The Shenzhen-based developer MetaChip has presented details on the progress of its Edge AI inference coprocessor, MEGA188. According to information shared at an event in Beijing on November 18, the chip has passed silicon spin and model testing stages. Work is currently underway on samples and engineering validation, with mass production scheduled for the first quarter of 2027, no sooner.
The company described the achieved milestones in an interview published by the Chinese tech publication Zhidx, and IT Home covered this development on November 20. It should be noted that this was not the chip's first inclusion in operation: MetaChip reports that the silicon was received on October 14, and it launched that day. PCIe connection was established on October 19, and stacked memory read/write verification followed at the end of October.
The chip design is implemented in a 28nm process technology and vertically integrates DRAM with logic via hybrid bonding and through-silicon vias. MetaChip claims a peak memory bandwidth of 600 GB per second, memory configurations of 2 GB, 4 GB, and 8 GB, and an average chip power consumption of about 5–10 watts. These figures are supplier specifications, not results of independent system efficiency measurements.
MEGA188 is intended to complement the host processor through interfaces such as PCIe and M.2, not to replace the computer's central processing unit. The main idea is that bringing memory closer to computing blocks can reduce the bottleneck associated with data movement during local language model inference.
In internal tests, the company achieved performance of approximately 293 tokens per second in high-performance mode using MiniCPM5-2B and nearly 80 tokens per second when running Qwen3-8B. In low-power mode, the rates were about 92 and 26 tokens per second, respectively. These modes cannot be combined into a single promise of maximum throughput at minimum energy consumption.
An important aspect is the difference between operating one card versus multiple cards. MetaChip reported that for Qwen3-30B-A3B on a single card, approximately 19 tokens per second are achieved, while with parallel operation of multiple cards, the rate increases to about 80 tokens per second. This multi-card rate is not the result of a single chip's operation, and the provided measurements do not establish performance for arbitrary queries, model settings, or applications.
The developer is also exploring the possibility of deploying a mixture-of-experts architecture with SSD support: full expert weights are stored on the drive, while stacked DRAM caches active experts. Larger configurations of the 120-billion-parameter class, utilizing multiple cards and distributed SSD storage, are still under verification.
A short-term hurdle remains the transition from engineering samples to reproducible and supported products. Interface compatibility, reliability testing, runtime software, and customer integration remain part of this work; the announced production period is the company's target, not a confirmed delivery volume.



