Cache and TLB Updates

One of the biggest changes in the new Sunny Cove core is the cache hierarchy. Throughout most of the last decade, Intel has kept the same cache configuration among its server cores with very few changes, aside from improving the micro-op cache through Sandy Bridge today, but with Sunny Cove, we get some key updates.

From our microarchitecture disclosure article, we know that Sunny Cove brings with it an increased L1 data cache, as well as a doubled L2 cache, from Skylake. With this comes more associativity as well.

Core Cache Comparison
Sunny
Cove
AnandTech Cannon
Lake
Skylake Sandy
Bridge
  AMD
Zen 2
48 KB
12-way
L1-D 32 KB
8-way
32 KB
8-way
32 KB
8-way
  32 KB
8-way
32 KB
8-way
L1-I 32 KB
8-way
32 KB
8-way
32 KB
8-way
  32 KB
8-way
512 KB
8-way
L2 256 KB
4-way
256 KB
4-way
256 KB
8-way
  512 KB
8-way
2 MB
16-way
L3/core 2 MB
16-way
2 MB
16-way
2MB
16-way
  4 MB
16-way
2304 uOp Cache 1536 1536 1536   4096

I’ve also included the size of the micro-op cache in this comparison, because we understand that the L1-I and the micro-op cache are typically located next to each other in the die design – increasing the size of one typically means limiting the size of the other, however given that both of them work to accelerate the same types of data, it is important to find the right balance: AMD for Zen 2 decided to halve the L1-I and double its associativity, and then double the micro-op cache from 2k to 4k entries. For Sunny Cove, Intel has increased the L1-D by 50%, but also increased the micro-op cache by 50%, taking advantage of the second generation 10nm process.

In the above latency graph comparison, we can notably see the new increased cache capacities. The L1’s increase from 32KB to 48KB is well visible, as well as the L2 increase from 256KB to 512KB.

On the part of the L2 it’s interesting to see that again the new core’s TLB structures don’t quite cover the whole capacity of the L2 and that’s why the latter half of the L2 has worse access latencies for L1D TLB miss patterns.

L3 latencies look similar, however we’ll dwell into that just in a bit. The DRAM latencies of course are apples and oranges in this case as Intel’s new LPPDR4X memory controller in the new ICL part doesn’t have a counter-part we can compare to, but as expected the memory latency is notably worse than a desktop part so no big surprises there.

What’s rather disappointing however is that we’re not really seeing any kind of change to the prefetchers and our access patterns between the two microarchitectures look identical. We had expected Intel to improve in this regard, but alas it seems, at least for our test cases, there’s been no improvements.

One of the major balancing acts with cache sizes is also cache latency. The bigger the cache, the longer it takes to retrieve data from it – increasing associativity can do this as well. Intel for the longest time has cited a 4-cycle latency to its L1 cache, and a 12-cycle latency to its L2 cache. This changes, in quite a substantial way.

Core Cache Latency (in cycles)
Sunny Cove AnandTech Cannon
Lake
Skylake Sandy
Bridge
  AMD
Zen 2
5 L1 4 4 4   4
13 L2 12 ~12 12   12
30-36 L3   26-37     34

Due to the increase in the L1 cache, that 4-cycle latency is now a 5-cycle latency. Intel is hoping that its double L1 store bandwidth will help hide that latency, however adding 20% latency onto every data fetch is a heavy hit to take. This is why a lot of the rest of the core is also beefed up, so work can still continue to cover for this extra cycle. It is worth noting that Intel’s Atom cores in the past had an L1 latency of 3 cycles, however they have also since moved to 4-cycle.

The L2 also changes, from 12 cycles to 13 cycles. You can’t double a cache for free, and it is interesting to note here that now AMD and Intel share the same L2 cache size and the same associativity, however AMD has a cycle advantage on Intel. This has a knock on effect for L3 cache latency too, which has increased from ~26-37 cycles to ~30-36 cycles on Intel, from Skylake. The latency here isn’t uniform because the cache isn’t uniform, and accessing non-local L3 slices comes with a latency penalty, as say opposed to AMD’s uniform latency L3.

One of the biggest additions Intel was keen to make in our Sunny Cove microarchitecture disclosure is the increase in the L1 bandwidth. Based on our testing, Skylake’s L1 data cache was capable of 2x32-byte read and 1x32-byte write per clock. For Sunny Cove this has increased, but it gets a bit more complicated. In one clock Sunny cove can perform 2x64-byte reads and either 1x64-byte write, or 1x32-byte write, or 2x16-byte writes. So both reads and writes, at peak, have doubled with bit width, with the write configurations being a bit more complex.

