GIGABYTE AI TOP ATOM Personal AI Supercomputer NVIDIA GB10 - 128GB / 4TB PCIe 4.0
PRODUCT DESCRIPTION
GIGABYTE AI TOP ATOM Personal AI Supercomputer, NVIDIA GB10 Grace Blackwell Superchip, 20-Core Arm CPU with Blackwell GPU, 128GB Unified Memory, 4TB M.2 NVMe PCIe 4.0 SSD, ConnectX-7 200GbE, 10GbE, Wi-Fi 7 (ATAGB10-9001) - 3 Year Local Warranty
GIGABYTE AI TOP ATOM 4TB PCIe 4.0 β NVIDIA GB10 Personal AI Supercomputer in a 1-Litre Desktop
The GIGABYTE AI TOP ATOM (ATAGB10-9001) is GIGABYTE's build of the NVIDIA DGX Spark platform: a personal AI supercomputer powered by the NVIDIA GB10 Grace Blackwell Superchip, pairing a 20-core Arm CPU with an integrated NVIDIA Blackwell GPU and 128 GB of coherent unified LPDDR5x memory inside a 1-litre, 150 Γ 150 mm chassis. GIGABYTE rates it at 1 petaFLOP of FP4 AI performance and states support for models of up to 200 billion parameters on one unit, or 405 billion with two units linked over ConnectX-7. This configuration fits a 4 TB M.2 NVMe PCIe 4.0 SSD.
For Singapore developers, data scientists, university labs and enterprise AI teams, the ATOM puts DGX-class local inference, fine-tuning, RAG and agentic AI on a desk, with no cloud GPU bill and no data leaving the premises. GIGABYTE adds its AI TOP Utility on top of NVIDIA DGX OS, and SourceIT supplies it with a full GST invoice and 3-year local warranty.
NVIDIA GB10 Grace Blackwell Superchip β 1 PetaFLOP FP4, 20 Arm Cores, 5th-Gen Tensor Cores
The GB10 Superchip combines a 20-core Arm CPU β 10 Cortex-X925 performance cores and 10 Cortex-A725 efficiency cores β with an NVIDIA Blackwell architecture GPU on one package, sharing a single coherent memory space rather than copying data across a PCIe bus. Fifth-generation Tensor Cores with native FP4 precision deliver up to 1 PFLOP (1,000 TOPS) of sparse AI throughput, fourth-generation RT Cores accelerate rendering and synthetic-data generation, and one NVENC and one NVDEC engine handle video. GIGABYTE's figures match NVIDIA's DGX Spark platform specification, so every DGX Spark Playbook and container applies directly to the ATOM.
128 GB Unified Memory at 273 GB/s β Run 200B-Parameter Models Locally
A discrete workstation GPU with 16 to 32 GB of VRAM stops long before a 70-billion-parameter model loads. The ATOM's 128 GB of LPDDR5x is one coherent pool on a 256-bit interface at 273 GB/s, shared by CPU and GPU, so Llama, Qwen, DeepSeek, Gemma, Mistral, gpt-oss and Nemotron-class models of up to 200 billion parameters load in full on one unit, and NVIDIA's platform figures allow fine-tuning of models up to 70 billion parameters. GIGABYTE notes that a portion of memory is reserved for system use. GIGABYTE also lists 64 GB ATOM variants; every tier SourceIT stocks is the 128 GB build.
4TB M.2 NVMe PCIe 4.0 SSD β What the PCIe Generation Does and Does Not Change
This configuration (ATAGB10-9001) carries the 4 TB M.2 NVMe SSD on a PCIe 4.0 interface β the same capacity as the 4TB PCIe 5.0 tier (ATAGB10-9000) on the previous-generation interface. A PCIe 4.0 x4 link provides roughly 8 GB/s of theoretical bandwidth against roughly 16 GB/s for PCIe 5.0 x4, so the PCIe 5.0 tier pulls a multi-hundred-gigabyte checkpoint off the drive faster; GIGABYTE does not publish the SSD model or its read and write ratings. The interface generation only affects that load step: once a model is resident in the 128 GB unified memory, inference and fine-tuning run at the same speed on every ATOM tier because the GB10 and its 273 GB/s memory bandwidth set the pace. For teams that load a model once and run it for hours, this tier gives the full 4 TB of space on one unit; the 1TB PCIe 4.0 tier (ATAGB10-9002) is the entry point where a smaller library of quantised models is enough.
