My First Impression on HP Zbook Ultra G1a: Ryzen AI Max+ 395, Strix Halo 128GB

HP’s Zbook Ultra sparks “toaster” jokes as it nips at desktop power

TLDR: A researcher’s HP Zbook Ultra laptop nearly matches a high-end desktop on memory-heavy workloads, but it draws big power and then throttles down. Comments split between awe at “portable supercomputer” and jokes about turning it into a toaster, debating heat, settings, and whether a desktop still makes more sense.

A forum post on the HP Zbook Ultra G1a claimed this slim laptop with 128GB memory and AMD’s new chip was pulling desktop-class numbers. In plain English: it chews through big math and science simulations, then tries local AI (the kind you run offline) with a huge context window. The jaw-dropper? It reportedly delivers about 80% of a beefy desktop’s performance on memory-hungry tasks. But it slurps power: heavy loads spike near 80 watts, then throttle down to ~45W after half an hour.

The crowd instantly split. Hype squad: “Finally, a portable supercomputer” and “128GB in a laptop—yes please.” The roast squad: “Is this a toaster?” Heat and throttling became the main drama—if it downs power over time, how “ultra” is it? There’s a mini-feud over Windows 11’s “Best performance” mode, with skeptics calling it a show-off setting and doubting the long-term numbers. AI fans cheered the local model result as “wild” while skeptics wondered if the GPU mode and settings were painting too pretty a picture. Gamers peeked at the benchmark charts and shrugged: great for work, maybe not a graphics monster. The vibe: applause vs. side-eye, with memes and popcorn at the ready.

Key Points

  • The HP Zbook Ultra G1a with AMD Ryzen AI Max+ 395 (Strix Halo) and 128GB RAM was tested under Windows 11 Pro 24H2 in Best performance mode.
  • Under sustained CPU or GPU load, power draw peaks near 80W, sustains ~70W briefly, then gradually reduces to ~45W after ~30 minutes, with ~10% all-core clock reduction.
  • Benchmarks run include CPU-Z, Cinebench R23, 7-Zip, 3DMark Fire Strike, and 3DMark Time Spy, with results shown via screenshots.
  • In a memory-bandwidth-bound FDTD simulation, the laptop achieved about 80% of a TR 5995wx workstation’s performance.
  • Local LLM testing via LM Studio with a Phi4 Q8 (15.5GB) model and 24k context ran on the GPU using “Vulcan,” successfully solving a complex integral; the author cautions possible detail inaccuracies.

Hottest takes

"are you hosting this on your toaster?" — cranberryturkey
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.