By Neil Tyler
Ambient Scientific has announced that its first AI SoC, the GPX10, designed to run cloud-free AI applications on portable battery-powered devices has gone into volume production.
The GPX-10 is manufactured at TSMC using 40 nm technology and comes with a complete development platform to enable On-Device/IoT AI Applications, that includes a Software Development Kit with Neural Network Compiler.
The Evaluation kit will be available for commercial purchase in Q1 2024.
Created using DigAn technology, GPX10 utilises a combination of high-speed digital and analogue circuit technologies to enable best-in-class AI throughput with very low power consumption.
According to Ambient, the GPX10 provides a number of significant benefits for edge AI such as:
Performance Efficiency >8 TOPS/W for on-device AI applications at 40nm (compared to lower efficiency from competitors at deeper process nodes)
<100 µW of for always-on sensor fusion applications to run on coin cell batteries
Compact form factor – smaller than a fingernail for tiny portable devices
Enhanced privacy & security – due to limited data transfer over internet
The GPX10 is also completely programmable and combined with Ambient Scientific’s in-house software toolchain including an AI compiler, development libraries, training toolkit and more, will enable on-device battery powered applications across application verticals (sensor fusion, audio, and vision) and across multiple industries such as automotive, healthcare, consumer wearables and industrials.
“We’re already partnering with customers to deliver AI applications powered by GPX-10, that includes a driver monitoring system, ECG wearables and always-on voice-based distress systems,” said Head of Strategy and Business Development at Ambient Scientific Satish Kutty M,.
The team at Ambient Scientific said that it was eager to see how the industry develops products over time, using its edge AI development platform with the GPX-10 AI processor.
AI has traditionally been relegated to the confines of expensive cloud computing, limiting accessibility for the masses. Ambient Scientific says it’s on a mission to correct the trend and democratise sustainable AI for all.
“To bring the power of AI to the edge, we needed to free it from the shackles of the cloud and internet which requires a complete rethinking of conventional computing architecture”, said founder and CEO of Ambient Scientific GP Singh.
Consequently, has looked to innovate at every level from circuit technologies to a completely new Instruction Set Architecture (ISA) including:
DigAn Matrix computer: A combination of high-speed digital and analogue circuit technologies to achieve very high-performance matrix computing at very low power consumption
3D memory with in-memory computing: A deep circuit innovation that assists the DigAn matrix computer in disbursement of large number of operands for matrix computer thereby allowing DigAn matrix computer to achieve very high throughput using power efficiently
Multi-channel Ultra Low power Sensor ADC: Efficient sensor ADC that can connect to multiple analogue sensors simultaneously for AI applications that require intelligent sensor functions
Always-on Sensor Fusion DMA: A special functional architecture that allows extremely low power (microwatt) intelligent sensor AI applications on multiple sensors data fusion.
“We have leveraged a combination of digital and analogue computing concepts to bring the best of both worlds tougher – scalability of digital and efficiency of analogue. Our processors are manufactured and can be programmed like any digital chip while carrying power and area efficiency of analogue,” said Singh.
Ambient’s MX8 AI core based on this technology, is hyper scalable – across any number of cores and down to any process nodes – to enable a robust roadmap of processors. These include GPX-1, a single core processor for integration with MEMs sensors to turn them “intelligent” all the way to GPX-64, a vision focused chip designed for high-speed video AI applications.