ConnectX-7 200 GbE, 10 GbE, Wi-Fi 7 and Four USB-C Ports
Networking is what separates the ATOM from an ordinary mini PC. An integrated NVIDIA ConnectX-7 NIC provides two QSFP ports at 200 Gbps for linking a second ATOM or joining a switched cluster, while a 10 GbE RJ-45 port handles the office network and model downloads; Wi-Fi 7 and Bluetooth 5.4 cover wireless. On the rear panel GIGABYTE lists three USB 3.2 Gen 2x2 Type-C ports at up to 20 Gbps, a fourth USB 3.2 Gen 2x2 Type-C port that serves as the power input (PD in) and one HDMI 2.1a output with multichannel audio; a Kensington Nano lock slot sits on the side. GIGABYTE does not publish DisplayPort Alt Mode on the USB-C ports, so plan on HDMI 2.1a as the display output.
1-Litre Chassis, 240 W Adapter, 1.2 kg β Cool and Quiet on the Desk
The ATOM measures 150 Γ 150 Γ 50.5 mm β a 1-litre chassis that fits under a monitor, on a lab bench or in an edge cabinet β and weighs 1.2 kg per GIGABYTE's specifications page. It draws power from a 240 W external adapter through the USB-C PD input, so it runs from a standard socket. GIGABYTE describes the thermal design as keeping the system cool and quiet during intensive workloads but publishes no fan count or acoustic figures.
NVIDIA DGX OS and GIGABYTE AI TOP Utility
The ATOM boots into NVIDIA DGX OS, an Ubuntu Linux-based build with the full NVIDIA AI software stack preinstalled β CUDA, TensorRT, NVIDIA NIM microservices, PyTorch and NGC containers β the same foundation as DGX Spark, DGX Station and DGX Cloud, so GIGABYTE states that work started on the ATOM migrates to larger NVIDIA infrastructure. GIGABYTE layers its AI TOP Utility on top, listing model download, inference, retrieval-augmented generation (RAG) and machine learning as its current capabilities, with fine-tuning on your own data among the applications it describes. NVIDIA's platform software also applies: NemoClaw wraps OpenClaw agents with OpenShell policy-based privacy and security guardrails, NVIDIA AI Enterprise for DGX Spark is available for production development, DGX Spark Playbooks provide step-by-step recipes, and GIGABYTE names the NVIDIA Agent Toolkit and NVIDIA Sync for multi-unit management.
Scaling From One ATOM to Two or Four
GIGABYTE states that connecting two ATOM systems scales to models of up to 405 billion parameters over ConnectX-7, and its page adds that users can cluster up to four DGX Spark-class systems with the built-in ConnectX-7 networking and NVIDIA Sync; NVIDIA's platform figures put four linked GB10 systems at up to 700 billion parameters. A two-unit link needs only a QSFP direct-attach cable; four units need an external 200 GbE switch. Beyond 128 GB, the same DGX OS stack carries over to the GIGABYTE W775-V10-L01 DGX Station GB300 system with 748 GB of coherent memory.
Available in Singapore Through SourceIT
SourceIT is an authorised IT reseller in Singapore and supplies the GIGABYTE AI TOP ATOM 4TB PCIe 4.0 (ATAGB10-9001) with a full GST invoice, 3-year local warranty and enterprise, research-grant and education procurement assistance. The ATOM is also available as the 4TB PCIe 5.0 (ATAGB10-9000) and 1TB PCIe 4.0 (ATAGB10-9002) on the same GB10 platform. Browse every GB10 and GB300 system we stock in the NVIDIA DGX Spark & DGX Station collection and the wider AI workstations and AI servers collection.
