Big day at Axelera today. Two things landed this morning and, while the hardware is the headline, I reckon the software is every bit as interesting, especially to everyone here in the community.
AxScript
The short version is, you’ll be able to compile your own models. Not your weights on one of our models. Your actual model, your operators, your data movement, your scheduling, up and running on Axelera silicon.
If you’ve ever hit the end of a supported model list and thought "well, that's that", this is the fix. AxeleraScript, or AxScript, is a Python-based DSL that programs the AI cores directly. Two ways to use it:
- Write it by hand. There's documentation and an agentic harness that reads your source model and helps produce the AxScript for it, with you keeping control of what comes out. Evaluation and quantization tooling ships alongside, so you can check accuracy at each step rather than at the end.
- Or don't. AxGraph, the graph compiler, still takes ONNX and PyTorch and handles it automatically. If one layer needs hand-tuning, you write that layer in AxScript and AxGraph compiles it in. Same backend either way, so you are not picking a stack.
During internal testing, Qwen3.5 went from a weeks-long port to one day, one engineer, with Wingman assistance. It's also one language across the roadmap, so the same code targets Metis and Europa now and carries forward to Titania once that’s with us.
Availability: Models built with AxScript should start shipping in Voyager SDK 1.9 shortly. Language access goes to selected customers first for testing, then opens publicly before the end of the year.
Europa
The other half of today. Europa is the second-generation AIPU architecture, and it's shipping in two cards.
Per AIPU you get 629 TOPS from eight second-generation AI cores, 16 RISC-V vector cores, 128MB of L2 SRAM and 200GB/s of DRAM bandwidth, with pre- and post-processing and a video decoder on board so your host CPU stops doing the boring work.
Axelera Edge 232p is a half-height, half-length PCIe card with one Europa AIPU. Axelera Server 250p is full-height, full-length with four of them, for 2500+ TOPS.
Press release: https://axelera.ai/news/axelera-ai-launches-europa-delivers-physical-and-enterprise-ai-through-growing-partner-ecosystem-including-dell-and-supermicro
Edge 232p: https://axelera.ai/axelera-products/edge#232p
Server 250p: https://axelera.ai/axelera-products/server#250p
Over to you
The blog ends with a question worth stealing. If you have a workload you assumed could never run at the edge, what is it? Drop it below. It's genuinely useful for us to know what people want ported next, and it’d be great to start the list right here.
