MetaChip promotes testing of Edge AI MEGA188 chip with production plans for early 2027
Read more
Pandaily
pandaily.com

MetaChip promotes testing of Edge AI MEGA188 chip with production plans for early 2027

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.

Similar stories

Google launches TPU 8tx and TPU 8i Max for data protection and increased memory in AI
Read more
olhardigital.com.br

Google launches TPU 8tx and TPU 8i Max for data protection and increased memory in AI

Google has unveiled two new artificial intelligence (AI) chip systems, named TPU 8tx and TPU 8i Max, which represent an expansion of the eighth generation of its processors. One of these new devices was specifically designed to ensure data protection during processing, while the other was created to manage large volumes of information used by AI agents.

With these launches, the corporation now offers four eighth-generation TPU systems in a single year, each tailored to different demands for model training and execution. These devices complement the TPU 8t and TPU 8i, which were presented in April during the Google Cloud Next event.

The TPU 8t is focused on large-scale model training, and the TPU 8i is intended for tasks requiring agile responses. According to Google, these new chips provide up to 2.7 times more performance in training and offer 80% better cost-performance ratio in inference compared to the previous generation, respectively.

The TPU 8tx is aimed at companies and government agencies that need to handle confidential information without it being exposed during the computation process. Meanwhile, the TPU 8i Max was designed for applications that require the simultaneous analysis of vast amounts of data, such as AI agents examining extensive code or long video recordings.

The TPU 8tx integrates confidential computing functionalities directly into its hardware. This allows data and internal model information to remain secure even when being utilized by the system. This feature is suitable for organizations that handle private data or need to comply with very strict security standards.

This equipment uses the same processing technology as the TPU 8t but combines it with fifth-generation AMD EPYC processors and the Boardfly network. Google reported that this combination results in a response time up to half that of the TPU 8t, without causing a significant impact on performance.

The architecture also supports adapting AI models using private data, creating reduced versions from larger models, and training agents without the need to transfer information between different server sets.

The TPU 8i Max consolidates two TPU 8i chips into a single physical structure, giving it double the processing capacity and memory resources compared to the standard model. Furthermore, Google mentioned that this system concentrates twice as many processing components in the same rack space, allowing for greater computational capacity in data centers.

In addition to the hardware, the company introduced tools to assist developers. TorchTPU, released as open source, allows applications developed with PyTorch and Hugging Face models to run on TPUs without requiring code rewriting. Another innovation is TPU Sync, which improves communication between chips, servers, and storage systems, optimizing access to information that AI reuses when generating responses. Google stated that this can decrease the data transfer time and waiting time for the first word or part of a response by over 90%.

Finally, Cloud Pathways became generally accessible on Google Kubernetes Engine (GKE). This tool facilitates the coordination of multiple sets of TPU processors as if they were a single unit, simplifying the execution of tasks that require high processing capacity.

Google reported that about 100 customers already use its TPUs, covering AI laboratories, financial institutions, retailers, and universities. The company itself employs artificial intelligence tools, such as AlphaChip and Gemini, to accelerate the development of its own processors.

Goke plans series of GK6950 Edge AI chips for robots supporting 10 TOPS and VLA
Read more
pandaily.com

Goke plans series of GK6950 Edge AI chips for robots supporting 10 TOPS and VLA

Goke Microelectronics is developing a series of Edge AI chips named GK6950, intended for robots capable of human interaction. This series is planned to be available next year.

Details of the plans were presented on November 2nd by Ni Yayu, Deputy Head of Digital Platform Development and AI Technical Lead, during a supply chain session held concurrently with the China Semiconductor Industry Conference 2026. The chip has not yet been released, and no information regarding samples, customers, or delivery timelines has been provided.

According to Ni, the GK6950 series utilizes a next-generation neural processor architecture that operates directly on the device, providing dense INT8 computation power of 10 TOPS. The chip supports various precision levels: mixed formats of INT4, INT8, INT16, and FP16. It is designed to handle traditional audio and vision perception, as well as inference of large models and vision-language-action models on a single architecture.

