AMD snaps up MK1 to accelerate inference and reasoning on Instinct GPUs – News

Chipmaker Advanced Micro Devices Inc. has added to a string of recent acquisitions, buying a startup called MK1 that develops software to enhance the inference and reasoning capabilities of its artificial intelligence processors.

The acquisition was announced today in a blog post by Anush Elangovan, AMD’s corporate vice president of software development, who said MK1’s software will help the company to “advanced AI performance and efficiency” across its computing stack. The plan is for MK1’s team to join AMD’s Artificial Intelligence Group, he added.

MK1 is led by a friend of Elon Musk – the Neuralink Corp. co-founder Paul Merolla. At Neuralink, Merolla helped lead that company’s chip design efforts and develop algorithms capable of decoding brain activity. He founded MK1 alongside Thong Wei Koh, who was formerly a team lead at Neuralink, working on the development of brain neural signal processing systems.

Their work at MK1 is different. There, they have created software that’s designed to accelerate AI inference and reasoning on AMD’s Instinct graphics processing units, and it has been a big hit. According to Elangovan, the startup’s Flywheel technology currently serves more than 1 trillion tokens per day. He explained that the software is uniquely able to take advantage of the Instinct chip’s memory architecture, and that the acquisition will support the delivery of more “accurate, cost-effective and fully traceable reasoning” at much greater scales.

“Together, we’ll accelerate the next generation of enterprise AI, enabling customers to automate complex business processes and unlock new opportunities in high-value applications,” he said.

AMD’s acquisition spree

AMD has gained a lot of momentum in the AI industry in recent months, notably agreeing a deal with OpenAI Group PBC to deploy more than six gigawatts of Instinct-powered compute resources at that company’s data centers. A big part of the company’s success is down to its successful acquisition strategy, which has seen the chipmaker buy up startups that can enhance and optimize its software capabilities.

In its biggest acquisition so far this year, AMD paid $4.9 billion to snap up a company called ZT Systems, which manufactures hardware such as computer servers and integrated server racks for cloud data centers. That deal has allowed AMD to accelerate the development of rack-scale systems powered by its Instinct GPUs, and is believed to be one of the main factors in OpenAI’s decision to invest in its hardware.

It was an expensive deal, but the company was able to recoup most of that investment last month when it sold ZT Systems’ manufacturing unit to the U.S. electronics manufacturing services giant Sanmina Corp. for $3 billion, while retaining its design and enablement teams.

In its latest earnings call last week, AMD revealed it has spent $36 million on various other acquisitions this year. For instance, in June it spent an undisclosed sum to buy a startup called Untether AI Inc., which sold energy-efficient inference chips for data centers and edge networks. That deal came just days after it purchased a compiler startup called Brium, which enables AI software and applications to run on non-Nvidia Corp. GPUs without rewriting their code. It’s an important capability for AMD, as many developers optimize their software specifically to run on Nvidia’s hardware.

Back in May, AMD snapped up Enosemi Inc., a developer of silicon photonics-based networking chips that help to connect massive clusters of GPUs together so they can work in concert to run the largest AI models.

Last week, AMD revealed it had seen a significant increase in sales of its central processing units, as well as its Instinct GPUs, generating record-breaking revenue of $9.25 billion during its last quarter.

Image: AMD