Gigabyte Launches Custom NVIDIA “DGX Spark” Mini-AI Supercomputer: The AI TOP ATOM Delivering 1,000 TOPS of AI Performance

Gigabyte, renowned for its innovative PC products, is making waves in the artificial intelligence sector with the launch of its distinct version of the NVIDIA DGX Spark supercomputer, dubbed the AI TOP ATOM.

Introducing Gigabyte’s AI TOP ATOM: A Compact Powerhouse in AI Computing

At this year’s Computex, NVIDIA unveiled an expansion of their DGX Spark offerings, moving from proprietary systems to allowing AIB partners to create customized versions. Gigabyte has seized this opportunity, presenting its unique supercomputer solution. We had the privilege of experiencing, firsthand, what may be one of the smallest supercomputers globally, which incorporates advanced components, including the ARM-based GB10 Grace Blackwell Superchip and memory from Micron. While details about the internal architecture remain sparse, let’s delve into what makes Gigabyte’s offering remarkable.

The AI TOP ATOM constitutes Gigabyte’s proprietary approach to the DGX Spark lineup. This compact device is designed to achieve an impressive AI performance of up to 1, 000 AI TOPS, significantly outperforming existing NPU engines by a factor of 20 to 50 times. Notably, the GB10 Grace Blackwell Superchip features the advanced NVIDIA Blackwell GPU, complete with fifth-generation Tensor Cores and FP4 support, enabling it to handle up to 1, 000 trillion operations per second, perfect for AI fine-tuning and inference tasks.

The innovative GB10 Superchip utilizes NVIDIA’s NVLink-C2C interconnect technology to facilitate a CPU+GPU-coherent memory model, yielding five times the bandwidth compared to fifth-generation PCIe systems. Additionally, the AI TOP ATOM incorporates Micron’s cutting-edge LPDDR5X memory chips. During our exclusive look inside, we noted that the GB10 Superchip was labeled as an engineering sample, hinting that the DGX Spark is still undergoing refinement. Gigabyte claims that the AI TOP ATOM can manage up to 70 billion parameters, and when paired with an additional ATOM, utilizing NVIDIA Connect-X, the system could reach up to 405 billion parameters for advanced AI training.

The system features various ports, including HDMI, Type-C, and ConnectX-T, but its aesthetics are relatively understated, likely not appealing to gaming enthusiasts seeking flashy designs. Nevertheless, the introduction of Gigabyte’s ATOM system marks a significant step forward in AI computation, embodying the future possibilities of AI hardware capabilities.

Source & Images