Wedoany.com Report on Feb 25th, MatX Inc., a chip startup founded by former Google engineers, recently completed a successful $500 million funding round to accelerate the market launch of its first product. This Series B funding was led by Jane Street and Situational Awareness, with participation from over six institutions, including chip manufacturer Marvell Technology Inc. and a co-founder of Stripe Inc. Previously, MatX had secured over $100 million from a consortium that included similar backers.

MatX focuses on developing a processor specifically optimized for large language models, named MatX One. The company states that this chip offers higher throughput compared to existing graphics cards and can connect hundreds of thousands of accelerators to form clusters for handling large-scale training and inference tasks. Its design is based on a partitionable systolic array architecture, allowing circuit configurations to be customized based on datasets to enhance efficiency.

In terms of memory, MatX One uses SRAM cells to store most model weights, leveraging their high-speed characteristics to reduce latency and accelerate processing. Simultaneously, the chip employs HBM memory to store KV cache data, reducing repetitive computations by caching frequently used calculation results, thus saving time. MatX's research blog also reveals that the chip will integrate optimization methods such as speculative decoding and block-sparse attention to further improve performance.

MatX co-founder and CEO Reiner Pope wrote in a blog post: "This chip combines the low latency of an SRAM-first design with the long-context support of HBM. These elements, along with new ideas for numerical processing, deliver higher throughput on large language models than any announced system, while matching the latency of SRAM-first designs." The company plans to use the new funds to finalize the chip design and aims to complete the tape-out process within a year, which is the final stage of semiconductor development.