NVIDIA's GB10 superchip is sold in several desktops: NVIDIA's own DGX Spark and versions from Asus, HP, Lenovo and MSI. Because they share the same chip, the choice is less about raw speed and more about storage, software and who supports it.
What is the same on all of them
- NVIDIA GB10 Grace Blackwell superchip: a 20-core Arm CPU plus a Blackwell GPU
- 128 GB of unified LPDDR5X memory shared by CPU and GPU (about 273 GB/s)
- Up to about 1 petaFLOP of FP4 AI compute
- NVIDIA ConnectX-7 networking with two QSFP ports (200 Gb/s), so any two can be linked into a cluster
- A small desktop box with an external power adapter
So for the same model at the same precision, expect very similar performance across all five.
What differs
Storage
Systems ship with 1 TB to 4 TB of NVMe SSD depending on brand and configuration. Model files are large (a 70-billion-parameter model is often 40 GB or more even quantised), so if you plan to keep several models, datasets or fine-tuned checkpoints locally, pick 4 TB.
Operating system
The DGX Spark Founders Edition and most partner systems come with NVIDIA DGX OS and the AI software stack preinstalled. Some configurations, such as the Asus model we stock, ship without an operating system. They're aimed at teams that install their own image. If you want to plug in and start, choose a model with DGX OS preinstalled.
Design, cooling and noise
Each brand uses its own chassis and cooling. They're all small desktop units, but they differ in size, look and fan behaviour under sustained load. If it will sit next to you all day, that matters.
Support and fleet management
HP, Lenovo and MSI fit into their existing business support contracts and IT tooling, which helps if your company already standardises on one of them. The DGX Spark comes straight from NVIDIA.
Quick picks
- Just want the reference machine: NVIDIA DGX Spark Founders Edition
- Your company buys HP: HP ZGX Nano G1n
- Your company buys Lenovo: Lenovo ThinkStation PGX
- Installing your own OS image: Asus Ascent GX10 (no OS)
- Planning two from day one: a 2-node bundle with the cable included
Is a GB10 system the right tool at all?
GB10 systems shine at running and fine-tuning large models in a small, quiet, low-power box, thanks to 128 GB of memory. If your models fit in less memory and you need maximum speed, a workstation with an RTX PRO 6000 has far more memory bandwidth. Not sure? Check how much memory your model needs first.
