Urban NOA on a new chip (AXERA M95)

Technical lead for porting QCraft's urban navigate-on-autopilot stack to a new chip family: a staged bring-up from a 700-TOPS demo board to the 350-TOPS M95, model adaptation and quantization, and road-test issue clustering. The v4 stack runs on test vehicles.

QCraft 2026 Technical lead (since June 2026)

350 TOPStarget chip, J6M class
v4stack running on test vehicles
0frame drops on key models in road tests
  • AXERA M95 / M97
  • Urban NOA
  • Bring-up
  • Quantization
  • Program leadership
On this page

Why this chip, and why early

Our stack ran on Horizon chips. Once Horizon started selling its own driving stack, depending on one chip vendor became a commercial risk. AXERA's M-series offered competitive compute per cost: the M95 (350 TOPS) is comparable to Horizon's J6M class, and few chips above 100 TOPS were available at the time.

Experience with J6M and Qualcomm showed that the value of an adaptation falls month by month. The first teams with a working stack on a new chip get the customer conversations. So the program started before the M95 was ready, on an M97 (700 TOPS) demo board.

Plan

bringing the urban NOA stack to a new chipGET BEHAVIOR RIGHT ON THE ROOMY DEV BOARD FIRST, THEN OPTIMIZE ON THE TARGET CHIP1M97 dev board700 TOPSvalidate the stack: drivingfirst, then parking;correctness beforelatency2Model adaptationoperators · quantizationadapt and quantize everymodel for the newtoolchain, reusing thestage-by-stage checks3Target chip M95350 TOPSport and tune latency untildriving models haveheadroom in the framebudget4Vehicle testsurban NOA v4the v4 stack on testvehicles, with dev-testanalysis5Issue triagewith evidenceroad-test issues groupedby theme, each handed toits owner with evidenceResultThe urban NOA v4 stack runs on test vehicles with no frame drops on key modelsRoad-test findings tracked by theme, kept apart from porting issues:lane-change safetyspeed behaviornavigation lane choice
  1. Prove the stack on the demo board, timing second. On the 700-TOPS board, the priority was correct behavior: driving first, then parking. Latency could wait.
  2. Adapt the models. Adapt and quantize each model for the new toolchain, and check numerical consistency with the same stage-by-stage method.
  3. Bring up on the M95 and tune latency until the driving models fit the frame budget with headroom.
  4. Test on vehicles with the v4 stack, and run development test analysis.
  5. Cluster road-test issues so each goes to the right owner with evidence attached, instead of a list of anecdotes.

Results

Road-test findings were grouped into themes, each with its owner:

  • Lane-change safety: some lane changes came too close to vehicles alongside, or were taken despite traffic approaching from behind.
  • Speed behavior: too conservative on some open roads, and slow to accelerate after yielding to vulnerable road users.
  • Navigation lane choice: wrong lane selection in some scenarios.

These are mostly model-behavior issues, which the model teams continue to improve. Separating "the port is wrong" from "the model needs work" keeps a bring-up moving.

Merging the chip-specific code back into the main branch also exposed integration bugs, which were fixed so the port did not become a long-lived fork.

What this taught me

  • Sequence by value. Correct behavior on a forgiving board first, then performance on the target.
  • Reuse the method, not just the code. The consistency checks built for Horizon chips transferred directly.
  • Triage is leadership work. Clustering issues with evidence and owners kept a multi-team effort moving.