| ??? 04/03/08 13:07 Read: times |
#152919 - cache lock won't really help Responding to: ???'s previous message |
Erik Malund said:
And, in the 50 and 100MHz derivatives things got a bit worse because of the caching.
.. can be "cache locked" Well, the cache is not unlimited, so the unlocked jumps will execute permanently longer (I never understood fully the explanation SiLabs gave for the cache operation but I'd say the penalty is 4 cycles per cache miss). I mean, only if you have a very very short program where the number of jumps is less than cache positions (where you don't need to lock them then either), the number of cycles - hence consumption per instruction - won't be affected by the fact that the cache needs to be reloaded upon cache miss. Erik Malund said:
I agree that, while 1 clock at 100 MHz sounds like 100 times faster that a 12MHz oldie, the real factor is something like 50. According to my so-often-mentioned mini-benchmark, it's slightly better: the factor is around 66 (without taking into account the cache misses). In other words, in average, there are 50% more instruction cycles per instruction in the singleclockers compared to the "vanilla". We all know the value of benchmarks, but you might have a look and judge yourself how representative that short piece of code might be. Erik Malund said:
It is, indeed a fact that those that write "sales literature" conveniently forget. At least we can judge the seriousness of intentions of the salespeople trying to push such stuff to those, many of whom they can quite safely assume are very well aware of the facts... Erik Malund said:
do you have Keil?, if not what form do you need C source, asm list or what? Preferrably hex, although I do have some demoKeil (but I am not sure I can handle it quickly - I tried it maybe thrice in the last two years). Thanks Jan |
| Topic | Author | Date |
| Low Power | 01/01/70 00:00 | |
| clocks/instruction - some hairsplitting | 01/01/70 00:00 | |
| not really 'quite' | 01/01/70 00:00 | |
| what the #$&* is "most"? | 01/01/70 00:00 | |
| critical code ... | 01/01/70 00:00 | |
| cache lock won't really help | 01/01/70 00:00 | |
| the issue is critical code | 01/01/70 00:00 | |
| a proper simulator would help | 01/01/70 00:00 | |
| In that case I doubt you will be able to buy ... | 01/01/70 00:00 | |
| This was not a "first" responder ... | 01/01/70 00:00 | |
| Challenger | 01/01/70 00:00 | |
| benchmark | 01/01/70 00:00 | |
| Experiment report: Dhrystone results | 01/01/70 00:00 | |
Nice work | 01/01/70 00:00 |