Improvements to the Cache Hierarchy

The biggest under-the-hood change for the Ryzen 2000-series processors is in the cache latency. AMD is claiming that they were able to knock one-cycle from L1 and L2 caches, several cycles from L3, and better DRAM performance. Because pure core IPC is intimately intertwined with the caches (the size, the latency, the bandwidth), these new numbers are leading AMD to claim that these new processors can offer a +3% IPC gain over the previous generation.

The numbers AMD gives are:

  • 13% Better L1 Latency (1.10ns vs 0.95ns)
  • 34% Better L2 Latency (4.6ns vs 3.0ns)
  • 16% Better L3 Latency (11.0ns vs 9.2ns)
  • 11% Better Memory Latency (74ns vs 66ns at DDR4-3200)
  • Increased DRAM Frequency Support (DDR4-2666 vs DDR4-2933)

It is interesting that in the official slide deck AMD quotes latency measured as time, although in private conversations in our briefing it was discussed in terms of clock cycles. Ultimately latency measured as time can take advantage of other internal enhancements; however a pure engineer prefers to discuss clock cycles.

Naturally we went ahead to test the two aspects of this equation: are the cache metrics actually lower, and do we get an IPC uplift?

Cache Me Ousside, How Bow Dah?

For our testing, we use a memory latency checker over the stride range of the cache hierarchy of a single core. For this test we used the following:

  • Ryzen 7 2700X (Zen+)
  • Ryzen 5 2400G (Zen APU)
  • Ryzen 7 1800X (Zen)
  • Intel Core i7-8700K (Coffee Lake)
  • Intel Core i7-7700K (Kaby Lake)

The most obvious comparison is between the AMD processors. Here we have the Ryzen 7 1800X from the initial launch, the Ryzen 5 2400G APU that pairs Zen cores with Vega graphics, and the new Ryzen 7 2700X processor.

This graph is logarithmic in both axes.

This graph shows that in every phase of the cache design, the newest Ryzen 7 2700X requires fewer core clocks. The biggest difference is on the L2 cache latency, but L3 has a sizeable gain as well. The reason that the L2 gain is so large, especially between the 1800X and 2700X, is an interesting story.

When AMD first launched the Ryzen 7 1800X, the L2 latency was tested and listed at 17 clocks. This was a little high – it turns out that the engineers had intended for the L2 latency to be 12 clocks initially, but run out of time to tune the firmware and layout before sending the design off to be manufactured, leaving 17 cycles as the best compromise based on what the design was capable of and did not cause issues. With Threadripper and the Ryzen APUs, AMD tweaked the design enough to hit an L2 latency of 12 cycles, which was not specifically promoted at the time despite the benefits it provides. Now with the Ryzen 2000-series, AMD has reduced it down further to 11 cycles. We were told that this was due to both the new manufacturing process but also additional tweaks made to ensure signal coherency. In our testing, we actually saw an average L2 latency of 10.4 cycles, down from 16.9 cycles in on the Ryzen 7 1800X.

The L3 difference is a little unexpected: AMD stated a 16% better latency: 11.0 ns to 9.2 ns. We saw a change from 10.7 ns to 8.1 ns, which was a drop from 39 cycles to 30 cycles.

Of course, we could not go without comparing AMD to Intel. This is where it got very interesting. Now the cache configurations between the Ryzen 7 2700X and Core i7-8700K are different:

CPU Cache uArch Comparison
Zen (Ryzen 1000)
Zen+ (Ryzen 2000)
Kaby Lake (Core 7000)
Coffee Lake (Core 8000)
L1-I Size 64 KB/core 32 KB/core
L1-I Assoc 4-way 8-way
L1-D Size 32 KB/core 32 KB/core
L1-D Assoc 8-way 8-way
L2 Size 512 KB/core 256 KB/core
L2 Assoc 8-way 4-way
L3 Size 8 MB/CCX
(2 MB/core)
2 MB/core
L3 Assoc 16-way 16-way
L3 Type Victim Write-back

AMD has a larger L2 cache, however the AMD L3 cache is a non-inclusive victim cache, which means it cannot be pre-fetched into unlike the Intel L3 cache.

This was an unexpected result, but we can see clearly that AMD has a latency timing advantage across the L2 and L3 caches. There is a sizable difference in DRAM, however the core performance metrics are here in the lower caches.

We can expand this out to include the three AMD chips, as well as Intel’s Coffee Lake and Kaby Lake cores.

This is a graph using cycles rather than timing latency: Intel has a small L1 advantage, however the larger L2 caches in AMD’s Zen designs mean that Intel has to hit the higher latency L3 earlier. Intel makes quick work of DRAM cycle latency however.

Talking 12nm and Zen+ Translating to IPC: All This for 3%?


View All Comments

  • prateekprakash - Friday, August 03, 2018 - link

    "We’ll cover these in the next few pages, as well as the results from our testing.

    Where is the overclocking result?
  • kithylin - Tuesday, September 04, 2018 - link

  • JRW - Thursday, December 06, 2018 - link

    Last year I upgraded from a 1st gen i7 920 to i7 8700K and even with spectre & meltdown performance has been amazing, also Asus has been recently updating the motherboard BIOS with further CPU performance improvements. Reply
  • WaltC - Friday, May 31, 2019 - link

    Had to chuckle a bit here reading the rather frightened comments about the "dangers" of a RAID drive disk simply because it has no fault tolerance. It's an incredible error, broaching almost upon the realms of superstition (unfortunately, this is no exaggeration), because all standard HDD disks are bereft of fault tolerance as well--simply putting an HDD in a JBOD configuration has absolutely no effect on the likelihood of the drive failing. That's an old wives' tail that really needs to be put to bed. The lack of fault tolerance is often cited as a legitimate reason to irrationally fear running drives with no fault tolerance protection. That is, of course, until you discover that every single *standard HDD* you have ever owned, including all of your boot drives to date, *also* had/have no fault tolerance protection...;) Ergo, all of those boot HDD's you've owned, about which you've never given a second thought as to their failure, are *exactly and precisely* as safe as any RAID-configured disk mentioned in this article...;) No, sorry, but your chances of drive failure do not and never will increase by 100% (or some other silly stratospheric number) every time you add in a like drive--no--trust me--the odds of drive failure with normal, everyday HDDs, IDE or otherwise (because a drive's format has nothing to do with its likelihood of failure--even remotely), is exactly the same for RAID drives as it is for standard HDDs. The amount of HDDs in a given system has zero--absolute zero--to do with the likelihood of any single drive failing--adding 1,000 drives doesn't increase the likelihood of a *single* drive failing one iota above the likelihood that a single drive, in a single-drive array, will fail. As well, in the one-drive system, versus the 1k-drive JBOD system--when *a* drive fails in either arrangement then you have lost all your data...;) Honestly, this article reads like someone interested in criticizing StoreMI as opposed to an honest evaluation of the product! I'm looking at using it, myself, which is how I got here in the first place. Instead of concentrating on slim-to-none chance remote probabilities of drive failure--how about a nice little article on how well it works, or doesn't work, as the case may be, eh?...;) Reply

Log in

Don't have an account? Sign up now