Archive 1997–2024

Intel's Centrino CPU (Pentium-M): Revolutionizing the Mobile World

Banias' Caches

We mentioned at last year's Fall IDF that Banias would be a 0.13-micron processor composed of no less than 77 million transistors. Almost half of the Banias die is reserved for cache and as you can expect, the cache is nothing normal.

Banias is outfitted with a 64KB L1 cache, twice the size of the L1 found on the Pentium III. The L1 is split into a 32KB data cache and a 32KB instruction cache; as we mentioned before, the gate-hungry trace cache was rejected as an option for a mobile CPU although it would have reduced branch mis-predict penalties.

What's most impressive about Banias' caches is its 1MB on-die L2 cache. Not only did the Israel design team carefully design the transistors that make up the gates within the cache to maximize power savings, but they also changed the way data is actually accessed within the cache. Banias employs an 8-way set associative L2 cache, and normally in such a cache when one of the "ways" is selected, the entire block is selected. With Banias, each "way" is further split into quadrants so that when one is selected, a separate multiplexer selects which quadrant the necessary data is located in and only activates that part of the cache. The end result is that much less of the power hungry L2 cache is consuming battery life, which helps extend that battery life significantly.

Out of all of the transistors in a CPU, those dedicated for cache are generally the most power hungry, and thus keeping them inactive as much as possible helps tremendously in reducing power consumption. Of course the downside to this approach is that cache accesses now take a little longer, but the power savings are significant enough to make up for the performance deficit by going to a larger cache - which is why Banias is able to have such a large cache and still be relatively low on the power consumption scale.

The higher latency of the cache is masked by the fact that a 1MB L2 cache improves overall system performance enough to hide the majority of the latency hit.

Originally published at https://www.anandtech.com/show/1083 (Wayback Machine).