Hacker News Reader

Comment by jcdavis | Hacker News Reader

Isn't 8mb of L3 cache a huge amount for a phone with 2 gigs of ram? A $260 Intel i5 6600 only has 6mb

Perhaps by comparison with desktop CPUs, but since the A8 had 4MB of L3, presumably there are good reasons for Apple to include more L3 in the A9. It may well be that Apple has found that a larger L3 cache contributes more to overall performance (or power efficiency) than, say, faster clocks or more RAM.

I suspect the L3 can also be accessed by the GPU since they share the memory management unit block. It probably contributes heavily into better GPU performance, seeing as the GPU in the SoC doesn't have its own dedicated high performance RAM, and the largest complaint I've seen from mobile game developers has been lack of raw memory bandwidth performance (ie, being unable to do stuff they can do anywhere else, even though the GPU itself can easily handle the task).

I suspect this is also why the iPhone 6S switched to a PCI-E SSD for flash instead of an SDIO/eMMC store like most devices use, just so high performance apps (such as games streaming in assets during runtime) don't choke the phone.

http://www.anandtech.com/show/9662/iphone-6s-and-iphone-6s-p...

According to them with an unoptimized benchmark (doesn't do queue depth >1 yet like desktop benchmarks do; required for real SSD benchmarking), Apple basically doubled the performance of the flash IO.

Intel's L3 cache is for CPU only, Apple's is propably shared with the GPU. Also some Intel CPU's have 128MB of "L4" cache to primarily speed up the GPU.

It isn't for CPU only, Haswell and up GPUs can zero-copy their command buffers out of it (thus making a large number of GPU compute situations "free"); they can't usefully use such a small amount of L3 for texture and framebuffer stuff (not like the highly optimized prefetch and caching systems in Radeons and Geforces can; that is sometimes a work of art).

But yes, I highly suspect the bigger L3 on these serves the same purpose as the L4 on Crystalwells.

Yea, I guess the interesting question is why desktop CPUs have relatively so little (I suspect the answer is Intel trying to price differentiate their Xeon chips, but who knows)

Do desktop systems have faster RAM? That would make the cache less important if so. Just a random guess.

Yes, they can.

Modern mobile devices have single or dual channel, DDR3L-800 through 1600. I'm not seeing any relevant mobile devices still shipping with DDR2, although sub-$100 tablets and anything that is a router still do.

Modern desktops and laptops are dual channel and DDR3-1600 and up, or now that DDR4 is a thing, DDR4-1866 or up.

Gaming machines/high performance workstations are DDR3-2133 and up, or with DDR4, 2400 and up is becoming popular.

LGA-2011-based stuff is quad channel, and LGA-2011 stuff can be multiple socket (its a waste not to populate dual socket E5s with at least 8 sticks).

Cache is a way to enhance performance in a way that can come at a relatively modest cost in power if not in chip area. 8T SRAM cells don't use much power statically, increasing size increases wire capacitance but you aren't switching any more transistors than with a 4mb cache save an extra level of muxing.

Perhaps they need it to handle the 12-megapixel photos and 4K video.

Those are basically streaming, I doubt it.

Oh, so when you take photos and record videos, the data doesn't touch the CPU?

It touches the CPU but it doesn't need to be cached. The CPU loads it, computes on it, then sends it on its way. Hence a 8mb cache wouldn't work any better than a 4mb cache or even a 1mb cache. It's the changes to the storage bandwidth that reflect the better camera.

The uncompressed data generally doesn't

And the Image Signal Processor doesn't share the L3 cache?

The A7 broke new ground by having an unprecedented 4mb L3 cache, which is what helped give it such amazing performance. (See Anandtech review of iPad Air/iPhone 5S)

Reply on news.ycombinator.com