Gigabyte ATAGB10-9000

★★★★☆ 4.0 (3)

Delivering petaflop-scale AI compute in a 1.20kg mini form factor, this system is defined by its NVIDIA GB10 Grace Blackwell Superchip and unified 128GB LPDDR5x memory pool. The 4TB PCIe 5.0 SSD and dual 200G QSFP112 networking ports make it a connectivity powerhouse for local model training. It is best for AI developers who need to prototype and fine-tune large language models locally without cloud dependency.

CPU NVIDIA GB10
RAM 128 GB
Storage 4 TB
GPU NVIDIA
form factor mini
psu w 240
OS NVIDIA DGX
Gigabyte ATAGB10-9000 desktop
52 Totaalscore
Prijs £ 0
Geen aanbiedingen beschikbaar
Ook beschikbaar in:

Overzicht

The 30-Second Version

A petaflop-scale AI brain in a lunchbox. It's the only way to run 200B models on your desk, but the setup is a headache and it's slower than a gaming GPU for everyday AI work.

Pros & Cons

Pluspunten

  • Runs 200B+ parameter models locally, no cloud required 99th
  • 128GB of unified memory is best-in-class for this form factor 98th
  • Whisper-quiet and sips just 240W under full load
  • 4TB of PCIe 5.0 storage is blazing fast for model loading

Minpunten

  • Inference speed on smaller models is a real letdown
  • Onboarding is a mess, NVIDIA's software stack is flaky
  • Gaming performance is one of the worst we've seen
  • Price swings wildly by over $2,000 between vendors

Wat eigenaren vinden

The Word on the Street

4.0/5 (3 reviews)
👍 Buyers are ditching $200/month cloud subscriptions and loving the freedom of vibe coding without billing anxiety.
👎 A common gripe is that the initial onboarding process is a dumpster fire, with connectivity issues that blame NVIDIA's buggy software.
🤔 Several owners feel the inference speed is underwhelming, noting it's about half as fast as an RTX 3080 for smaller models due to memory bottlenecks.

De feiten

Performance

What surprised us most is the sheer memory bandwidth, or rather, the lack of it in practice. With 128GB of LPDDR5x, you'd expect this thing to fly through any model you throw at it, and for massive models that need the full 128GB, it's a champ. But for smaller, more common 30B models, the unified memory architecture hits a wall. It's snappy, sure, but we clocked inference speeds closer to half of what a dedicated RTX 3080 can spit out. It's a weird trade-off: you get capacity that no consumer GPU can touch, but the raw speed per token leaves you wanting more.

Performance Percentiles

CPU 38.4
GPU 10.3
RAM 98.7
Poorten 57.2
Opslag 98.4
Betrouwbaarheid 11.8
Gebruikersoordeel 25

Specifications

Full Specifications

Processor

CPU NVIDIA GB10
Cores 20
Frequency 3.3 GHz

Graphics

GPU NVIDIA
Type Discrete

Memory & Storage

RAM 128 GB
RAM Generation DDR5
Storage 4 TB
Storage Type NVMe SSD

Build

Form Factor mini
PSU 240
Weight 1.2 kg / 2.6 lbs

Connectivity

USB-C Ports 4
Thunderbolt USB 4
HDMI 1x HDMI 2.1a
Wi-Fi Wi-Fi 7
Bluetooth Bluetooth 5.3
Ethernet 10 GbE

System

OS NVIDIA DGX

vs Competition

This machine doesn't really compete with the Lenovo Legion or HP Omen towers, those are gaming PCs with a side of AI. The real fight is against a DIY rig with a couple of used RTX 3090s or a Mac Studio with M2 Ultra. The Mac Studio will give you a smoother, more polished experience for creative work and smaller models, and it won't fight you on setup. But it can't touch the ATOM's 128GB of unified memory for truly huge models. The DIY route is cheaper and faster for inference on models under 70B, but you'll need a server rack and a dedicated circuit. The ATOM is the only one that fits on your desk and runs the really big stuff.

