|
Listen to this story
|
Ever dabbled with DIY electronics projects? If yes, chances are you’ve already met Raspberry Pi – a name known to hobbyists and industry users alike. Recently, the company announced a strategic partnership with Sony’s semiconductor division to bring a line of Sony’s edge AI devices to the Raspberry Pi ecosystem.
“We see a lot of people using Raspberry Pi for AI/ML applications,” said Eben Upton, CEO of Raspberry Pi Ltd. The use cases vary from hobbyists using it to classify animals they saw in their backyard to actual industrial applications such as classifying objects on the production line into good or bad.
People until now were either using third-party accelerators such as the Google USB 3.0 accelerated device which could be directly connected to Raspberry Pi, or were doing the AI work on the CPU. “For a lot of AI applications, the CPU is actually sufficient,” said Upton. However, it has its own advantages and disadvantages.
“What this [partnership] is getting us access to is solely a suite of AI-accelerated vision products. So particularly, product line X500, which is an image sensor that has an integrated CNN inference accelerator, and other AI products and services,” he added.
Open, but not so open
Raspberry Pi has previously faced the heat for not being completely open. Almost all of their products have proprietary silicon on them. Thus, although at the application layer it may be open, the internal systems, including the HAL layer and drivers, are closed and not accessible to users.
Speaking about this, Upton stressed, “I think what’s important is not whether the hardware is open-source or proprietary, but whether you can run standards-based ML models on the hardware.”
Sony, for instance, has a DSP (Digital Signal Processor) which accelerates convolution operations by providing the tooling required to convert TensorFlow for closed systems. “TensorFlow is kind of a de facto standard, I guess, for model representation.” Or they provide tooling that will convert from an open standard, or a pseudo open standard or something like TensorFlow, which is technically a Google proprietary standard but open. The toolings act as a bridge between the open world and closed systems.
“And so it [the hardware] doesn’t need to be open. It merely needs to implement an open standard. It needs to provide access to its proprietary capabilities through an open standard and we do this with openMAX for video, likewise for imaging purposes and KMS composition. Similarly, we’ll do it for TensorFlow for machine learning acceleration,” he added.
RISC-V-based Raspberry Pi?
Although Raspberry Pi is a member of RISC-V International, Upton says that the value proposition provided by RISC-V isn’t a slam dunk. While the upside of having the flexibility to add instruction sets for domain specific applications is certainly big, he stressed, “there is a formidable barrier between RISC-V and volume success, in particular around core maturity and software maturity”.
There is thus a trade-off between freedom and certainty, and depending on one’s viewpoint of this trade-off, they can pick an architecture.
But, Upton is still hopeful. He gives the example of ARM which used to have the same problem – in relation to core maturity and software maturity – but they resolved it, so, he says, “there is no reason to believe that in ten years’ time, there wouldn’t be RISC-V based products coming out of Raspberry Pi.”
“If ARM maintains their current trajectory in terms of the fraction of the value they want to extract from the system, it’s going to be very hard for them to justify the trouble and expense of migrating to a different instruction-set architecture. So, from this point of view, I would say it is ARM’s to lose.”
And while ARM has publicly downplayed the risks of open-source ISA to their chip empire, they do see the customer demand shifting. In 2019, the company announced Custom Instructions, a new feature of its Armv8-M architecture for embedded CPUs. So, we can expect them to reorient themselves against the changing tide.
What next?
Over the years, the evolution of Raspberry Pi has been such that now over 70% of sales come from industrial and commercial applications. This is why one of the challenges in the past year for them has been trying to balance providing the hobbyist space with material against sustaining their existing OEM customers.
The company, which is recovering from the supply chain constraints over the past two years, is trying to gauge the underlying demand for its products. Once they get into full production capacity, only then can a realistic estimate be given.
Discussing the path ahead, Upton said they are doing a lot of software work in the existing generation of products. “We try to see our progressions as a series of steps. However, every new generation of the product is not a stepped change. But, what we do is we still pursue pouring sand onto the steps, so that you get a bit of a slope,” remarked Upton. He likens the software investments to the metaphor of a sand, which will give them just enough leverage to make the big move to the next generation.
