krishnan questions SiMa.ai Palette 3.0's single-chip robot AI claims
SiMa.ai launched the Palette 3.0 edge inference platform on November 17, claiming its Modalix single chip can concurrently run perception, depth, and inference models (including LLMs/VLMs) under tight power and memory constraints, with speculative decoding delivering roughly 3.5x faster token output. Analyst krishnan followed with an 8-tweet thread systematically questioning the solution's viability in real-world robotics.
Confirmed
- SiMa.ai released Palette 3.0 on 11-17, claiming the Modalix single chip concurrently runs perception, depth, and inference models under power and memory constraints, with speculative decoding delivering a 3.5x speedup.
- krishnan's core objection: robots cannot pause visual perception while a language model thinks, and vendors' throughput claims are not results for robot safety or task completion rates.
- SiMa provided details of its competitive benchmarks via email, but krishnan considers them vendor throughput marketing.
- Even with container isolation, vision, depth, and inference models still contend for the same silicon's compute and memory — "shared silicon is still shared."
- krishnan proposed benchmark claims for robotic edge AI: test the actual shipped model combination rather than summing isolated scores; run continuously for hours (including thermal throttling) and then measure observed data age, worst-case response latency, and energy per completed task, with independent stop mechanisms participating in the tests.
- He also proposed a thought-experiment test: whether a warehouse robot encountering an unexpected obstacle mid-LLM-inference can brake in time (this is krishnan's proposed test scenario, not a deployment reported by SiMa).
Not Yet Confirmed
- No public data exists on Palette 3.0's safety performance or task completion rates in real robot deployments.
- Whether the 3.5x inference speedup from speculative decoding translates into "fresh" perception data is unverified; krishnan argues inference acceleration does not equal perception freshness.
Why It Matters
- Single-chip multi-model integration is a hot selling point for edge robotics AI, but krishnan reminds buyers: before purchasing a faster inference stack, make the vendor prove the rest of the robot doesn't stall while inference is running.
- He argues hardware separation remains a serious option — a dedicated controller with an independent safety path may be better; single-chip integration only delivers its cost savings if it never makes critical behaviors wait on a busy inference model.
- He also notes that enforcing resource budgets for robots is itself a product opportunity.
2026-11-18 ~ 2026-11-18 · 8 related posts
Primary sources
- SiMa.ai Palette 3.0 runs perception and LLMs concurrently on one chip, claims 3.5x via speculative decoding — krishnan ·
- Hardware separation for robot safety remains serious option; resource budgets a product opportunity — krishnan ·
- Before buying a faster reasoning stack, ask what the rest of the robot does while it reasons — krishnan ·
- A robot can't pause perception while its LLM thinks: SiMa Palette 3.0 single-chip claim under scrutiny — krishnan · 2026-11-18
- Speculative decoding doesn't prove perception stays fresh during long robot reasoning — krishnan · 2026-11-18
- [source] SiMa.ai Palette 3.0 runs perception and LLMs concurrently on one chip, claims 3.5x via speculative decoding — krishnan · 2026-11-18
- Vendor throughput claims aren't robot-safety results: shared silicon is still shared — krishnan · 2026-11-18
- Proposed test: can a warehouse robot still stop in time while its LLM is mid-reasoning? — krishnan · 2026-11-18
- Benchmark the model mix that ships: observation age, worst-case latency, energy per task — krishnan · 2026-11-18
- [source] Hardware separation for robot safety remains serious option; resource budgets a product opportunity — krishnan · 2026-11-18
- [source] Before buying a faster reasoning stack, ask what the rest of the robot does while it reasons — krishnan · 2026-11-18