AI News

DeepX's second-generation edge AI chip DX-M2 tapes out, with PIM boosting performance by 6.7x

Semiconductor Industry ObserverUpdated Nov 9 18:25

AI summary

South Korean edge AI chip startup DeepX announced that its second-generation AI accelerator DX-M2 has taped out. It can be equipped with Samsung's custom LPDDR5x-PIM, running models with 20 billion to 100 billion parameters at 5 watts of power, with an inference speed of 10 to 30 tokens per second. Simulations show that after offloading the GEMV operations of Llama3.1-8B to PIM, tokens-per-second output increases by 6.7x, but overall power consumption rises to 2.7 times the original. DX-M2 is expected to sample in March 2027 and enter mass production in fall 2027. The third-generation DX-M3 is planned for launch in 2028 and will support standard LPDDR6 PIM.

Why read this

AI score 55

The original text discloses the tape-out timeline for DX-M2, the 6.7x inference improvement brought by PIM, and its power cost, which can be used to gauge the pace of deployment for edge AI chips.

Topics

Tags

Body · AI translation

DeepX's second-generation edge AI chip taped out, performance boosted 6.7x with PIM support Source: Ban Dao Zong Heng Published: 2026-11-09 16:54 AI chip South Korea Generate poster DX-M2 is expected to sample in March 2027 and enter mass production in autumn 2027.

South Korean edge AI chip startup DeepX has announced that its second-generation AI accelerator has taped out, and the chip can be equipped with Samsung's custom processing-in-memory (PIM) technology. The DX-M2 targets robots, drones and other edge devices for deploying generative AI, and can run large models with up to 100 billion parameters at under 5 watts of power consumption.

DeepX's first-generation chip, the DX-M1, is now in mass production and focuses on vision applications. The newly taped-out DX-M2 is designed to make local AI viable on devices with limited power and thermal conditions. Depending on the model, the DX-M2 can run models with 20 billion to 100 billion parameters (about 3 billion to 20 billion effective parameters), at an inference speed of 10 to 30 tokens per second, with power consumption of just 5 watts, the same as the DX-M1.

DeepX founder and CEO Lokwon Kim said in an interview: "The most core change between the M1 and M2 is memory bandwidth." Memory bandwidth has increased from 50GB/s to 150GB/s, a threefold increase, along with multiple energy-efficiency optimizations. DeepX says that through architectural optimization, the NPU's energy efficiency has doubled; combined with the 2nm process, energy efficiency doubles again, with peak compute reaching 80 TOPS.

Kim said: "This is critical. 5 watts of power consumption means our chip can be used in all kinds of scenarios, whether drones or other battery-powered devices. Our goal is to enable high-performance AI to be deployed on all kinds of terminal devices." The DX-M2 supports the mixture-of-experts (MoE) architecture for large models, and is also compatible with new technologies such as speculative decoding and TurboQuant.

Processing-in-memory (PIM)

The vast majority of AI inference tasks are limited by the movement of weights between memory and compute units, especially general matrix-vector multiplication (GEMV), an operation that requires frequent random memory access and slows down computation. Memory vendors have introduced PIM solutions that perform memory-intensive operations such as GEMV directly inside the memory chip. Performing computation in memory can effectively boost equivalent memory bandwidth and reduce time-consuming memory accesses.

PIM is about to become part of JEDEC's LPDDR6 standard. DeepX has already partnered with Samsung to integrate the DX-M2 with Samsung's LPDDR5x-PIM solution. Kim explained: "We have been working with Samsung's memory team on this technology for a year and a half. Samsung has developed its own custom PIM adapted for LPDDR5x. We have received LPDDR5x-PIM samples and are internally testing the compute logic and performance."

DeepX's simulation results show that offloading the GEMV operations of the Llama3.1-8B model to LPDDR5x-PIM improves tokens-per-second output performance by 6.7x compared with running the DX-M2 alone. Kim said: "The DX-M2 with LPDDR5x-PIM is already close to the optimal solution. However, the next-generation DX-M3 will support standard LPDDR6 PIM, at which point we can unleash full theoretical performance, with performance improving by another 5 to 8x."

Kim also noted that the performance gain is not without cost: achieving the 6.7x performance increase raises overall power consumption to 2.7 times the original. Even so, DeepX believes that a 2x performance improvement can still be achieved while maintaining the same power cap as non-PIM mode.

Kim mentioned that the compiler requires a great deal of manual adaptation to support this operating mode. DeepX has already run a prototype compiler based on simulation of the PIM version of LPDDR5x. "We are now building hardware modules to verify in a real environment, so the compiler is not fully complete yet, but the team is pushing forward at full speed."

