Comparison · hardware
A lasting local mind needs memory more than anything: a mid-size open model (about 30 billion parameters) needs roughly 17 GB at 4-bit, plus room for long context and everything else a robot runs. That rules out most robot computers (8–16 GB). This comparison covers every robot module and small computer we found with 64 GB or more, plus the latest Mac mini, the smallest Apple computer that qualifies.
Prismatic's two chosen hosts come first, labelled; the others follow alphabetically and are ranked by nobody. Speeds are estimates from memory bandwidth unless a benchmark is shown.
| Chip | Role | Memory | Bandwidth | AI compute | Power | Price | Est. speed, ~30B model | Robot-ready |
|---|---|---|---|---|---|---|---|---|
| Mac mini with M5 Pro, 64 GB · Apple — Prismatic's retrofit hostWe make it | Desk-size computer: Prismatic's retrofit site control centre | 64 GB Published | 307 GB/s Published | Up to 18-core CPU, 20-core GPU with Neural Accelerators | 6 W idle, 145 W CPU maximum (Apple, wall); 155 W maximum continuous rating; not yet measured by Prismatic Published | From US$1,699; about US$2,699 with 64 GB (estimate from Apple configuration prices) Estimate | ~9.9 tokens/s Estimate | No: mains-powered desk computer; sits beside the machines |
| Jetson AGX Orin 64 GB · NVIDIA — Prismatic's greenfield hostWe make it | Robot module: Prismatic's greenfield host | 64 GB Published | 204.8 GB/s Published | Up to 275 TOPS; 2048-core Ampere GPU, 12-core Arm CPU | 15–60 W modes Published | Module US$2,999 (1,000+ units); developer kit US$3,499 Published | ~6.6 tokens/s Estimate | Yes: built for robots; industrial version available |
| DGX Spark (GB10) · NVIDIA | Desk AI computer | 128 GB Published | 273 GB/s Reported | Up to 1 petaFLOP FP4; 20-core Grace CPU | Not disclosed here Not disclosed | US$4,699 (after a February 2026 price rise, reported) Reported | ~8.8 tokens/s Estimate | No: desk computer |
| Dragonwing IQ10 · Qualcomm | Robotics reference design | 64 GB Published | Not disclosed Not disclosed | 700 TOPS (up to 2,000 with expansion); 18-core Oryon CPU | 12 V or 24 V input; -40 to 70 °C Published | Not disclosed Not disclosed | Not estimated | Yes: for AMRs, industrial machines and humanoids |
| Jetson AGX Orin Industrial · NVIDIA | Rugged robot module | 64 GB Published | 204.8 GB/s Published | Up to 248 TOPS | 15–75 W Reported | US$3,199 (1,000+ units) Published | ~6.6 tokens/s Estimate | Yes: wide temperature, vibration, extended life |
| Jetson T4000 (Thor family) · NVIDIA | Robot module | 64 GB Published | Not disclosed Not disclosed | 1,200 FP4 TFLOPS (Blackwell) | 40–70 W Reported | US$2,999 (1,000+ units) Published | Not estimated | Yes: for humanoids and advanced robots |
| Jetson T5000 (AGX Thor) · NVIDIA | Robot module | 128 GB Published | 273 GB/s Published | 2,070 FP4 TFLOPS; 14-core Neoverse-V3AE CPU | 40–130 W Published | Module US$4,999 (1,000+ units); developer kit US$5,499 Published | ~8.8 tokens/s Estimate | Yes: built for humanoids |
| Ryzen AI Max+ 395 (mini PCs) · AMD | Mini PC chip (e.g. GMKtec EVO-X2) | 128 GB Published | 256 GB/s Reported | 16-core CPU; 40-CU RDNA 3.5 GPU; up to 96 GB to the GPU | Up to ~120 W in performance modes (reported) Reported | About US$3,299 for 128 GB (GMKtec EVO-X2, reported) Reported | ~8.3 tokens/s Estimate | No: mini PC |
Speed for one conversation is limited mainly by memory bandwidth; long prompts are limited by compute. Power matters for batteries; price matters for fleets. Prismatic's two hosts: a 64 GB Mac mini (M5 Pro) beside machines that already work (retrofit), and the Jetson AGX Orin 64 GB inside new machines (greenfield sites). The 16 GB Mac mini it was proven on is too small for industrial work. Neither 64 GB host is yet measured by Prismatic.
