A former executive leading Qualcomm's Snapdragon computing platforms has departed the company to become CEO at an AI chip startup.
Keith Kressin will lead product commercialization for MemryX, which was founded in 2019 and makes memory-intensive AI chiplets.
The company is now out of stealth mode and will soon commercially ship its AI chips to non-tech customers. The company was testing early generations of its chips with industries including auto and robotics.
The previous CEO and company co-founder, Wei Lu, will slip into the role of CTO.
Qualcomm last week introduced new Snapdragon chips for Android smartphones and Windows PCs at the company's tech summit in Hawaii. Kressin bolted Qualcomm on Friday, once the announcements and festivities were complete, and joined MemryX this week.
The MemryX AI chip is more like a programmable mesh, with memory and logic coupled tightly that could be mapped to an AI model. The company's initial focus is on computer vision and on smaller networks with localized memory, with no need to access external DRAM, which could slow down AI applications.
The chip includes embedded floating point execution units, which are typically found on larger AI chips like GPUs. That will aid in faster low-level processing to analyze and identify patterns in data.
"Part of my job is to figure out which customers we're gonna prioritize and how to scale -- that's what we look forward to doing in 2022," Kressin told The Register.
The company's target is mainly edge devices and commercial customers in markets like auto, many of which are taking chip development in their own hands to take control of the supply chains. The chiplet model allows MemryX to resize the chips for larger deployments.
"We'll take their model, run it and show them their performance of their model in their office. That'll be something that others can't do," Kressin said.
AI chips are a hot market, with Cerebras Systems last month raising $250 million in its latest round of funding. Cerebras has the world's largest chip, and while MemryX's chip comes nowhere close in size, they share a common theme -- a big focus on memory.
MemryX's chip is power-efficient and can be tweaked for specific workloads, while Cerebras can handle multiple AI models on a single chip. The MemryX chip cuts of the switching fabric typically found in larger chips, which helps reduce the size and other overheads.
- Qualcomm serves up trio of new chips garnished with lavish ladles of AI gravy
- Qualcomm taps Samsung to make next-gen 10nm Snapdragon
- Intel updates mysterious ‘software-defined silicon’ code in the Linux kernel
- Alibaba teases a breakthrough chip, merging processor and memory
The AI chip market is dominated by Nvidia, but many AI chips are being implemented in edge devices and the cloud to run inferencing and deep-learning applications. Amazon's AWS recently announced it was announced it was bringing Habana's AI chips to its cloud offerings.
But the challenge for non-tech companies has been figuring out the cost involved in deploying an AI model, which includes costs related to time, labor, electricity. The goal of CIOs is to figure the outcome is worth the time and money spent.
Kressin's biggest challenge will be to contend with many other AI chip startups trying to address those fundamental AI questions posed by CIOs, said Nathan Brookwood, a chip analyst.
AI is a hodge-podge of techniques, and there's no analytical way to sort out the best approaches to solve problems, Brookwood said. AI chip startups have to first prove that AI software works, and even more challenging, that it works on their hardware.
"Right now we're in the hype cycle of AI. You go from the hype cycle to people getting disappointed, and reality sets in, and real growth starts to take place," Brookwood said. ®