Core markets

The DX-M2 targets several key tracks. DeepX is working with ODM manufacturers in Taiwan, China, to explore deployment scenarios for AI PCs (laptops, workstations). Tim Park, DeepX's director of strategic marketing, said: many robot makers are still using GPUs such as Nvidia Thor for prototyping, but they need lower power consumption and higher on-board compute; automakers need low-power chips to run AI agents inside vehicles. In the IoT space, DeepX has partnered with Samsung to develop a new generation of smart gateways. The company has also joined the AWS IoT ecosystem, using Greengrass to enable remote model updates.

Park said most robotics customers first develop with GPUs, but are forced by thermal and power constraints to look for alternatives. DeepX is pushing customers to migrate from Nvidia's Isaac robotics platform to its own Newton platform. "Customers care about our hardware, and they also care about how to migrate from Nvidia. The answer is: customers only need to change the API to call Newton instead of Isaac, and they can run the exact same pipeline on the DeepX platform."

Park said that 17 companion libraries have now been adapted. The team has demonstrated a robot based on the DX-M1, reusing source code originally intended for Nvidia Jetson and replacing only the libraries with DeepX's.

Kim said: "After customers switch from Nvidia's solution to DeepX, battery life can improve by 20 times at lower cost. And the switch takes less than a second—you just need to change the API name from Isaac to Newton."

A complete robotics software stack for DeepX hardware is still under development. DeepX is engaging with robotics customers in the United States, China, Vietnam, and South Korea, and Hyundai Motor Group's robotics lab is already using the DX-M1 on its robot platforms.

Meanwhile, the company continues to expand its ecosystem, working with Taiwan ODM manufacturers and leaders in various niche segments, and recently reached a partnership with Intel. Through its collaboration with Intel, DeepX is integrating OpenVINO into the DeepX SDK to reach Intel's x86 customer base. The DX-M1 can complement the NPU built into Intel Core Ultra CPUs, and can also offload or take on the entire AI compute load for Intel edge CPUs that lack an NPU.

Kim revealed that the company recently held talks with AMD to jointly build a robotics platform based on DeepX accelerators and AMD Kria development boards, targeting the South Korean ecosystem first. In addition, DeepX is engaging with heterogeneous ecosystems such as Raspberry Pi and Baidu PaddlePaddle.

Commercialization Progress

Kim said the DX-M1 has already received more than 100 orders, covering major global markets, with applications including drones, robotics, smart factories, security cameras, aerospace, healthcare, IoT, and other physical AI scenarios. The DX-M1 is now in mass production.

Kim said that China was previously the most active market for physical AI, and this year Japan has become a new major market, where local companies show a strong willingness to deploy humanoid robots. "More than ten leading Japanese companies are evaluating our chips, and the feedback has been very positive. The U.S. market is also performing well, with several major companies giving us very good feedback."

DeepX is also advancing smart city projects in South Korea, Vietnam, and Singapore. Park said: "Our long-term judgment is that as energy efficiency continues to improve, AI compute will move toward distributed deployment. The PC and smartphone eras proved that compute can penetrate massive numbers of endpoints. AI will follow its own path, but when applications need to read sensor data in real time or operate without a continuous network connection, the value of local inference is very clear."

Park believes edge AI is a huge opportunity for chips, especially at a time when power resources for newly built data centers are very tight. Kim's view is that just as only a quarter of CPU compute is now deployed in data centers, the AI processor market will also shift toward the edge over the next 5 to 10 years. He added that many AI data center projects are now delayed, and the compute gap cannot be solved by increasing investment alone. An AI chip with just 5 watts of power consumption can move AI compute to endpoints. "If endpoint devices can deploy models with up to 100 billion parameters, AI traffic in data centers can be reduced by more than 80%. In other words, without adding new AI servers, the scale of compute infrastructure can expand fivefold. That is our vision."

DeepX's third-generation product, the DX-M3, is planned for launch in 2028, using a chiplet architecture with LPDDR6-PIM, targeting 1 ePOPS (INT8) of compute and aimed at automakers and humanoid robot manufacturers, scenarios that require multi-chiplet solutions. The DX-M2 is expected to sample in March 2027 and enter mass production in fall 2027.

This article is reposted from media reports or online platforms and reflects the author's personal stance or views. Our reposting is for sharing only and does not mean we endorse or agree with it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service in a timely manner, and we, as a neutral platform service provider, will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

"Homogeneous + Heterogeneous" dual launch! MetaX, the only domestic GPU at the China Pavilion 2026-11-09 17:55

Major photoresist acquisition! $12 billion, JSR changes hands 2026-11-09 17:51

Targeting monthly output of 700,000 memory wafers, Musk's Terafab goes straight after Samsung 2026-11-09 17:48

Source

Semiconductor Industry Observer