Pick a chip, then the model size. Speed is estimated from memory bandwidth: the main limit for answering one person at a time.
How it works: weights = parameters × bits ÷ 8 × 1.05; needed = weights + context × memory per 1K + 4 GB overhead; speed ≈ bandwidth × efficiency ÷ weights. Estimates for one conversation at a time. Prefill (reading a long prompt) depends on compute, not bandwidth, and is not shown.
What the computer costs to run, around the clock, and over five years.
How it works: kWh = watts × hours × 365 ÷ 1,000; cost = kWh × price per kWh.
For robots and off-grid machines.
How it works: hours = battery × efficiency ÷ watts.
Our review. The latest Mac mini (released 22 Sept 2026) and the fastest memory here at the lowest price: the natural site box for machines that already exist.
Gaps. Not a robot computer (mains power, no rugged or wide-temperature version); Apple software stack.
Benchmark found: 15–22 tokens/s reported for a 32B 4-bit model with MLX (reported)
Sources: apple.com · macworld.com · promptquorum.com
Our review. The proven robot module, used today inside cobots, tractors and AMRs, with a 15 W mode for battery life.
Gaps. Slowest memory here; an older (Ampere) generation.
Benchmark found: Runs 70B-class models at 4-bit per community reports; ~30B speed not published (reported)
Sources: developer.nvidia.com · thinkrobotics.com
Our review. Twice the memory of the others, for bigger models on a desk.
Gaps. Not a robot computer; the most expensive box here.
Benchmark found: Not published here
Sources: nvidia.com · notebookcheck.net
Our review. A serious new robot platform with 64 GB ECC memory and rugged ratings; general availability September 2026.
Gaps. Memory bandwidth and large-model speed not published; model support for mid-size open models still arriving.
Benchmark found: Not published
Sources: docs.qualcomm.com · hothardware.com
Our review. The Orin for harsh machines: farms, mines, vehicles.
Gaps. Same memory speed as the Orin 64 GB.
Benchmark found: As Orin 64 GB
Sources: developer.nvidia.com · developer.nvidia.com
Our review. Blackwell-generation robot module with 64 GB at the Orin's module price.
Gaps. Higher minimum power (40 W); memory bandwidth not found.
Benchmark found: Not published
Sources: developer.nvidia.com · waveshare.com
Our review. The most powerful robot module: 128 GB and Blackwell, 7.5x Orin's AI compute.
Gaps. The most power-hungry module (40 W minimum) and the highest price.
Benchmark found: Qwen3 30B-A3B mixture-of-experts at 61 tokens/s in MLPerf (published by NVIDIA; a sparse model, not a dense 30B)
Sources: developer.nvidia.com · cnx-software.com · developer.nvidia.com
Our review. Lots of memory at a mini-PC price; strong community support.
Gaps. Not a robot computer; x86 mini PCs vary in build quality.
Benchmark found: ~31 tokens/s on a sparse 120B model (reported)
Sources: runaihome.com · layer3labs.io
The maker's view
For patient investors
Retrofit sites pay first on a 64 GB Mac mini (M5 Pro) beside machines that already work. Greenfield machines carry the same mind inside, on the Jetson AGX Orin 64 GB, licensed per machine. The horizon (blackouts, wilderness, space) is where a mind that needs no cloud is not a feature but the point.
Memory bandwidth of the Jetson T4000 and Qualcomm IQ10; typical power of the Mac mini M5 Pro and DGX Spark under AI load.
A quick introduction to the names on this page. Staff counts describe the stated organisation and reporting period; they are not live totals.
Designs computer processors and graphics chips for PCs, servers and AI systems. [1]
Sources checked 2026-10-05. Profiles are current background, not a statement of who held a role on the article date.
Makes iPhones, Mac computers and other devices, along with their software and online services. [1]
Sources checked 2026-10-05. Profiles are current background, not a statement of who held a role on the article date.
Builds human-shaped robots intended to perform physical work. [1]
Sources checked 2026-10-05. Profiles are current background, not a statement of who held a role on the article date.
Designs processors and software used for AI, graphics and large computing workloads. [1]
Sources checked 2026-10-05. Profiles are current background, not a statement of who held a role on the article date.
Designs chips and wireless technology used in phones, vehicles and other connected devices. [1]
Sources checked 2026-10-05. Profiles are current background, not a statement of who held a role on the article date.
Prismatic Agents