GIGABYTE AI TOP ATOM 4TB PCIe 4.0 (ATAGB10-9001) Datasheet β Technical Specifications
General
| Product Name | GIGABYTE AI TOP ATOM Personal AI Supercomputer β 4TB PCIe 4.0 |
| SKU / Part Number | ATAGB10-9001 |
| Category | Personal AI supercomputer / AI developer desktop (NVIDIA DGX Spark GB10 platform) |
| Form Factor | 1-litre compact desktop, 150 Γ 150 mm footprint |
| Local Warranty (Singapore) | 3 Year Local Warranty; procurement and GST invoice by SourceIT |
Processor & GPU β NVIDIA GB10 Grace Blackwell Superchip
| Chipset | NVIDIA GB10 Grace Blackwell Superchip |
| CPU | 20-core Arm: 10 Γ Cortex-X925 performance cores + 10 Γ Cortex-A725 efficiency cores |
| GPU | NVIDIA Blackwell architecture GPU (integrated) |
| Tensor Cores | 5th generation, FP4 / FP8 (NVIDIA platform specification) |
| RT Cores | 4th generation (NVIDIA platform specification) |
| Video Engines | 1 Γ NVENC, 1 Γ NVDEC |
AI Performance
| Tensor Performance | 1 PFLOP (1,000 TOPS) FP4 with sparsity (theoretical, per NVIDIA footnote) |
| Model Support (single unit) | AI models up to 200 billion parameters, per GIGABYTE; fine-tuning up to 70 billion parameters, per NVIDIA platform figures |
| Model Support (two units) | Up to 405 billion parameters with two ATOM systems connected over NVIDIA ConnectX-7, per GIGABYTE |
| Model Support (four units) | Up to 700 billion parameters with four GB10 systems, per NVIDIA |
Memory
| System Memory | 128 GB LPDDR5x coherent unified system memory, shared by CPU and GPU |
| Memory Interface | 256-bit |
| Memory Bandwidth | 273 GB/s |
| Upgradeability | Soldered β not user-upgradeable |
Storage
| SSD | 4 TB M.2 NVMe PCIe 4.0 SSD |
| Encryption | Self-encrypting NVMe M.2 drive, per GIGABYTE datasheet |
| Interface Bandwidth | PCIe 4.0 x4 β approximately 8 GB/s theoretical link bandwidth; SSD model and sequential read/write ratings not published by GIGABYTE |
Networking
| High-Speed NIC | NVIDIA ConnectX-7 NIC β 2 Γ QSFP, 200 Gbps |
| Ethernet | 1 Γ RJ-45, 10 GbE |
| Wi-Fi | Wi-Fi 7 |
| Bluetooth | Bluetooth 5.4 |
I/O & Display
| USB (data) | 3 Γ USB 3.2 Gen 2x2 Type-C, up to 20 Gbps (rear) |
| USB (power in) | 1 Γ USB 3.2 Gen 2x2 Type-C, PD in (rear) |
| Display Output | 1 Γ HDMI 2.1a (rear); DisplayPort Alt Mode over USB-C not published by GIGABYTE |
| Audio Output | HDMI multichannel audio |
| Security | 1 Γ Kensington Nano slot (side) |
Software & Security
| Operating System | NVIDIA DGX OS (Ubuntu Linux-based), arm64 |
| AI Stack (preinstalled) | Full NVIDIA AI software stack β CUDA, TensorRT, NVIDIA NIM microservices, PyTorch, NGC containers |
| GIGABYTE Software | GIGABYTE AI TOP Utility β model download, inference, retrieval-augmented generation (RAG), machine learning |
| Agentic AI | NVIDIA NemoClaw with OpenShell guardrails; NVIDIA Agent Toolkit; NVIDIA Sync |
Power, Physical, Acoustics & Certifications
| Power Supply | 240 W external adapter, USB-C PD input |
| Chassis Volume | 1 L |
| Dimensions (H Γ W Γ D) | 50.5 Γ 150 Γ 150 mm |
| Weight | 1.2 kg (2.64 lb) per GIGABYTE's specifications page; GIGABYTE's datasheet PDF lists 1.5 kg |
| Cooling / Acoustics | Fan configuration and noise figures not published by GIGABYTE |
| Certifications | Not published by GIGABYTE |
| In the Box | AI TOP ATOM unit, power adapter, power cord (varies by country) |
Compatibility & Software Support
What Runs on the AI TOP ATOM
Anything in the NVIDIA and open-source AI ecosystem published for arm64: Hugging Face Transformers, vLLM, Ollama, llama.cpp, LangChain and LlamaIndex RAG pipelines, NVIDIA NIM microservices and Blueprints, NeMo, TensorRT-LLM, ComfyUI, RAPIDS and the NVIDIA Isaac, Metropolis and Holoscan frameworks. NVIDIA's DGX Spark Playbooks apply directly because it is the same GB10 platform, and GIGABYTE's AI TOP Utility adds a guided path for users who prefer not to start from the command line.
Arm64 Linux β Plan Your Software Accordingly
The GB10's CPU is Arm-based and the ATOM ships with NVIDIA DGX OS, so drivers and containers are arm64 (aarch64) builds. Every major AI framework and NGC container is published for arm64, but x86-only proprietary software will not run natively and Windows is not supported. Most teams drive the ATOM from a laptop over SSH, Jupyter or VS Code Remote. For a Windows 11 Arm desktop in the same collection, see the ASUS Ascent QN10.
