Analog AI chip startup aims to achieve 1 POPS per Watt: 1000 TOPS/W is ‘within reach,’ says company co-founder as he looks at Gigabit-sized analog-first SLMs
- Blumind introduces ultra-efficient analog AI chip, reaching 10 nJ/inference
- Focused on wearables, healthcare, automotive and always-on AI
- Scale for larger models, aiming for 1000 TOPS/W performance
Blumind, an analog AI chip startup, has demonstrated a chip designed for low-power applications that achieves an impressive 10 nJ per inference, paving the way for the company’s ambition to scale analog computing to new heights .
The company showed off its test silicon for an ultra-efficient keyword spotting chip at Electronics 2024where co-founder Niraj Mathur told it EE times“What’s especially gratifying is that there has been more traction in the past year than we pushed.”
“People come to us specifically asking for analog AI solutions because they believe something new needs to be done.”
1000 TOPS/W is within reach
Blumind has already seen interest from the wearables, automotive and healthcare sectors. One of the examples the company gave involved a tire pressure monitoring system (TPMS) that could analyze road conditions.
The customer needed this to provide “extreme energy efficiency, because it’s in the tire, it has to last the life of the tire, you don’t want to open the tire to change the battery,” Mathur explains. Another potential use was detecting heart signals through a pacemaker sensor powered by energy harvested from muscle movements, requiring only a few hundred nanowatts of power.
The startup’s first product, an analog keyword spotting chip, is expected to be widely produced in 2025. It will be available as both a standalone chip and a chiplet that can be integrated into packages of microcontroller units. “Chiplets are the other way of integration for our customers,” Mathur said in his interview with EE times. This approach allows Blumind’s technology to complement fully programmable MCUs, focusing on always-on AI tasks.
Looking ahead, Blumind plans to scale its analog architecture for applications that require much larger models, such as vision CNNs and eventually gigabit-sized small language models (SLMs). Mathur said the company’s goal of achieving 1,000 TOPS/W is within reach, highlighting the potential of analog-first, multi-die solutions.
Despite his company’s ambitious roadmap, Mathur emphasized the importance of a pragmatic approach. “No one has really taken analog computers to high-volume production and delivered on its promise. We want to be the first to do that, but we want to walk before we try to run,” he said.