The Raspberry Pi Pico was launched in January, and whereas this modest US$4 microcontroller did not make massive waves within the PC group it was an vital first – because the debut platform of the primary ‘Pi Silicon’ CPU, the RP2040. Obviously at that worth level it is a low energy processor (40nm dual-core Arm Cortex-M0+ at 133MHz) with restricted scope, however solely able to the microcontroller duties for which will probably be purposed. The Raspberry Pi Pico is reportedly promoting very nicely with 1 / 4 million models offered and three quarters of one million on backorder.
It is just some weeks later and Raspberry Pi co-founder Eben Upton has already began to talk about what may be subsequent for Pi Silicon – and the reply is a big uplift in efficiency for machine studying duties. Upton and his in-house ASIC growth staff reckon it is extremely worthwhile now to pursue the event of a light-weight accelerator for use for extremely low energy machine studying purposes.
Above is the state of play evaluating the present (dangerous) ML efficiency of RP2040 powered boards with rivals. In explicit, you may see the Pico can full a ‘detect an individual’ process utilizing ML processing in 2,200ms. That result’s a bit sluggish in comparison with rivals. In the longer term, inside 6 months or so in accordance with Upton, the goal is for ML accelerators to make it into Pi Silicon to finish this detection work in 200ms or much less.
The key forward-looking slide from Upton’s presentation is reproduced above and exhibits the aggressive setting for an upcoming Pi Silicon primarily based board. The ultimate bullet level outlines how the second Pi Silicon processor may be made up. It shall be designed as a ‘light-weight accelerator’ and have 4 to eight multiply-accumulates (MACs) per clock cycle, in accordance with this data.