3 ms·
Is your processor a good match for ml workflows, or will GPUs still be the way to go there? Does your system use distributed memory? shared memory? How do you
by K0balt 2y ago
Is your processor a good match for ml workflows, or will GPUs still be the way to go there?
Does your system use distributed memory? shared memory? How do you deal with memory bound tasks?
I’d love to see this released in an RF-SOC like the esp32 or. NRF52840, in my experience, RF capabilities are a requirement for the vast majority of edge applications.
What are we looking at in energy savings over best-in-class Harvard designs like the NRF52840?
- blacklion 2y agoI've thought EMF32 are most efficient, but I've looked up NRF52840 and shocked: 150 uA/MHz fo EMF32 and 30 (!) uA/MHz for Nordic part, if I goggled it right.
- K0balt 2y agoYeah, the Nordic part is basically magic if you are careful. If these guys can do 10x on that, a whole market segment could move to energy harvesting / single use battery and battery life for other segments could double or triple. (Often the uC is not the big power draw)
- picture 2y agoIt's not hard at all to get good uA/MHz numbers, it just increases die size and cost. You can push it pretty far by tuning the synthesis, for example the same Cortex-M4 core on the Ambiq Apollo3 Blue advertises "6 μA/MHz executing from flash or RAM at 3.3 V" These are magnitudes of difference across the Espressif, Nordic, and Ambiq parts. The difference in price is also proportional. https://ambiq.com/apollo3-blue/ https://ambiq.com/apollo3-blue/
- K0balt 2y agoCool, il check that out!
- keyi 2y ago1. Yes we are a great fit for edge ML inference, but we are also great for DSP and other, general-purpose application logic. 2. Our chip has an integrated memory & integrated non-volatile storage that are shared across the tiles in our architecture. 3. We agree, RF applications are very interesting and are a great fit for our architecture in a lot of cases 4. It's hard to say what the energy savings would be versus that specific chip without doing some benchmarking. We have measured 10-100x improvement in energy consumption against several very competitive Arm implementations that we've tested and I'd expect to see a similar advantage.