Asus Ascent GX10

Original Article ↗ Hacker News Discussion ↗

Overall Impression and Pricing

  • Many see the Ascent GX10 as essentially a rebadged DGX Spark/GB10 box with 128GB unified memory and 1TB SSD, priced around $3,000–4,000 depending on vendor and region.
  • Some are tempted by the form factor and RAM; others argue that for the same money you can build a more powerful traditional system (e.g., HBM Xeon, multi‑GPU desktop) or rent GPUs cheaply.

Hardware Specs and Memory Bandwidth

  • Unified 128GB memory is widely appreciated for fitting very large models or experimentation without sharding.
  • The revealed memory bandwidth (~270–300 GB/s LPDDR5X) is heavily criticized as “laptop‑class” and far below high‑end GPUs (e.g., 3090, RTX 5090, M‑series Macs).
  • Several commenters argue this makes large LLM inference slow and full training unrealistic; others counter that with high batch sizes it can still be fine for certain training/finetuning workloads.

Comparisons: DGX Spark, Macs, Ryzen AI / Strix Halo

  • Treated as the same architecture as DGX Spark; prior complaints about underdelivered performance and thermals are referenced, though some claim those early critiques misunderstood the hardware.
  • Compared to Mac Studio / MacBook Pro (M3/M4/M4 Max): Macs win on bandwidth, portability, resale value; GX10 wins on Linux and CUDA support.
  • Compared to AMD Strix Halo / Ryzen AI Max mini‑PCs: AMD options are cheaper and often competitive or faster on token/s benchmarks; GX10’s main advantage is CUDA and 200GbE clustering.

Use Cases and Niche

  • Consensus: this is not an optimal “fast home LLM box” for pure inference.
  • Seen more as a CUDA dev workstation / ARM Linux workstation with lots of RAM, and as a local prototyping node before scaling to cloud A100/H100/H200 clusters.
  • Chaining multiple units over 200GbE is considered interesting; gaming or general desktop use is seen as poor value.

Software, OS, and Ecosystem

  • Runs Nvidia’s DGX OS (Ubuntu-based). People have successfully installed other distros; tool support is still maturing.
  • Some report flaky UI/graphics behavior and dislike Nvidia’s extra management software layer, preferring simple SSH access.

Marketing, Website, and FAQs

  • The product page and FAQ are widely mocked: evasive non‑answers to “memory bandwidth” and heavy marketing jargon lead some to suspect LLM‑generated copy.
  • The ASUS site UX (images, popups, AI chat bot) and ASUS software/firmware quality in general draw strong criticism.