Benchmark · on identical silicon

AMD wrote the manual. We built the machine.

AMD's own playbooks prove a consumer-grade AMD APU can run frontier-scale models on hardware you own. Getting there is a multi-step project that ends at a command line. Arsenale is the finished product on the same class of silicon: a faster inference stack, and a complete system that works the moment you switch it on.

Reserve a unit · £2,499 Talk to us

The DIY route, and the product

AMD's playbooks are a real milestone: the silicon vendor now documents, officially, how to run frontier-scale models on a consumer-grade AMD APU. That validates the entire category. The question is what stands between that documentation and a system a person or a department can actually use. Here is the honest, head-to-head.

Arsenale AMD's documented playbook
(do it yourself)
Inference throughput, identical hardware 2.5× prompt processing, faster generation The documented setup (baseline)
Memory setup Automatic at boot. Full unified pool. Manual reconfiguration, then reboot
Scale beyond one box Cluster units for 350B+ models, configured and supported Documented, but you wire and tune the cluster yourself
Setup to first answer Switch on. Ready. Multi-step build: drivers, runtime, container, model download, networking
Power under load About 100W, standard wall socket Depends on your build
The software Integrated: OS, model, agent runtime, dashboard You assemble the runtime, server, models and a web interface
What you end up with A working system and an agent workforce A raw API endpoint
Support UK support, warranty, 5 years of updates Community, or open an issue

Throughput compares the Arsenale inference stack against AMD's documented playbook setup on the same class of consumer-grade AMD APU, with identical model and settings: 370 versus 150 tokens per second prompt processing, and 23 versus 18 generation. Power measured at the wall under sustained load. Independent verification available under NDA. Last verified 2026-04-04.

AMD's published playbooks prove the platform. Arsenale is the finished product built on it: the same class of silicon, a faster inference stack, and the years of kernel, driver and inference engineering that turn a documented build into an appliance that simply works.

What the playbook actually asks of you

None of this is a criticism of AMD. Their playbooks are good, and they prove the platform. But this is the work that still stands between the documentation and a working assistant:

  1. Reconfigure system memory with a separate tool, then reboot.
  2. Install and maintain the GPU compute stack: drivers and runtime.
  3. Stand up an inference server, by container or built from source.
  4. Download the model by hand, often hundreds of gigabytes.
  5. Configure networking, and for the larger models, a second machine wired to a switch.
  6. Connect and manage a separate chat interface yourself.

The end state is a raw API endpoint that you own and maintain. Arsenale ships the destination, not the directions: the same class of hardware, a faster stack, a model, an agent runtime, a dashboard, and UK support, in one unit that works out of the box.

Ready?

Reserve a unit at the launch price, or write to us with the questions this page didn't answer. Either is fine.

Reserve a unit · £2,499 Write to us

More on the product: the platform overview, the personal page, and the trust page.