Page 51 - EE Times Europe Magazine | April2019
P. 51

EE|Times EUROPE — Boards & Solutions Insert   51

                                                            Adapting the Microcontroller for AI in the Endpoint


          With viable competitors to Arm now out
        there, including the up-and-coming
        instruction-set architecture offered by the
        RISC-V foundation, why did Eta Compute
        choose to use an Arm core for ultra-low-
        power machine-learning acceleration?
          “The simple answer is that the ecosystem for   The architecture of
        Arm is just so well-developed,” Eta Compute   GreenWaves’ GAP9
        CEO Ted Tewksbury told EE Times Europe.   ultra-low-power AI
        “It’s just much easier to go to production [with   chip now uses 10
        Arm] than it is with RISC-V right now. That   RISC-V cores.
        situation could change in the future … RISC-V   (Image: GreenWaves)
        has its own set of advantages; certainly, it’s
        good for the Chinese market. But we’re looking
        primarily at domestic and European markets   That capability, in the right hands, may   impossible or it was going to cost you a
        right now with the ecosystem for [our device].”  also offer the opportunity to reduce power   fortune. And the fortune it was going to cost
          Tewksbury noted that the major challenge   consumption.               you was essentially your investor’s money
        facing the AIoT is the breadth and diversity   Eta Compute can’t take advantage of it just   going to a different company, and that is very
        of the applications. The market is rather   yet, because the new policy does not apply   difficult to justify.”
        fragmented, with many relatively niche   retroactively to existing Arm cores, so it is not   GreenWaves’ custom extensions alone give
        applications commanding only low volumes.   applicable to the M3 core that Eta is using.   its cores a 3.6× improvement in energy con-
        Altogether, however, this sector potentially   But could Tewksbury see Eta Compute using   sumption over unmodified RISC-V cores. But
        extends to billions of devices.     Arm custom instructions in future product   Croome also said that RISC-V has fundamen-
          “The challenge for developers is that they   generations to cut power consumption fur-  tal technical benefits that are simply due to
        cannot afford to invest the time and the   ther? “Absolutely, yes,” he said.  its being new.
        money in developing customized solutions                                  “It’s a very clean, modern instruction set;
        for each one of those use cases,” Tewksbury   ALTERNATIVE ISAs          it doesn’t have any baggage,” he said. “So
        said. “That’s where flexibility and ease of use   RISC-V has been getting a lot of attention this   from an implementation perspective, the
        become absolutely paramount. And that’s   year. The open-source ISA allows the design   RISC-V core is actually a simpler structure,
        another reason why we chose Arm — because   of processors without a license fee, whereas   and simple means less power.”
        the ecosystem is there, the tools are there,   designs based on the RISC-V ISA can be pro-  Croome also cited control as an important
        and it’s easy for customers to develop prod-  tected as with any other type of IP. Designers   factor. The GAP8 device has eight cores in its
        ucts quickly and get them to market quickly   can pick and choose which extensions to add,   compute cluster, and GreenWaves needs very
        without a lot of customization.”    including their own customized extensions.  fine, detailed control over the core execution
          After keeping its ISA under lock and key for   French startup GreenWaves is one of sev-  to allow maximum power efficiency. RISC-V
        decades, Arm finally announced in October   eral companies using RISC-V cores to target   enables that, he said.
        that it would allow customers to build their   the ultra-low–power machine-learning space.   “In the end, if we could have done all of
        own custom instructions for handling spe-  Its devices, GAP8 and GAP9, use eight- and   that with Arm, we would have done all of that
        cialist workloads such as machine learning.   nine-core compute clusters, respectively. Each   with Arm; it would have been a much more
                                                          device also has an addi-  logical choice … because no one ever got fired
                                                          tional core that handles   for buying Arm,” he joked. “The software tools
                                                          control functions.    are there to a level of maturity which is far
                                                           Martin Croome, vice   higher than RISC-V … but, that said, there’s
                                                          president of business   now so much focus on RISC-V that those tools
                                                          development at Green-  are increasing in maturity very fast.”
                                                          Waves, explained to     In summary, while some see Arm’s hold
                                                          EE Times Europe why    on the microprocessor market as weakening,
                                                          the company uses RISC-V   in part because of increased competition
                                                          cores.                from RISC-V, the company is responding by
                                                           “The first reason is   allowing some customized extensions and
                                                          RISC-V gives us the ability   developing new cores designed for machine
                                                          to customize the cores at   learning from the outset.
                                                          the instruction-set level,   In fact, there are both Arm and non-Arm
                                                          which we use heavily,”   devices coming to the market for ultra-low-
                                                          said Croome, adding that   power machine-learning applications. As
                                                          the custom extensions are   the TinyML community continues to work
                                                          used to reduce the power   on reducing neural network model size
                                                          of both machine-learning   and developing dedicated frameworks and
                                                          and signal-processing   tools, this sector will blossom into a healthy
                                                          workloads. “When the   application area that will support a variety of
                                                          company was formed, if   device types. ■
        Eta Compute’s ECM3532 uses an Arm Cortex-M3 core plus   you wanted to do that
        an NXP CoolFlux DSP core. The machine-learning workload   with any other processor   Sally Ward-Foxton is a staff correspondent
        can be handled by either or both cores. (Image: Eta Compute)  architecture, it was either   at AspenCore.

                                                                                       www.eetimes.eu | APRIL 2020
   46   47   48   49   50   51   52   53   54   55   56