
Today we’re announcing XPU Grasshopper, our AI architecture co-designed with AI!
It went from its first tokens to 816x faster in 13 weeks on FPGA.
With a very small team and autonomous AI agents, we were able to design the hardware and software together. The agents rewrote kernels, reworked the hardware overlay, and ran and tested their own FPGA builds.
Running gemma-3-1b-it (yes, it runs other models too, but that’s for the next launch :)) at more than 51 tok/s and over 90% utilization at batch 1, on four tiles on an FPGA, measured end to end, to prove out our micro-architecture and software stack before silicon.
We wrote up the full story, including the day-by-day optimizations the agents found, the Mutation Tile Unit architecture, and the benchmarks: https://zscc.ai/announcing-xpu-grasshopper
100 numbered FPGA devkits are open for pre-order, with early cluster API and SSH access: https://zscc.ai/grasshopper
Read more: https://zscc.ai/announcing-xpu-grasshopper