400-12 TOPS
Intel Core Ultra 9 285H (Arrow Lake H, Core Ultra 200H)128 GBLPDDR5/X-8400 or DDR5-6400-13 TOPS
Qualcomm Snapdragon X2 Elite Extreme X2E-96-100128 GBLPDDR5X-9523, 192-bit228 GB/s80 TOPS
Qualcomm Snapdragon X2 Elite X2E-90-100128 GBLPDDR5X-9523, 128-bit152 GB/s85 TOPS
Intel Core Ultra X9 388H (Panther Lake, Core Ultra Series 3)96 GBLPDDR5X-9600-50 TOPS
Apple M4 Pro64 GB-273 GB/s-
Apple M5 Pro64 GB-307 GB/s-
Qualcomm Snapdragon X Elite X1E-84-10064 GBLPDDR5X-8448135 GB/s45 TOPS
Qualcomm Snapdragon X Plus X1P-64-10064 GBLPDDR5X-8448135 GB/s45 TOPS
Apple M432 GB-120 GB/s38 TOPS
Apple M532 GB-153 GB/s-
Intel Core Ultra 9 288V (Lunar Lake, Core Ultra 200V)32 GBLPDDR5X-8533 (on-package)-48 TOPS

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.

Free quotes

Get quotes for AI hardware or services

Free, no obligation. Pricing depends on your site and volume, so a quote is the only real price. Your request comes to us, and we pass it to an authorized dealer or installer serving your area.

By submitting you agree we may pass your name, company, email and ZIP to an authorized dealer serving your area so they can contact you about this request. They may pay us for the introduction. See our privacy policy.