5 ms·

Apple chips are fast mostly because they have a lot of cache to spare. Take for example the A12Z. It has 8MB of L2 (not L3, it's L2!) cache. An Intel Core i7-1

Apple chips are fast mostly because they have a lot of cache to spare. Take for example the A12Z. It has 8MB of L2 (not L3, it's L2!) cache. An Intel Core i7-1068NG7 present in the latest Macbook Pros (that performs akin to the A12Z according to Geekbench) has only 2MB of L2 cache. No other ARM CPU has this level of L2 cache. Apple chips are not "magical", Apple just can afford packing up lots and lots of cache because they are not in the silicon "race to the bottom" like Qualcomm and Intel are, for example. L2 cache is very expensive and Apple is just hacking its way up by packing as much L2 cache as they can. Don't take me wrong, it's not that Apple is "right" or "wrong" by doing it. They just can and did it. However, it's needless to say that their CPUs are not so different from other ARM ones, they just happen to have a budget and a business model that let's them ignore the price/performance ratio when designing chips in order to achieve the maximum performance.

kllrnohj 6y ago
> Take for example the A12Z. It has 8MB of L2 (not L3, it's L2!) It's not really that clear cut. You could also argue that the A12Z has 8MB of L3 and 0MB of L2. The L2 in the A12Z is shared while the L2 in the Intel CPU is not. Similarly the latency of the A12Z's L2 is a lot higher than the latency of Intel's L2, but also then still lower than Intel's L3. https://images.anandtech.com/doci/13661/A12X-lat.png https://images.anandtech.com/doci/13661/A12X-lat.png & https://images.anandtech.com/doci/14664/ICLlat.png https://images.anandtech.com/doci/14664/ICLlat.png So it's not "traditional" L2 as you're familiar with it, it's more like an L2.5 or something. Although still accurately called L2 as it is the second level of cache, it's just that Apple went with a rather different cache hierarchy & latency structure than Intel did. But it's really not at all accurate to compare the 8MB of L2 on the A12Z to the 2MB of L2 on Ice Lake. Those are very different caches.
- leonidasv 6y ago
  Latency increases as cache size increases. That's well a well known fact.
hn_check 6y ago
Why do you quote magical as if someone said that. Further, your point is that Apple chips are faster because of easily replicated reasons. Intel is charging hundreds of dollars for their chip -- you should charge them some consulting fees and tell them how easy this is to boost their performance! This isn't even considering that your whole analysis is flawed to begin with and you're comparing apples and oranges.
- imtringued 6y ago
  I think you misunderstood the comment. Intel CPUs are already performing well and it is Apple that is using the same strategy to reach Intel level performance (or even go slightly above it). What you missed is the fact that other SoC vendors like Qualcomm don't follow this strategy and therefore end up with cheaper but also lower performance SoCs. Since Intel doesn't manufacture ARM chips, you are now forced to go with Apple if you want good performance from an ARM chip.
  - hn_check 6y ago
    The comment was very literal in comparing the cache on the Apple chip to the Intel chip. And just to be clear, cache is "expensive" in die size. They aren't putting an order in for L2 cache to Samsung or something. That Apple chip has a die size less than half the size of Intel chips than it outperforms. So the whole "expensive" claim is debunked before it even gets started. Further we are very explicitly comparing Apple silicon to Intel because that is exactly the transition that's happening here.
    - ece 6y ago
      The Apple chip doesn't have all big cores and only 6 total cores in the A12, Intel or AMD (CCX) have 8 big cores with SMT. Apple's multithread performance is accordingly slower. Once you account for these two big omissions, you'll likely find Apple take as much or more die area than AMD's Zen2 CCX for similar MT performance. It'll probably be more die area for equivalent performance, which for Apple might not be an issue given it's margins. Of all the ARM designs we've seen, cache is by far the unique factor in Apple's design, so comparing die size with equivalent cores+features makes complete sense. Like others have mentioned, maybe Apple will just focus on implementing new instructions, but at that point, they will likely diverge enough from the ARM ecosystem that developers and users should be somewhat worried.
azinman2 6y ago
In college my comp. architecture prof. used to say 'cache is king'.

kllrnohj 6y ago

> Take for example the A12Z. It has 8MB of L2 (not L3, it's L2!) It's not really that clear cut. You could also argue that the A12Z has 8MB of L3 and 0MB of L2. The L2 in the A12Z is shared while the L2 in the Intel CPU is not. Similarly the latency of the A12Z's L2 is a lot higher than the latency of Intel's L2, but also then still lower than Intel's L3. https://images.anandtech.com/doci/13661/A12X-lat.png https://images.anandtech.com/doci/13661/A12X-lat.png & https://images.anandtech.com/doci/14664/ICLlat.png https://images.anandtech.com/doci/14664/ICLlat.png So it's not "traditional" L2 as you're familiar with it, it's more like an L2.5 or something. Although still accurately called L2 as it is the second level of cache, it's just that Apple went with a rather different cache hierarchy & latency structure than Intel did. But it's really not at all accurate to compare the 8MB of L2 on the A12Z to the 2MB of L2 on Ice Lake. Those are very different caches.

leonidasv 6y ago
Latency increases as cache size increases. That's well a well known fact.

hn_check 6y ago

Why do you quote magical as if someone said that. Further, your point is that Apple chips are faster because of easily replicated reasons. Intel is charging hundreds of dollars for their chip -- you should charge them some consulting fees and tell them how easy this is to boost their performance! This isn't even considering that your whole analysis is flawed to begin with and you're comparing apples and oranges.

imtringued 6y ago
I think you misunderstood the comment. Intel CPUs are already performing well and it is Apple that is using the same strategy to reach Intel level performance (or even go slightly above it). What you missed is the fact that other SoC vendors like Qualcomm don't follow this strategy and therefore end up with cheaper but also lower performance SoCs. Since Intel doesn't manufacture ARM chips, you are now forced to go with Apple if you want good performance from an ARM chip.
- hn_check 6y ago
  The comment was very literal in comparing the cache on the Apple chip to the Intel chip. And just to be clear, cache is "expensive" in die size. They aren't putting an order in for L2 cache to Samsung or something. That Apple chip has a die size less than half the size of Intel chips than it outperforms. So the whole "expensive" claim is debunked before it even gets started. Further we are very explicitly comparing Apple silicon to Intel because that is exactly the transition that's happening here.
  - ece 6y ago
    The Apple chip doesn't have all big cores and only 6 total cores in the A12, Intel or AMD (CCX) have 8 big cores with SMT. Apple's multithread performance is accordingly slower. Once you account for these two big omissions, you'll likely find Apple take as much or more die area than AMD's Zen2 CCX for similar MT performance. It'll probably be more die area for equivalent performance, which for Apple might not be an issue given it's margins. Of all the ARM designs we've seen, cache is by far the unique factor in Apple's design, so comparing die size with equivalent cores+features makes complete sense. Like others have mentioned, maybe Apple will just focus on implementing new instructions, but at that point, they will likely diverge enough from the ARM ecosystem that developers and users should be somewhat worried.

azinman2 6y ago

In college my comp. architecture prof. used to say 'cache is king'.