Prove every optimization on real hardware.
KernelX proposes, runs, verifies and improves kernel and graph optimizations on your accelerator.
An agentic loop for kernels and compute graphs
KernelX is an agentic optimization tool for accelerator kernels and compute graphs.
It proposes an optimization, runs it on real hardware, verifies correctness, and keeps improving it. An end-to-end, evidence-driven loop.
- Propose. Selects an optimization for a kernel or graph pass.
- Run. Executes it on real hardware.
- Verify. Checks correctness before it counts.
- Improve. Keeps what's faster and loops.
Bring your chip's capability model to life.
Tell us about your accelerator.
How KernelX optimizes performance
It builds a capability model and runs a loop that:
Analyze capability
Models what the chip can do.
Select optimizations
Chooses kernel-level techniques to test.
Run graph passes
Fusion, parallelization, scheduling.
Embed & verify
Verifies performance end-to-end in a runtime like llama.cpp.
Record & catalogue
Logs what works for the next pass.
What a hardware team gets
Full model enablement
A complete model running end-to-end, verified in a real runtime.
Automated optimization
Graph and kernel optimization run as a loop, not a manual pass.
Faster time to develop
Development and iteration with minimal human intervention.
Capability model & catalogue
Reusable knowledge of what works on the chip.
Feedback on the silicon
Evidence that informs the next hardware revision.
Deploying ML models on custom hardware, with continuous improvement
Fewer abstractions, open platforms, and LLMs working under expert guidance make continuous improvement on new silicon fast, explainable and repeatable.
New hardware needs a scalable AI compiler, fast.
- Costly and complicated. Stacks with many IRs that don't scale with changing AI workloads.
- Slow development. Long bring-up latency and a long tail of performance work.
- Lack of experts. Few people can reason across workload, compiler and hardware at once.
- 01Collapse unnecessary abstractions across the stack.
- 02Few experts, many agents, for quick enablement.
- 03Increase performance through KernelX — self-improving kernel generation with expert oversight, built on open-source platforms.
A performant MoE model on a 1,000+ core system, in days, through llama.cpp
Automatic self-improvement with expert oversight, under a limited token budget.
LPDDR4 DRAM, Minion cores at 600MHz. Esperanto Technologies' internal stack reached 11.51 tok/s on a Llama 3 8B model for comparison.
2x the raw decode throughput on a 4x bigger model, versus the native stack. No in-house algorithmic tricks or runtime customization, just better hardware utilization, workload understanding and KernelX. Repeatable for your hardware.
Let's talk about your hardware
Tell us about your accelerator and where performance is falling short.