Linking Two or Four ATOM Units
Two ATOM units connect back-to-back over their ConnectX-7 QSFP ports with a single QSFP112 direct-attach cable β SourceIT stocks the NVIDIA QSFP112 400G DAC cable, 40 cm (X0101G00400A) β for models of up to 405 billion parameters per GIGABYTE. Four units need an external 200 GbE switch and NVIDIA Sync, reaching 700 billion parameters per NVIDIA. The 10 GbE RJ-45 port stays free for site networking.
Where the ATOM Sits in the GB10 Family
The AI TOP ATOM is one of several GB10 desktops SourceIT stocks alongside the NVIDIA DGX Spark itself, from ASUS, Dell, Lenovo, HP, MSI, Acer and xFusion. All share the GB10 Superchip, 128 GB unified memory, 1 PFLOP FP4 and ConnectX-7 networking; the ATOM is distinguished by GIGABYTE's AI TOP Utility, a three-tier storage range with a PCIe 5.0 option at 4 TB, and the 3-year local warranty supplied through SourceIT.
Package Contents & Ordering Information
What Ships in the Box
- GIGABYTE AI TOP ATOM (ATAGB10-9001) with NVIDIA GB10 Grace Blackwell Superchip, 128 GB unified memory and 4 TB M.2 NVMe PCIe 4.0 SSD
- 240 W power adapter and power cord (varies by country)
- NVIDIA DGX OS preinstalled with the NVIDIA AI software stack and GIGABYTE AI TOP Utility
Monitor, keyboard, mouse and the QSFP112 clustering cable are not included.
GIGABYTE AI TOP ATOM Range at SourceIT Singapore
- GIGABYTE AI TOP ATOM 4TB PCIe 5.0 NVMe β ATAGB10-9000
- GIGABYTE AI TOP ATOM 4TB PCIe 4.0 NVMe β ATAGB10-9001 (this listing)
- GIGABYTE AI TOP ATOM 1TB PCIe 4.0 NVMe β ATAGB10-9002
- NVIDIA QSFP112 400G DAC Cable, 40 cm β X0101G00400A (links two units)
All three tiers share the same GB10 Superchip, 128 GB unified memory, ConnectX-7 networking, ports and chassis; only SSD capacity and PCIe generation differ. The GIGABYTE AI TOP ATOM 4TB PCIe 4.0 is in stock at SourceIT Singapore with GST invoice and 3-year local warranty.
Other NVIDIA GB10 and DGX Station Systems at SourceIT
- NVIDIA DGX Spark 4TB β NVIDIA's own first-party GB10 unit
- ASUS Ascent GX10 β 1TB / 2TB / 4TB
- Dell Pro Max with GB10 β 2TB / 4TB
- Lenovo ThinkStation PGX β 1TB / 4TB
- HP ZGX Nano G1n β 4TB
- MSI EdgeXpert MS-C931 β 4TB
- Acer Altos BrainSphere GB10 F1 β 4TB
- xFusion FusionXpark GB10 β 1TB / 2TB / 4TB
- GIGABYTE W775-V10-L01 (NVIDIA DGX Station GB300) β step up to 748 GB coherent memory and 20 PFLOPS FP4 for trillion-parameter models
Browse the full NVIDIA DGX Spark & DGX Station collection. For volume, project, research-grant or education pricing, GST invoicing and current lead times, email sales@sourceit.com.sg quoting ATAGB10-9001.
Datasheet, Downloads & Setup Guide
Official GIGABYTE and NVIDIA documentation:
- GIGABYTE AI TOP ATOM β product page (Singapore)
- GIGABYTE AI TOP ATOM β full specifications
- GIGABYTE AI TOP ATOM datasheet (PDF)
- GIGABYTE AI TOP ATOM β support and downloads
- NVIDIA DGX Spark (GB10) β official platform page (Singapore)
- NVIDIA DGX Spark / DGX OS user documentation
- NVIDIA DGX Spark Playbooks β recipes for LLMs, fine-tuning, RAG and agents
- NVIDIA DGX Spark Setup and Enterprise Deployment Guide β SourceIT (applies to the AI TOP ATOM)
For the GIGABYTE user manual, email sales@sourceit.com.sg quoting ATAGB10-9001.
