Eta Compute Brings Ultra-Low Power AI SoCs to Edge Devices

Article By : Sally Ward-Foxton

Synaptics becomes Eta Compute investor and licenses Tensai software, Eta Compute will sell boards based on Synaptics AI SoC...

AI chip startup Eta Compute has pivoted away from silicon to become a system solutions company specializing in AI at ultra-low power in edge devices. The startup has also announced a strategic partnership with Synaptics, wherein Synaptics has invested in Eta Compute and will exclusively license the company’s Tensai Flow software. Moving forward, Eta Compute will continue to sell its existing ECM3532 ultra-low power AI chip, though there will be no further generations of Eta Compute silicon. The company will also develop boards and modules based on both the ECM3532 and Synaptics’ new Katana AI SoC.

Eta Compute also announced it has closed a series C round of funding, raising $12.5 million, bringing the total funds raised to $31.5 million. The round was led by Synaptics, along with participation from existing investors. Synaptics’ chief strategy officer, Satish Ganesan has joined the Eta Compute board of directors.

Eta Compute’s ECM3532 is a heterogeneous multi-core SoC optimized for ultra-low power AIoT applications which can perform AI inference with only microwatts of energy. The company’s secret sauce is in a proprietary voltage and frequency scaling technique achieved without a phase locked loop (PLL), combined with Arm Cortex-M3 and NXP CoolFlux DSP cores which both operate voltage and frequency scaling independently. Either or both cores can be used to accelerate the AI workload. The company’s software stack, Tensai Flow, includes both a compiler for the ECM3532 and a “model zoo” of neural network models optimized for low-power operation.

Edge intelligence
Synaptics has been diversifying its portfolio into the IoT space, acquiring Broadcom’s wireless IoT business this summer (alongside docking station IC and software company DisplayLink).

Satish Ganesan, chief strategy officer at Synaptics told EE Times that the company is shipping SoCs into products like smart speakers, set top boxes, smart displays and headsets, where the trend is towards more intelligence at the edge. AI inference on battery power is quite a challenge; Synaptics has launched a platform for this market called Katana. The platform includes an AI-capable multi-core SoC optimized for low power consumption in edge devices, and will now also include Eta Compute’s Tensai Flow software.

“The intent is to add a family of processors as the market expands, because the scope of the market is pretty wide,” Ganesan said.

Synaptics Katana SoC
The first SoC in Synaptics’ Katana family is an AI-capable multi-core design (Image: Synaptics)

Like Eta Compute’s ECM3532, the Katana SoC uses a hybrid multi-core architecture (in this case, a standard Arm core and a custom DSP optimized for low power), but Katana offers a higher performance option that is sufficient to run more than one kind of AI inference in parallel. While there may be some architectural similarities with the ECM3532, the design is Synaptics’.

Why not just acquire Eta Compute and bring the team and software in-house?

“An investment was the right vehicle for us at this moment in time,” Ganesan said. “We’ll explore the market together and determine future steps at a later time.”

The two companies plan to go to market together.

“Eta Compute will have access to our Katana silicon to be able to develop modules, boards and platforms, depending on the customers, to go and tackle the market from their perspective,” Ganesan said. “For certain customers or certain applications, we’ll be able to pair up with the Eta Compute software and compiler and provide that as a bundle to our customers… it’s a joint development and joint promotion deal that we have reached.”

Energy efficiency
Following the launch of Eta Compute’s ECM3532, customers said their top priority other than energy efficiency was ease of use, Ted Tewksbury, CEO of Eta Compute told EE Times. This is especially true of heterogeneous processing architectures which can be challenging to program.

Eta Compute Ted Tewksbury
Ted Tewksbury (Image: Eta Compute)

“What we heard from the market very clearly is that you need a very easy to use, push button approach to interface these complex processors with TensorFlow, PyTorch and other AI frameworks,” Tewksbury said. “We’re working very closely with Synaptics to do co-optimization of the hardware and software to provide the most efficient code on the Synaptics Katana processor.”

Tewskbury said that Tensai Flow’s compiler can get an order of magnitude better energy efficiency than other compilers on the market, or proprietary ones. This is done through a variety of memory optimization techniques, as well as performance and energy optimization, but the software must be optimized for the specific target processor.

Tensai Flow’s compiler, originally optimized for the ECM3532, will be ported over to the Katana processor. Synaptics is an exclusive licensee – the compiler will not be optimized for other target hardware. Given the heterogeneous multi-core (Arm microcontroller plus DSP) nature of both chips, there is a lot of synergy between TensaiFlow and Katana, Tewksbury said.

“That’s one of the reasons for this partnership, our software is unique in the industry in being able to translate and compress and optimize from TensorFlow to hybrid multi-core processors, which are the most difficult to do embedded programming on,” Tewksbury said. “As we looked across the industry, Synaptics with Katana really had the most advanced, most sophisticated and most energy efficient hybrid multi-core processor out there. And so there’s a lot of synergy, because we’ve already developed those techniques for optimizing and compiling from TensorFlow to a hybrid multi-core architecture.”

Eta Compute Tensai Flow
Eta Compute’s TensaiFlow toolchain (Image: Eta Compute)

Business model
While Eta Compute has licensed its software to Synaptics, the company’s new business model is not software licensing. Instead the company will focus on more of an end-to-end solution, designing modules around its own, and Synaptics’, silicon.

“At this time we don’t plan to do any further Silicon development,” Tewksbury said. “We plan to rely on Synaptics for Katana, for the current generation, and then work co-operatively towards next generation products,” Tewksbury said.

Eta Compute sensor board
Eta Compute’s AI-enabled sensor node board (Image: Eta Compute)

Ditching silicon development will help Eta Compute to reduce its overheads. Going forward, Tewksbury plans to grow the company by building sales of the existing ECM3532, along with boards and modules based on the ECM3532 and Katana which use the Tensai Flow software.

“Eventually, we will wrap additional application software, whether it’s dashboards, analytics, business insights as a service, we’ll wrap those services around our modular solutions and sell those as complete solutions to the customer,” Tewksbury said. “That gives us an order of magnitude or more increase in our [average selling price]. So we’ve essentially pivoted from being a pure play semiconductor company to being a system solutions company.”

The first member of the Katana SoC family is available now, as is Eta Compute’s ECM3532. A version of Eta Compute’s Tensai Flow software tailored for Synaptics’ Katana hardware will be available in the first half of 2021.

Leave a comment