We can see the direct effects of the bandwidth change in our bandwidth tests. The L1 store bandwidth is extremely evident as it measures in at 222GB/s on the ICL part, while only reaching 139GB/s on the 28% higher clocked 9900K. If we normalise for frequency, we indeed see a doubling of the store bandwidth of the new core.

The L2 cache bandwidth looks unchanged from Skylake – the part here lags behind with an equal amount to the frequency difference. Surprisingly, the L3 bandwidth fares quite well here and is almost at parity to the desktop 9900K even though the frequency difference. The DRAM bandwidth is inferior on the ICL mobile chip, but again that’s to be expected due to the apples and oranges DRAM architecture comparison.

Moving on to the TLB configurations, and it turns out that Intel has moved where this data is stored. This data is typically stored in CPUID.02h, however it is now found in CPUID.18h, and the info between different page sizes and capabilities is now given as subdivisions. The TLBs look as follows, and broken down into configurations:

Sunny Cove TLB
AnandTech Type Entries Associativity
L1-D 4K 64 4-way
L1-D 2M 32 4-way
L1-D 1G 8 full
L1-I 4K+2M 8 full
L1-I 4K+2M+1G 16 full
L2 4K+2M 1024 8-way
L2 4K+1G 1024 8-way

This means that for 4K and 2M L1-I entries, there are a total 8+16 = 24 possible, but only 16 1G possible. For 4K L2 entries, there are 1024 + 1024 = 2048 entries possible, however only 1024 each of 2M/1G are possible. The peak numbers align with what Intel has told us in our microarchitecture disclosure, however it is good to dive into what is possible.

Intel 10nm Ice Lake, Sunny Cove, and Gen11 Tested: Details Inside Security Updates, Improved Instruction Performance and AVX-512 Updates
POST A COMMENT

261 Comments

View All Comments

  • andrewaggb - Thursday, August 1, 2019 - link

    I'm glad you did the benchmarks. At this point everybody knows intel has manufacturing problems and that availability will be extremely limited for these parts. It's nice to know they have updated designs in a working state and that even with limited clock speeds they're pretty fast. I'm disappointed intel isn't simultaneously releasing these parts on 10nm and 14++++ but hopefully that's because they are more confident about their future roadmap. Reply
  • eastcoast_pete - Thursday, August 1, 2019 - link

    I think they ran out of space to fit any more "++++" after the 14 nm. Reply
  • StormyParis - Thursday, August 1, 2019 - link

    Don't they have cross-licensing with MS, allowing them to use '#' for 3+ '+' ? Reply
  • brunis.dk - Friday, August 2, 2019 - link

    Tell me why i need to prepare months in advance? It's great you get a paper launch product in hand. But to me it's still a paper launch. What's different here from you just releasing the benchmarks on availability day? If i was in the mood for a new cpu/mobo purchase i would start saving up and buy whatever is best on the day. Agree with brakdoo! Reply
  • Kvaern1 - Saturday, August 3, 2019 - link

    Blink if someone is forcing you to read AT articles and you need help. Reply
  • close - Tuesday, August 6, 2019 - link

    @Kvaern1, your point that people should not criticize. So you can either applaud or move along? If nobody ever calls you out for anything then when do you learn? Pointing out something like a paper launch coming from a manufacturer with a history of deception is no without merit.

    Now whether this applies to this article is a different matter. But your low quality sarcasm response and the "only open for praise" attitude shows some troubling mediocrity.
    Reply
  • eva02langley - Friday, August 2, 2019 - link

    Ian, those companies do things like that for a reason. IMHO, they want to make the public believe that 10nm is a reality, when in fact it is still a fairy tale. Intel knows that the frequency drop, the yield and the price doesn't worth the trouble to switch to 10nm. Sure, they told investors that 10nm exist and they are releasing products to claim this, but in reality, 10nm will be just there so Intel doesn't need to admit to investor that 10nm is a fiasco. If they did, the price of the stock will lose significant value and it would affect the public perception of the company.

    This is what I can take from this whole thing.
    Reply
  • Klimax - Friday, August 2, 2019 - link

    Is sky blue in your world or is it some version of red? You may have had point a year or two back... Reply
  • 0ldman79 - Friday, August 2, 2019 - link

    Uh...

    They reworked the entire 10nm process to get it going.

    This isn't impossible, it is expensive and time consuming.

    They've spent the $$$ and at least a year working on it.

    It is hardly unheard of that the single most successful tech company on the planet figured out a problem.
    Reply
  • close - Tuesday, August 6, 2019 - link

    @0ldman79 "the single most successful tech company" - that's the overstatement of the millennium. And of course I'm no stranger to overstatements. Reply

Log in

Don't have an account? Sign up now