Ambiq has unveiled the first member of its Apollo5 family of SoCs, the Apollo510. Addressing the need for greater power efficiency and fast performance to deploy today’s speech, vision, health and industrial AI models on battery-powered devices, the company claims that the MCU delivers 30× better power efficiency and 10× faster performance.
(Source: Ambiq)
The 30× energy improvement can support most of today’s endpoint AI calculations, including low-power sensor monitoring, always-on voice commands and telco-quality audio enhancement. IoT devices performing AI/ML inferencing, such as next-generation wearables, digital health devices, AR/VR glasses, factory automation and remote monitoring devices, can expand their power budget and add more capabilities with the Apollo510’s SPOT-optimized design. The SPOT platform is Ambiq’s proprietary technology to optimize energy efficiency and power consumption.
Ambiq used the MLPerf TinyAI benchmark suite to measure the improved power efficiency and faster performance of the new Apollo510. “We used the industry-standard MLPerf Tiny AI benchmark suite and compared [the Apollo510] against published numbers for Cortex-M4 devices running the same software,” said Carlos Morales, VP of AI at Ambiq. “There are multiple benchmark models; the 30×/10× [improvements] are averages.
Ambiq’s Apollo510 MCU. Click for a larger image. (Source: Ambiq)
“The additional performance and power efficiency help in three ways,” he added. “It allows us to increase the sophistication of the model without making it run slower or consume more power. For example, we can create models to find multiple heart conditions rather than just Afib.”
It also increases the accuracy of the model, Morales said. “Our speech-enhancement model will do a much better job on Apollo5 at eliminating background noise because we have 10× as many CPU cycles to spend on it. We’re getting big gains with only about 5×, so that has the added benefit of leaving more cycles available for other tasks.
“Finally, it allows us to run faster—analyzing video at 50 fps is more useful than analyzing it at 5 fps,” he added.
To achieve the improvements in the Apollo510 MCU, Ambiq overhauled its hardware and software, leveraging the Arm Cortex-M55 CPU with Arm Helium. The result is processing speeds up to 250 MHz, up to 10× improved latency and lower energy consumption by approximately 2× when compared with Ambiq’s Apollo4.
“If your previous generation of the device had a two- to three-hour battery life while running AI, you could expect an Apollo510-powered device to run much longer, even lasting up to a full day,” Morales said. “If you had to throttle your device’s AI capabilities because you were concerned about battery life, you can now enable them or make them more sophisticated.”
The Apollo510 also integrates a 2.5D GPU with vector graphics acceleration that provides crystal-clear graphics and a 3.5× overall performance enhancement over the Apollo4 Plus family. Additional features include 4-MB on-chip NVM, 3.75-MB on-chip SRAM and high-bandwidth interfaces to off-chip memories. For extra-large neural network models or graphics assets, the high-bandwidth off-chip interfaces are individually capable of peak throughputs up to 500 MB/s and sustained throughput over 300 MB/s.
Building on its secureSPOT platform, security features include Arm TrustZone technology with a physical unclonable function, tamper-resistant OTP and secure peripherals for a trusted execution environment. Arm Helium technology supports up to 8 MACs per cycle and half-, full- and double-precision floating point operations for AI calculations and general signal-processing operations.
When asked how the Apollo510 differs from competitive offerings, Morales said: “Honestly, our biggest competition is and has always been ‘good enough.’ For some device manufacturers, it’s good enough to extend the battery life to two, three or five days when they could extend it to two weeks.
“We are a bit more expensive than the other MCU/SoC manufacturers who don’t care much about battery life or performance,” he added. “However, once the manufacturers tested and saw the big differences in battery life and performance enabled by an Ambiq chip versus other chips, it became more difficult for them to forgo what they could have had/gained by adopting the Ambiq chip. The energy efficiency based on Ambiq’s SPOT platform is not just 30% over competitors but 30×—the latest Apollo510 has proven that. This is why Ambiq always becomes the single-sourced vendor for many device manufacturers. When they see the performance comparison, it’s difficult for them to choose which product line adopts Ambiq.”
The Apollo510 MCU is currently sampling. Availability is expected in the fourth quarter of 2024. The Apollo510 MCU won the Embedded Award in the hardware category at embedded world 2024.
Advertisement
