Today's silicon chips can trace the origin and fundamental design back to 1945 when mathematician John von Neumann first outlined the concept.
getty
Every silicon chip can trace its ancestry back to an architecture that was first outlined in 1945 by mathematician John von Neumann. Silicon chips process instructions a step at a time; at any given moment, only a small piece of code is actually running. This means that memory and processor can be physically separated, allowing programs to reside in memory while the CPU handles one instruction at a time, executing each one before moving on to the next step.
Since its inception, this processing model has worked relatively well, but electronics company Anker says that AI breaks that assumption. A neural network doesn’t chop up problems into steps, processing millions or even billions of learned parameters from end to end. Every inference requires one of those parameters to make a round trip between the memory and the processor.
In a data center, all that fetching of parameters translates into a huge engineering cost. In a wearable device powered by a tiny battery, this is an obstacle because more than 90% of the power a chip consumes is spent on moving data around, leaving very little over for computation.
Hungarian-born American mathematician Dr. John von Neumann (1903 - 1957) highlighted the need to program a device in terms of the fundamental computational functions it can perform.
Getty Images
Anker says the solution to the power and data paradox already exists in nature. Neurons in the human brain don’t separate information that is stored from where it is processed. Both are handled in the same place and that’s the principle used in Anker’s new THUS processing platform. Apple has done something similar in reverse by integrating memory in its M series processors.
Instead of shuttling model parameters between memory and processor, THUS embeds the computation directly inside NOR Flash memory cells. The model parameters aren’t moved anywhere. All the energy previously consumed by shuttling data is conserved and for computation. NOR Flash-based compute-in-memory requires just a sixth of the physical footprint of SRAM-based alternatives, making it a viable technology the tiniest of devices.
“Every AI chip built until now stores the model on one side and does the computation on the other. To think, the device has to carry all those parameters across, many times per second, every single inference. THUS puts the computation where the model already lives. The model never has to move again,” says Anker’s CEO and founder, Steven Yang.
Anker claims that THUS is the world’s first neural-net CIM AI audio chip and Anker has chosen to implement it first in a pair of earbuds, a demanding application for an AI chip because earbud batteries are so small.
ANKER
Anker claims that THUS is the world’s first neural-net CIM AI audio chip and Anker has chosen to implement it first in a pair of earbuds, a demanding application for an AI chip because earbud batteries are so small. Earbuds operate on milliwatts of power with almost no room for silicon.
Earbuds must continuously run noise cancellation features and the constraints ensured earbuds could only access small neural networks of a few hundred thousand parameters at the most, hardly enough to handle complex, real-world noise effectively.
Anker says its THUS chip removes that ceiling and can support several million parameters across multiple workloads, delivering up to 150 times more AI computing power on the environmental noise cancellation task than its previous generation of flagship earphones.
Anker wants to use its THUS chip in products like noise-canceling headphones from its audio brand Soundcore.
SOUNDCORE
The chip's first disclosed feature, Clear Calls, addresses a problem every earbud user encounters: what the person on the other end of a call hears. Traditional Environmental Noise Cancellation relies on signal processing algorithms or small neural networks that break down in loud and complex environments like a busy bar, airport and noisy streets.
Clear Calls replaces those rules with a large neural network that runs entirely on the device, anchored by eight MEMS microphones and two bone conduction sensors that isolate the speaker’s voice from its physical vibration.
Anker claims the result is a significantly cleaner phone call under any conditions. Additional AI-powered features introduced with THUS will include Signature Sound and Voice Control, which will be unveiled at Anker Day, being held on May 21, 2026, in New York.
The THUS AI chip is fabricated in Germany as part of Anker’s global manufacturing network. It will be a multi-year platform designed to bring local AI to Anker’s full product roadmap, everything from audio devices to mobile accessories and IoT devices. Future chips in the platform are intended to serve additional functions as the roadmap expands.
The first Anker THUS AI chip will be integrated into Soundcore’s upcoming flagship true wireless earbuds, due to be announced at Anker Day on May 21, 2026, in New York.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。