The chip's design is oriented towards the specific workload of a robot: simultaneously executing a large interactive model and three-dimensional spatial reconstruction. Goke asserts that the same NPU operators can support different precision levels for each task. Language model inference is performed using 4-bit weights, FP16 activations, and an INT8 key-value cache with page management, which helps reduce memory load and bandwidth when dealing with long contexts and parallel streams. Spatial geometry reconstruction remains in FP16 format, including intermediate features and outputs, to prevent error accumulation in pose and depth estimations.

Image preprocessing is carried out in dense INT8 mode to maintain high throughput when processing data from multiple cameras. In addition to the chip itself, Goke plans to develop a lightweight streaming spatial reasoning algorithm called GK-VGGT-SLAM. This algorithm allows scenes to be reconstructed causally as video arrives, using only current and previous frames for pose estimation and real-time 3D map construction, instead of collecting all material and processing it later.

Ni framed the problem in terms of the information robots must constantly know while moving, grasping, and placing objects: their own pose, the structure of the environment, and the location of the target object. Errors in pose, depth, or point matching determination can lead to artifacts and incorrect maps, negatively affecting collision avoidance and manipulation, while onboard computational power, memory, bandwidth, and power consumption remain limited.

Information regarding energy consumption, the technological node used, memory configuration, and performance on robotic test benches was not disclosed. Until silicon shipment, the accuracy scheme and SLAM claims remain design targets.

AutoChips introduces SoC AC8035 for cabin control, plans engineering samples in Q1 2027
Read more
pandaily.com

AutoChips introduces SoC AC8035 for cabin control, plans engineering samples in Q1 2027

AutoChips, a subsidiary of NavInfo specializing in automotive chips, has unveiled the AC8035—a sixth-generation automotive-grade microchip. This platform is designed to consolidate cabin control, driver assistance, and parking functions into a single chip.

The announcement was made at the NavInfo conference on November 6, according to a report by the Chinese publication Sina. Engineering samples are scheduled for release in the first quarter of 2027.

The proposed single-chip configuration is capable of handling smart cabin tasks, supporting L2/L2+ level driver assistance features, and automated parking. Furthermore, it supports running large artificial intelligence models directly on the device.

A development platform and sample release schedule were presented during the announcement, but neither mass production nor the use of AC8035 in existing vehicles was confirmed.

Claims Regarding Chip Capabilities

NavInfo CEO Chen Peng noted that chip development must account for future needs, as the semiconductor development cycle for automotive applications is longer than the rapid pace of AI model updates. According to Sina's report, the new design incorporates headroom in areas such as memory bandwidth to allow for subsequent model updates.

The Chinese industry resource Chedongxi reported that Chen claims the design can enable continuous multimodal interaction without the need for an activation phrase. He also mentioned memory bandwidth that is two to three times higher than comparable products, although the report does not specify the test configuration or an independently assessed benchmark. These statements remain management opinions rather than demonstrated performance in mass-market vehicles.

The integration of components aims to reduce the number of individual controllers and associated hardware, which could potentially simplify the vehicle's computing architecture. However, the reports do not provide a quantitative assessment of system-level savings for AC8035, nor do they indicate a customer vehicle program. The final chip price or mass delivery date has also not been disclosed.

Data Management Approach

Chen also emphasized the importance of data management as part of a broader offering for the automotive industry. Chedongxi quotes his foreign approach as 'data localization, model globalization': this means adapting data delivery and processing processes to local regulations while simultaneously creating reusable model capabilities. He added that compliance and privacy protection must be embedded in the data processing cycle, not added after the vehicle is ready for sale.

The broader conference included cabin-specific models and collaboration with Microsoft to create a multilingual ecosystem of large models for automakers expanding internationally. These initiatives fall under NavInfo's wider strategy and are not additional features of the AC8035 chip itself. For the chip, the next stated milestone is the release of engineering samples; validation and transition to production are still pending.

Popular