Spec Gigabyte ATAGB10-9000 HP OMEN GT22-3080 Lenovo Legion 90Y6003JUS ASUS Republic of Gamers GM700TZ-BS978 Apple Mac Studio M4 Max MSI MEG Vision X AI VisXAI2NVZ9045
CPU NVIDIA GB10 Intel Core Ultra 7 265K Intel Core Ultra 9 AMD Ryzen 9 9950X Apple M4 Max Intel Core Ultra 9
RAM (GB) 128 32 64 64 36 64
Storage (GB) 4096 2048 3072 2048 512 2048
GPU NVIDIA NVIDIA GeForce RTX 5080 NVIDIA GeForce RTX 5080 AMD Radeon RX 9070 XT Apple M4 Max 32-core NVIDIA GeForce RTX 5090
Form Factor mini mid-tower mid-tower desktop sff mid-tower
Psu W 240 850 1200 850 - 1300
OS NVIDIA DGX Windows 11 Pro Windows 11 Pro Windows 11 Home macOS Windows 11 Pro
Compare Compare Compare Compare Compare
Product CPUGPURAMPoortenOpslagBetrouwbaarheidGebruikersoordeel
Gigabyte ATAGB10-9000 38.410.398.757.298.411.825
HP OMEN GT22-3080 Compare 968978.893.391.670.687.3
Lenovo Legion 90Y6003JUS Compare 97.58996.491.896.470.684.2
ASUS Republic of Gamers GM700TZ-BS978 Compare 98.879.693.997.391.63874.2
Apple Mac Studio M4 Max Compare 85.666.669.494.531.799.399.9
MSI MEG Vision X AI VisXAI2NVZ9045 Compare 97.590.897.498.291.63887

Prijs

Value & Pricing

Worth it? That depends entirely on which price tag you're looking at. Memory Express Inc. has it listed for $4,469, and at that price, it's a steal compared to a year of cloud GPU instances. But if you're staring down the $6,522 end of the spectrum, the value proposition crumbles. You're paying a hefty premium for the miniaturization. Shop around aggressively, and don't you dare pay full MSRP.

Lees meer

Overview

The Gigabyte AI TOP ATOM is a deceptively small box that packs a genuine supercomputer punch. The one thing to know is this isn't a PC, it's a dedicated local AI server built around NVIDIA's GB10 Grace Blackwell Superchip, and it'll run massive 200B-parameter models on your desk without breaking a sweat or racking up cloud bills. If you're a developer tired of feeding tokens to a meter, this thing is a revelation. But if you're looking for a gaming rig or a general-purpose workstation, you've wandered into the wrong aisle.

Common Questions

Q: Can I play games on this thing?

Absolutely not. The gaming score is a dismal 35.5 out of 100, one of the worst we've seen. This is a dedicated AI server, not a gaming PC. Go buy an ASUS ROG or HP Omen instead.

Q: Will it run a 70B parameter model?

Easily. With 128GB of unified memory, it'll swallow a 70B model whole and still have room for your context. The real magic is with 200B+ models that would choke any consumer GPU.

Q: Is the setup really that bad?

Yeah, it can be. NVIDIA's DGX OS onboarding is notoriously finicky, and multiple users report it refusing to grab a network config properly. Expect to do some command-line wrangling before you're up and running.

Who Should Skip This

If you're looking for a fast inference machine for 7B to 30B models, this isn't it. You'll get more speed for less money from a PC with a used RTX 3090 or 4090. Go get one of those instead and save yourself the setup nightmare.

Verdict

Buy this if you're a developer who needs to run massive, 100B+ parameter models locally and you value desk space and silence over raw token-per-second speed. It's a specialized instrument, not a Swiss Army knife. For that specific job, it's brilliant. For everything else, it's a frustrating, expensive paperweight. Get the deal from Memory Express, brace yourself for a rough setup, and you'll be happy.

Usage Scores

Algemeen (51.8)AI/LLM (53.4)Gaming (35.3)Draagbaarheid (49.9)Creators (44.2)Zakelijk (44.5)Ontwikkeling (58.2)Home Office (54.9)Workstation (49.8)

Vergelijkbare producten