Frequently Asked Questions β GIGABYTE AI TOP ATOM 4TB PCIe 4.0
What is the GIGABYTE AI TOP ATOM?
GIGABYTE's personal AI supercomputer built on the NVIDIA DGX Spark GB10 platform: a 1-litre desktop with the NVIDIA GB10 Grace Blackwell Superchip, 128 GB of unified memory and 1 PFLOP of FP4 AI performance, running NVIDIA DGX OS with GIGABYTE's AI TOP Utility for local LLM inference, fine-tuning, RAG and agentic AI. This configuration (ATAGB10-9001) has a 4 TB M.2 NVMe PCIe 4.0 SSD.
What is the difference between the AI TOP ATOM and the NVIDIA DGX Spark?
Same NVIDIA GB10 Superchip, 128 GB unified memory, 1 PFLOP FP4, ConnectX-7 networking and 240 W adapter, so AI performance is identical. The ATOM is GIGABYTE's build with the AI TOP Utility preinstalled and three storage tiers; the NVIDIA DGX Spark is NVIDIA's first-party unit in 4TB only.
What is the difference between the three ATOM part numbers?
Only the SSD. ATAGB10-9001 (this listing) has 4 TB on PCIe 4.0, ATAGB10-9000 has the same 4 TB on PCIe 5.0 and ATAGB10-9002 has 1 TB on PCIe 4.0. Compute, memory, networking, ports, chassis and software are identical, so AI throughput is the same on all three; this tier gives the full 4 TB of space with a drive on the previous-generation interface.
Does PCIe 5.0 storage make the ATOM faster at AI?
Only at load time. PCIe 5.0 x4 doubles the theoretical link bandwidth of PCIe 4.0 x4 β roughly 16 GB/s against 8 GB/s β which shortens the time to read a large model from the drive into memory, but once the model is resident, inference and fine-tuning are bound by the GB10 and its 273 GB/s memory bandwidth. GIGABYTE does not publish the SSD model or its read and write ratings. This 4 TB PCIe 4.0 tier gives the same AI performance and the same capacity as the PCIe 5.0 tier, with a slower initial load from disk.
What AI models can the ATOM run?
GIGABYTE states support for models up to 200 billion parameters on one unit and 405 billion with two units linked; NVIDIA's platform figures add fine-tuning up to 70 billion parameters and up to 700 billion parameters across four units. In practice that covers Llama, Qwen, DeepSeek, Gemma, Mistral, gpt-oss and Nemotron-class models.
Can I link two ATOM units together?
Yes, back-to-back over the ConnectX-7 QSFP ports with the NVIDIA QSFP112 400G DAC cable, 40 cm, sold separately. Four units need an external 200 GbE switch and NVIDIA Sync.
What does the GIGABYTE AI TOP Utility do?
GIGABYTE's guided front end on top of the NVIDIA AI software stack. GIGABYTE lists model download, inference, retrieval-augmented generation (RAG) and machine learning as its current capabilities, and describes fine-tuning pre-trained models on your own data among the system's applications.
Does the ATOM run Windows or x86 software?
No. It ships with NVIDIA DGX OS (Ubuntu Linux-based, arm64) and is a Linux AI development system; proprietary x86-only desktop applications will not run natively.
Does it need special power or cooling?
No. It runs from a standard socket through the supplied 240 W adapter via the USB-C PD input, and GIGABYTE designs the 1-litre chassis to stay cool and quiet under sustained workloads. GIGABYTE does not publish noise figures.
Can I upgrade the memory or storage?
Memory is soldered and fixed at 128 GB. The SSD is an M.2 NVMe drive, but GIGABYTE publishes no user-upgrade procedure, so choose the tier you need at purchase and confirm any later drive change with SourceIT to keep the warranty intact.
Who is the GIGABYTE AI TOP ATOM for?
AI developers prototyping and fine-tuning before scaling to DGX Station or DGX Cloud; university labs and research groups; data scientists; robotics and vision teams; and Singapore enterprises in finance, healthcare, legal and government that must keep models and data on-premise.
What warranty does the ATOM come with in Singapore?
A 3-year local warranty in Singapore, supplied through SourceIT. Every unit comes with a GST invoice and local support.
How do I buy the GIGABYTE AI TOP ATOM in Singapore?
Order on this page or email sales@sourceit.com.sg quoting ATAGB10-9001. SourceIT is an authorised IT reseller in Singapore and supplies the ATOM with GST invoice, local warranty and the QSFP112 clustering cable.
