Our take on the Gigabyte AI TOP ATOM AI Workstation

The Gigabyte AI TOP ATOM is a compact desktop workstation built on the NVIDIA DGX Spark GB10 platform. It pairs the GB10 Grace Blackwell Superchip with 128GB of unified LPDDR5x memory and a 4TB PCIe 5.0 NVMe SSD, so you can train, fine-tune and run inference on models locally. It handles models up to 200 billion parameters, and two linked units reach 405 billion through the dual 200 Gb/s QSFP112 ports. It suits developers, researchers and small teams who want local AI work without a rack of GPUs; against discrete-GPU workstations it trades raw throughput for a small footprint, unified memory and low power draw.

Best forDevelopers running local LLM inference and fine-tuningData scientists prototyping on a deskEdge AI and computer vision workSmall teams that want shared local AI capacity

Pros

128GB unified memory fits large models on one machine
Dual 200 Gb/s QSFP112 ports let two units pool capacity
4TB PCIe 5.0 NVMe storage is fast and roomy
Wi-Fi 7 and 10 Gb/s Ethernet cover modern networking
Compact, low-power desktop format with a 240W USB-C adapter
Ships with NVIDIA DGX OS and the Gigabyte AI TOP utility

Cons

ARM CPU and DGX OS limit some x86 software and drivers
No discrete GPU, so raw training throughput trails big GPU rigs
Scaling past 200B parameters needs a second unit
One-year limited warranty is short for a workstation

About the Gigabyte AI TOP ATOM AI Workstation

The Gigabyte AI TOP ATOM is a compact desktop AI workstation built on the NVIDIA DGX Spark GB10 platform. It uses the NVIDIA GB10 Grace Blackwell Superchip with a 20-core ARM CPU and fifth-generation Tensor Cores. It has 128 GB of unified LPDDR5x memory and a 4 TB NVMe M.2 PCIe 5.0 SSD.

It is meant for training, fine-tuning and running inference on models locally. It supports models of up to 200 billion parameters. Linking two systems through the 200 Gb/s ports on the NVIDIA ConnectX-7 SmartNIC scales support to 405 billion parameters.

Connections include Wi-Fi 7, a 10 Gb/s RJ45 Ethernet port, four USB-C 3.2 Gen 2x2 ports and an HDMI 2.1a output. It comes with NVIDIA DGX OS preinstalled and works with the Gigabyte AI TOP utility. It powers from a 240 W USB-C adapter.

Model

ATAGB10-9000

Product type

AI workstation

Platform

NVIDIA DGX Spark GB10

Processor

NVIDIA GB10 Grace Blackwell Superchip

CPU

20 cores: 10 Grace / Cortex-X925 up to 3.8 GHz, 10 Grace / Cortex-A725 up to 3.25 GHz

GPU

NVIDIA

Memory

128 GB LPDDR5x, onboard and not upgradeable

Storage

1 x 4 TB M.2 NVMe SSD

Storage expansion

1 x M.2 slot (PCIe 5.0), occupied

Operating system

NVIDIA DGX OS

Ports

4 x USB-C 3.2 Gen 2x2 (1 with Power Delivery, 3 with Video Alt Mode and Power Delivery), 1 x HDMI 2.1a output, 1 x RJ45 10 GbE

High-speed network

2 x QSFP 200 GbE (NVIDIA ConnectX-7 SmartNIC)

Wi-Fi

Wi-Fi 7 (802.11be), tri-band 2.4/5/6 GHz, MU-MIMO

Bluetooth

5.3

Security

Kensington lock slot

Power

240 W USB-C power adapter

Not included

No optical drive, external bays, speakers, microphones, memory card slot, keyboard or pointing device; no audio I/O

Dimensions

5.9 x 5.9 x 2 in

Weight

2.6 lb

Warranty

Limited 1-year manufacturer warranty

Questions and answers

Can it really run a 200B parameter model?

Yes, within the 128GB of unified memory, and two linked units reach 405B parameters over the QSFP112 ports. Actual headroom depends on quantization and context length.

How does it compare with a GPU workstation?

It trades raw throughput for a small, low-power desktop with large unified memory. A multi-GPU tower trains faster; this fits big models on one desk.

Does it run standard x86 software?

No. It uses a 20-core ARM CPU and NVIDIA DGX OS, so check that your stack and drivers support ARM Linux before buying.

What is in the box?

The workstation, a 240W USB-C power adapter and the preinstalled NVIDIA DGX OS. A display and peripherals are not included.

Can I link two units?

Yes. The two 200 Gb/s QSFP112 ports on the ConnectX-7 SmartNIC connect two systems to scale to 405B parameter models.

Is the memory upgradeable?

The 128GB LPDDR5x is unified with the GB10 Superchip, so treat it as fixed when planning workloads.

Where is the Gigabyte AI TOP ATOM AI Workstation cheapest?

At Linco, for $4,409.51, as of our latest check.

Gigabyte AI TOP ATOM AI Workstation

Best price now

$4,409.51at LincoChecked Nov 11

Prices are checked against each store and labelled with when.

Visit Linco

Ask our agent about this product

It checks today's price at every store, compares the alternatives and answers in about a minute.

Price facts

Last seen at $4,409.51 at Linco on Nov 11.