Sunday, June 26, 2005

Why Speedy does not feel the T5 update

Jeff Kivin recently discussed the T5 update in his 1src podcast. There, he said the following:
Speedy does not see any benefit from the T5's recent system update. But, we all know how Speedy does not measure ARM Processor's performance accurately!

Indeed, this looks true at the first glance. But, if one truly understands what a benchmark does, it all becomes wrong. Lets go back in history a bit to clear this up.
When the first XSCALE PocketPC's debutd, they had CPU benchmarks that were off the roof. However, most applications ran way slower on them than they did on the StrongARM machines because Windows CE was not optimized for the new CPU architecture at that time.
So, most benchmarks measure the capabilities of hardware. They usually contain a few loops that have very little to do with the OS. They essentially stress-test the processor and other hardware for their native capability-this is, what they can do physically!
But applications need more than that! They usually interfere with the operating system, and thus put an extra layer betwen them and the CPU/hardware. This abstraction layer costs extra performance.
So, when you update the abstraction layer to make it faster, the hardware benchmark does not see any benefits. However, applications can get way faster-just like what T5 users are experiencing right now! The data manager routines were made faster, but how should this affect the system bus, the processor speed, ...?
Do you have an explanation?


Post a Comment

<< Home