3 ms·
Well there still is a case for the AVR -- you get zero pipelining effects, no flash latency etc; it's like a tiny, slow, scalar CPLD for all intent and purpose.
by buserror 11y ago
Well there still is a case for the AVR -- you get zero pipelining effects, no flash latency etc; it's like a tiny, slow, scalar CPLD for all intent and purpose.
Cortex can't beat that, you can tight loop an AVR to cycle precision and get pretty cool return for your running frequency; the Cortex are nowhere as linear. [0]
Altho, I agree that the bigger parts are useless... xmegas and such have absolutely no reason to exist IMO.
On the other hand the CM3 and especially M4's are awesome beast; and the parts like the STM32's are just peerless in terms of features -- they are just borderline scary.
On the other hand, I was worried a little while back that there was nothing between these trusted 8 bits cores and the ARMs -- but nowadays there is; Espressif esp8266 is pretty cool, but the next one in the pipeline is awesome [1]; 2 (well 3 but) cores, and more peripherals than you can shake a stick at. Thats the first SoC that for me can give the STM32 (and NXP equivalent) a good run for their money (and win, in my opinion).
Dislaimer:
[0] I know the AVR very well, I'm the author of simavr: http://github.com/buserror/simavr http://github.com/buserror/simavr -- [1] I'm also one of the beta tester for the ESP31/32 new chipset.
- revelation 11y agoPeople care about time, not cycles, and chances are with whatever crystal is running your AVR, you will still not be comfortable despite the lack of post-1990 CPU features in ATmegas. And finally, the idea of tight loops engineered to cycle precision is dying a slow death in general. You just smother the hard timing requirements with excess frequency and performance. Curious thing about the ESP8266: it doesn't use ARM or any of the other popular kids on the block but a Tensilica Xtensa architecture.