If your organisation is starting to use AI tools for writing, analysis, customer service, research or internal workflows you are probably doing it through a cloud service right now. That works, but it comes with trade-offs: ongoing subscription or usage costs, reliance on cloud infrastructure and, depending on the service and configuration, considerations around how and where your data is processed. Local AI hardware offers another option: running AI models on-site, giving your organisation greater control over the infrastructure and data involved.

The GIGABYTE AI TOP ATOM is one of the first compact machines designed specifically for this. It is built on NVIDIA's DGX Spark platform and is small enough to sit on a desk.

Independent reviews of the AI TOP ATOM have been published by ServeTheHome and StorageReview. This article draws on those reviews alongside GIGABYTE's published specifications to explain what the system is, how it performs and who it is likely to suit. TechVerse carries the AI TOP ATOM in multiple storage configurations.

Specifications

The following specifications are taken from GIGABYTE's published product documentation.

Processor	NVIDIA GB10 Grace Blackwell Superchip
CPU	20 Arm CPU cores
GPU	NVIDIA Blackwell GPU (integrated in GB10)
Memory	128 GB LPDDR5X unified (shared CPU and GPU)
AI compute	1 PFLOP FP4 (per NVIDIA GB10 specification)
Networking	NVIDIA ConnectX-7 200 GbE
Power supply	240 W USB-C PD (external adapter included)
Dimensions	150 x 150 x 51 mm
Weight	Approximately 1.2 kg
Operating system	Ubuntu (DGX OS base)

 

TechVerse currently carries the AI TOP ATOM in the following storage configurations:

• GIGABYTE AI TOP ATOM with 4 TB Gen5 NVMe
• GIGABYTE AI TOP ATOM with Gen4 NVMe (contact us for current options)

Storage capacity affects how many model weight files can be held locally. For teams working with multiple large models, the 4 TB variant avoids the need to reload weights from a NAS or external source between sessions.

What the GB10 unified memory architecture means in practice

The GB10's 128 GB LPDDR5X memory is shared between the CPU and the GPU. In a conventional discrete GPU setup, the CPU has its own system RAM and the GPU has separate VRAM; moving data between them involves a PCIe transfer that adds latency. The unified approach removes that step.

For AI inference, the practical result is that the GPU can access all 128 GB without a separate VRAM ceiling. NVIDIA states the GB10 platform supports models of up to approximately 200 billion parameters in suitable quantisation formats. Actual model fit depends on quantisation level, architecture, context window size and runtime overhead, so this figure should be treated as a general indication rather than a hard limit.

Design and cooling

The AI TOP ATOM measures 150 x 150 x 51 mm and weighs approximately 1.2 kg. Its footprint is similar to a portable hard drive enclosure, which makes it practical for desk deployment, lab benches or situations where space is constrained. Power is supplied through a 240 W USB-C PD external adapter, which is included with the system.

The chassis uses a combination of passive external surfaces and an internal fan array. Under sustained workloads, the fans increase speed noticeably. The system is not silent under full load, which is worth considering for open-plan office environments.

Thermal performance

ServeTheHome and StorageReview both conducted sustained workload testing. The following temperatures were recorded in their testing:

• CPU: peak approximately 90 degrees C (ServeTheHome)
• GPU: peak approximately 81 degrees C during the Prefill phase of inference (ServeTheHome)
• NVMe (Samsung PM9E1): remained below 60 degrees C throughout sustained testing (StorageReview)
• ConnectX-7 NIC: peak approximately 72 degrees C (StorageReview)

All of these are within the system's operating envelope. ServeTheHome and StorageReview note the system performed consistently throughout their test runs. Ensure the unit has adequate airflow clearance on all vented sides in deployment.

Storage: GPU Direct Storage support

The AI TOP ATOM includes a Samsung PM9E1 NVMe SSD. StorageReview's testing measured GPU Direct Storage (GDS) read performance, which allows the GPU to access NVMe data directly over PCIe without routing through the CPU. In their testing, GDS sequential read throughput reached approximately 11 GiB/s at 8 parallel threads with 1 MB block sizes, at which point throughput plateaued, indicating the drive rather than the GDS interface was the limiting factor.

GDS is most relevant at model load time, where weights need to be transferred from storage into memory. StorageReview's published results provide the relevant throughput figures for different thread counts and block sizes.

Inference performance

StorageReview benchmarked inference throughput using vLLM Online Serving, a standard framework for evaluating LLM server performance. They tested across multiple models at different sizes and quantisation levels. Their results, published on StorageReview.com, show throughput in tokens per second for each configuration.

GPU power draw during inference peaked at approximately 75.54 W in StorageReview's testing. Combined with CPU and platform overhead, total system power remained within the 240 W supply headroom throughout their tests. Full model-by-model results and methodology are available in StorageReview's published article.

Who is this suited for?

The AI TOP ATOM is primarily designed for local AI development and inference rather than large-scale model training from scratch. The combination of 128 GB unified memory, a compact form factor and a USB-C power connection makes it suitable for:

• AI developers and researchers who want capable local inference without the space and power requirements of a conventional GPU server
• Organisations that process sensitive data locally, where data residency policies or operational preferences make cloud API inference unsuitable. This covers a range of situations, from internal preference to formal compliance requirements, and the right approach depends on the organisation's specific controls and the data involved
• Professional and technical users who need local AI compute at client sites or in environments where reliable network connectivity cannot be guaranteed

It is not a replacement for a GPU cluster where high concurrent throughput is the primary requirement, and it is not designed for large-scale model training. Within its intended scope, the combination of the GB10 unified memory and the form factor is distinctive at this class of device.

TechVerse: AI TOP ATOM availability

TechVerse carries the GIGABYTE AI TOP ATOM in both 4 TB Gen5 NVMe and Gen4 NVMe configurations. If you are evaluating which storage variant suits your workload, or considering a multi-system deployment (two units can be connected directly via a DAC cable for a 256 GB combined-memory configuration), contact our team to discuss requirements and current availability. Buy now