The ASUS Ascent GX10 is a compact, high-performance AI workstation that redefines what’s possible in a 150 x 150 mm footprint. Built around NVIDIA’s GB10 Superchip—an Arm v9.2-A CPU paired with Blackwell integrated graphics and 128GB of unified LPDDR5x memory—the system targets AI workloads with up to 1 petaFLOP of FP4 compute, fine-tuning models up to 200 billion parameters. Unlike bulkier workstations, the GX10 prioritizes efficiency without sacrificing performance, making it a standout for small businesses where space and heat are critical constraints.
ASUS has taken a minimalist approach to design, housing the GB10 hardware in a 51mm-tall, 1.48kg chassis finished in Stellar Grey. The front features vertical vents that span the entire width, with a subtle power button integrated into the grille—a detail that sets it apart from competitors who often omit power indicators entirely. All connectivity is rear-mounted, including three USB 3.2 Gen 2×2 Type-C ports (with DisplayPort 2.1 alternate mode), one USB Type-C port for 180W PD input, HDMI 2.1, and a dual-port NVIDIA ConnectX-7 SmartNIC for high-speed networking. A Kensington lock slot rounds out the rear I/O.
Storage flexibility is a key differentiator. The GX10 supports M.2 NVMe configurations ranging from 1TB PCIe 4.0 to 4TB PCIe 5.0, catering to workloads that demand either capacity or speed. Power comes from a 240W external USB-C adapter, though the system itself draws up to 180W—a balance that reflects ASUS’s focus on thermal efficiency without sacrificing performance.
Performance: Balanced Burst and Sustained Stability
Thermal testing reveals the GX10’s strength in managing heat under sustained workloads. CPU temperatures peaked at 87.3°C during burst phases but stabilized quickly, avoiding prolonged saturation. GPU thermals followed a similar pattern, reaching 82°C before leveling off. The NVMe drive remained cool (peaking at 56.8°C), while the NIC stayed within expected ranges (73°C max). This behavior suggests ASUS has optimized airflow to prevent throttling, even when pushing the system near its limits.
Performance benchmarks using vLLM online serving confirm the GX10’s consistency across AI models. Whether running GPT-OSS-120B or Llama 3.1 8B, the system demonstrated linear scaling with minimal dips—particularly in Prefill Heavy workloads, where throughput climbed aggressively from batch sizes of 4 to 64. Decode Heavy scenarios showed steady gains, though lower batch sizes exhibited slight variability, a common trait across Spark-based systems.
Storage and GPU Direct Storage: A Tradeoff Worth Noting
The GX10’s storage performance hinges on the included Phison ESL04TBTLCZ 1TB Gen4 SSD, which is the slowest write-speed drive tested in this platform. While read throughput scaled well (peaking at 5.12 GiB/s with 128 threads), write performance plateaued early, indicating saturation before maximum concurrency. Latency spikes at higher thread counts further highlight the bottleneck. A faster PCIe 5.0 SSD would mitigate these issues, but for now, users prioritizing high-bandwidth storage may find themselves constrained.
Direct Storage (GDS) testing underscores this tradeoff. Read operations scaled efficiently (up to 4.28 GiB/s), but write throughput stagnated at 0.72 GiB/s, with latency climbing sharply beyond 32 threads. This behavior is consistent across Spark systems using Gen4 SSDs, reinforcing the need for better storage pairings in future configurations.
The GX10’s real-world impact lies in its ability to deliver GB10 performance without the thermal overhead of larger workstations. Small businesses running AI inference or model fine-tuning will benefit most from its compact size and balanced thermals, though those relying on high write speeds should opt for PCIe 5.0 storage if available. The system’s predictable scaling and moderate power draw make it a practical choice for environments where efficiency is as critical as performance.
