-
Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!
CPU cache size not so important today?
- Thread starter User12
- Start date
- Joined
- Mar 17, 2002
The cache size continues to grow in new generations of processors, I would argue that the chip makers would not be adding additional cache if they did not think it had a benefit. Especially considering the silicon real-estate the larger L2 and L3 caches take up.
- Joined
- Nov 3, 2008
In short the cache is why the CPU performs as it does. Look at the cacheless AMD processors. They are dependent on low timings.
So CPU cache sizes affect performance more today than ever.
Small or no or ineffective cache put everything off on memory. Good cache and no matter the memory you get good performance.
- Joined
- Oct 21, 2012
- Location
- Phoenix, Az
I seem to recall Tom's concluding that cache had a lot less to do with CPU performance then it used to. That mega quantities of cache didn't seem to contribute anything to general performance... that there were other bottlenecks in chip construction these days. (which i know was not the case 5-7 years ago, when cache had a lot to do with chip performance)
I remember a separate article that pointed out how it was "odd" that the A10-5800k didn't seem to be inhibited by it's lack of L3 cache vs a similar built FX chip. Proving, at least with AMD, the bottleneck in the system was somewhere else.
though considering how dependent trinity chips are on ram timings, that's probably where the lack of the L3 cache shows up. I think we probably can point to cache being the culprit behind why ram timings mean so little these days for most chips... but that the general boost it gives is likely peaked at the moment till they clear whatever is bottlenecking the chips now.
I remember a separate article that pointed out how it was "odd" that the A10-5800k didn't seem to be inhibited by it's lack of L3 cache vs a similar built FX chip. Proving, at least with AMD, the bottleneck in the system was somewhere else.
though considering how dependent trinity chips are on ram timings, that's probably where the lack of the L3 cache shows up. I think we probably can point to cache being the culprit behind why ram timings mean so little these days for most chips... but that the general boost it gives is likely peaked at the moment till they clear whatever is bottlenecking the chips now.
- Joined
- Nov 3, 2008
Exactly. The AMD chips seem to be lacking in predict flush efficiency. Many times the cacheless CPUs do better than the L3 processors.
- Joined
- Jul 24, 2002
- Location
- Seattle, WA
Even though memory access is faster it's still around 10x slower than cache access, depending on what caching layer we're talking about. It's unlikely we'll ever do away with a layer; in fact, we keep adding more. L1, L2, L3, main memory, now SSD, then spinning disk... There's an advantage to layered stores like that.
Caches being bigger doesn't have much effect so long as the application has enough cache to do what it wants. It's like memory in that (currently) if you have around 8GB you're set for most applications and going to 16GB or 32GB won't bring a measurable performance benefit. More memory is useless until you find a use for it.
Everything's great until you start swapping to disk, and then life sucks. Similarly, you don't get a benefit from more cache until you fill it up and have to go to RAM.
Cache sizes probably keep increasing to stay ahead of the curve and make sure you don't get into a worst-case performance scenario.
Caches being bigger doesn't have much effect so long as the application has enough cache to do what it wants. It's like memory in that (currently) if you have around 8GB you're set for most applications and going to 16GB or 32GB won't bring a measurable performance benefit. More memory is useless until you find a use for it.
Everything's great until you start swapping to disk, and then life sucks. Similarly, you don't get a benefit from more cache until you fill it up and have to go to RAM.Cache sizes probably keep increasing to stay ahead of the curve and make sure you don't get into a worst-case performance scenario.
- Joined
- Jun 6, 2002
johan if you look at intel cpus since core 2, L2 has gone down alot. the celeron based core 2 cpus had atleast 512k L2 vs every celeron and higher i series having only 256kb. the only thing on these newer cpus from intel that is increasing is L3. i need to loook but i think off the top of my head the L1 from core 2 to i series has stayed the same.
looking at the older core 2's you can see the increase of core 2 making a difference in performance. the doubling of L2 increased performance per clock in the apps i talked about above. once we got to to a certian size you could see how it wanst worth adding more. on current i's i would gladly give up 3mb on a 6mb L3 for doubling the L2 on the cpus.
looking at the older core 2's you can see the increase of core 2 making a difference in performance. the doubling of L2 increased performance per clock in the apps i talked about above. once we got to to a certian size you could see how it wanst worth adding more. on current i's i would gladly give up 3mb on a 6mb L3 for doubling the L2 on the cpus.
There is big difference in some tests between 2MB L3 and 3MB L3 so between Pentiums/i3 and Celerons. There is almost no difference between 3570K and 3770K where is 6MB vs 8MB. So cache is for sure helping in performance but there is some limit at least for desktops. I don't think that 8MB is helping AMD FX series at all
SB/IB memory controllers don't need really tight memory timings or really high clocks to perform good just because of large and fast cache.
There is some reason why server cpus have 15-20MB+ L3. Probably there is much bigger difference in multithreading the same as you almost won't see difference between 2 and 4 memory channels in desktops but in servers there is up to 50% higher bandwidth.
SB/IB memory controllers don't need really tight memory timings or really high clocks to perform good just because of large and fast cache.
There is some reason why server cpus have 15-20MB+ L3. Probably there is much bigger difference in multithreading the same as you almost won't see difference between 2 and 4 memory channels in desktops but in servers there is up to 50% higher bandwidth.
Last edited:
