Extended review
The full write-up on this product
Gigabyte AI TOP ATOM is a compact personal AI desktop — 150 × 150 × 50.5 mm, roughly one liter in volume — powered by the NVIDIA GB10 Grace Blackwell chip. This model, ATAGB10-9001, comes with 128GB of unified memory and a 4TB drive, designed for local AI development and inference on your desktop.
Memory128GB LPDDR5x
Storage4TB NVMe
Networking10GbE + ConnectX-7
Inside the Chip
The GB10 combines a 20-core Arm CPU (10 × Cortex-X925 and 10 × Cortex-A725) with a Blackwell architecture GPU and 5th-generation Tensor Cores. According to NVIDIA, the GB10 platform features 6,144 CUDA cores. NVLink-C2C handles the CPU-to-GPU interconnect directly on the chip.
| Processor | 20 Arm cores: 10 × Cortex-X925 and 10 × Cortex-A725 |
|---|
| Graphics | Blackwell with 5th-gen Tensor Cores, 6,144 CUDA cores (GB10 platform specification) |
|---|
| AI Performance | Up to 1 petaflop in FP4 — theoretical figure relying on sparsity |
|---|
| Memory | 128GB LPDDR5x unified, shared between CPU and GPU, 256-bit interface, 273GB/s. Part of the memory is reserved for the system |
|---|
| Storage | 4TB M.2 NVMe, PCIe 4.0 |
|---|
| Wired Networking | RJ-45 at 10GbE standard |
|---|
| High-Speed Networking | NVIDIA ConnectX-7 with two QSFP ports, up to 200Gbps per port |
|---|
| Wireless | Wi-Fi 7, Bluetooth 5.4 |
|---|
| USB | 4 × USB Type-C 3.2 Gen 2x2 (up to 20Gbps). One serves as the power input, and three are available for devices |
|---|
| Video | HDMI 2.1a with multi-channel audio. The USB-C ports support DisplayPort Alt Mode |
|---|
| Dimensions and Weight | 150 × 150 × 50.5 mm, 1.2 kg |
|---|
| Power | External 240W adapter included. The GB10 chip TDP is 140W — a chip specification, not total system consumption |
|---|
| Software | NVIDIA DGX OS (Ubuntu Linux-based) with NVIDIA AI software stack, plus Gigabyte AI TOP Utility |
|---|
| Physical Security | Kensington Nano slot |
|---|
🔗ConnectX-7 with QSFPTwo QSFP ports, up to 200Gbps per port. Two units can be linked using a single QSFP cable, sold separately
📦Large ModelsAccording to Gigabyte, a single unit supports quantized models up to 200 billion parameters, and two linked units support up to 405 billion
🖥️DGX OSPre-installed NVIDIA Linux OS with the AI software stack
Things to Keep in Mind- The 1 petaflop figure is a theoretical FP4 value with sparsity — not a general performance benchmark
- 200Gbps is the port rating, not measured throughput
- Part of the memory is reserved for the system, so available memory is less than 128GB
- The processor is Arm and the pre-installed OS is Linux (DGX OS) — not a standard x86 PC
