MetaChip says its MEGA188 edge-AI coprocessor has reached engineering samples and model testing, with mass production targeted for the first quarter of 2027. Its pitch is to make local AI inference faster by placing memory close to computing, though the reported performance figures are company tests, not independent benchmarks.
Watch Desk analysis
What happened
The 28-nanometre chip combines logic and stacked DRAM using hybrid bonding and through-silicon vias. It is designed to work alongside a host processor through interfaces including PCIe and M.2, rather than replace the computer’s CPU.
PandaDaily reports that MetaChip presented the development at a Beijing event on 18 November. The company says the chip first powered on in October and has since passed milestones including PCIe connectivity and stacked-memory read/write checks. Samples and engineering validation are under way; volume production has not begun.
MetaChip claims peak memory bandwidth of 600GB per second and average chip power consumption of roughly 5–10 watts. In its tests, high-performance mode reached about 293 tokens per second with MiniCPM5-2B and nearly 80 with Qwen3-8B. Low-power mode produced about 92 and 26 tokens per second respectively. These are separate operating modes, not one combined promise of maximum speed at minimum power.
The distinction matters for larger models, too: MetaChip reports about 19 tokens per second for Qwen3-30B-A3B on one card and about 80 using multiple cards in parallel. The faster figure is not a single-card result. The company is also exploring storage-backed deployment for mixture-of-experts models, but says larger configurations remain under verification.
Why it matters
Local AI can be limited by how quickly a system moves model data, not simply by the processor doing the calculations. MEGA188’s approach is to bring more memory closer to the compute, potentially letting smaller or lower-power systems run inference workloads that would otherwise be awkward to fit.
That makes this a concrete hardware development, not just another chip roadmap. But the useful test will be whether engineering samples become reliable products with compatible interfaces, usable software and repeatable performance outside a company test. The production date remains a target, not a delivery record.
Our read
The interesting idea is the memory architecture and its focus on local inference. The numbers are worth watching, but tokens per second alone do not tell a buyer how a system performs on real applications, what it costs or how much power the whole setup draws. Promising silicon is a start; dependable hardware and software are the less photogenic part that makes it useful.
What to watch
- Whether MetaChip meets its first-quarter 2027 mass-production target.
- Independent tests of throughput, power use and model compatibility.
- Whether the multi-card and storage-backed configurations become supported products.
Discussion spark: For local AI hardware, would you prioritise faster inference, lower power use or the ability to run larger models on fewer cards?
Sources and evidence
- MEGA188 Advances Edge AI Chip Tests, Targets Production in Early 2027 (20 November 2026, 08:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.