Machines · 5 boxes · 4 chip makers

Five boxes.One job each.

NVIDIA, AMD, Intel and Apple, 1,263 TOPS between them. Each machine has a role that suits its silicon. Here they are one at a time, with the numbers they actually post.

  1. 01DGX SparkModel host
  2. 02Strix HaloImage & video
  3. 03Intel NUCVoice
  4. 04Mac mini M4 ProControl plane
  5. 05Mac mini M6Dev server

TOPS are vendor peak. All figures read from LabScope, 26 Oct 2026.

01 / 05NVIDIARole: model host

The one that thinks.

Lid
Gold-mesh chassis, lifted away
Chip
GB10 Grace Blackwell · 20 Arm cores
Memory
128 GB unified

1,000TOPS

79% of the lab in one box.

AI compute · up to 1 PFLOP FP4

Runs

  • vLLMServing Qwen 3.8 27B and Qwen 3.6 35B (mixture-of-experts, 3B active)
  • SparkChatPrivate chat with an off-the-record mode

Benchmarks · tokens per second

A 35-billion-parameter model that only wakes 3 billion per token. That's why it's the fast one.

Illustrated render of the AMD Strix Halo mini PC under a single overhead light

02 / 05AMDRole: image & video

The one that draws.

Chip
Ryzen AI Max+ 395 with Radeon 8060S
Cores
32 threads
Memory
64 GB + 64 GB GPU carve-out
AI compute
126 TOPS

Benchmarks · ComfyUI and Whisper

357

Qwen Image · images

0

Failures

32s

Median per image

19s

Best per image

~4.7min

LTX 2.5 · per clip, median

100%

Whisper · word accuracy

Runs

  • ComfyUIQwen Image and LTX 2.5 22B video
  • OllamaLocal models
  • Unsloth StudioFine-tuning
  • WhisperSpeech-to-text

357 runs, not one failure.

03 / 05IntelRole: voice

The one that listens.

Speech goes direct on the lab network by design. Latency matters more than a gateway hop.

Illustrated render of the Intel NUC mini PC

Whisper · transcript

Speech stays on the LAN.

Chip
Core Ultra 9 285H
Cores
16 threads
Memory
64 GB
AI compute
99 TOPS

Runs

  • WhisperSpeech-to-text
  • KokoroText-to-speech
  • Dify, n8n, FlowiseWorkflows
  • AgentsAgent apps

Where the 99 TOPS come from

NPU Arc iGPU CPU 99 TOPS

236

speech calls across the lab in the last 30 days

Gateway and control plane for the other four boxes

04 / 05AppleRole: control plane

The one that routes.

Chip
Apple M4 Pro
Cores
12
Memory
24 GB
AI compute
38 TOPS Neural Engine

Runs

  • LiteLLMThe AI gateway
  • AILogicThe control plane
  • AppsHosted alongside

Last 30 days through it

10,635

AI calls

62.4M

tokens — 62.4 million

5ms

median gateway overhead

The smallest TOPS figure here, and the box everything else answers to.

05 / 05AppleRole: dev server

The one that watches.

Chip
Apple M6
Cores
12
Memory
24 GB
AI compute
Not published

Runs

  • LabScopeThe lab's own monitor
  • DmotionAI video studio
  • DpagesPrompt to animated websites, decks and infographics
  • CbuseHosted here
Illustrated render of the Mac mini M6 dev server

LabScope · one scope channel per server

Metrics every

15s

Inventory every

60s

History kept

30days

Uptime · 30 days

99.993%

See LabScope

All five · one table

Side by side.

The five machines of DEVNET AI LAB compared
MachineMakerCoresMemoryAI computeRole
DGX SparkNVIDIA20 Arm128 GB unified1,000 TOPS79% of the labModel host
Strix HaloAMD32 threads64 + 64 GB126 TOPS10%Image & video
Intel NUCIntel16 threads64 GB99 TOPS8%Voice
Mac mini M4 ProApple1224 GB38 TOPS3%Control plane
Mac mini M6Apple1224 GB—Not publishedDev server

Five boxes, each with one job, and none of them a spare.

Next

AI FlowHow every call finds its way across them.