Memory Subsystem: Latency

AMD chose to share a core design among mobile, desktop and server for scalability and economic reasons. The Core Complex (CCX) is still used in Rome like it was in the previous generation. 

What has changed is that each CCX communicates with the central IO hub, instead of four dies communicating in 4 node NUMA layout (This option is still available to use via the NPS4 switch, keeping each CCD local to its quadrant of the sIOD as well as those local memory controllers, avoiding hops between sIOD quadrants which encour a slight latency penalty). So as the performance of modern CPUs depends heavily on the cache subsystem, we were more than curious what kind of latency a server thread would see as it accesses more and more pages in the cache hierarchy. 

We're using our own in-house latency test. In particular what we're interested in publishing is the estimated structural latency of the processors, meaning we're trying to account for TLB misses and disregard them in these numbers, except for the DRAM latencies where latency measurements get a bit more complex between platforms, and we revert to full random figures.


(ns @ 3.4GHz)

(ns @ 3.2GHz)
Intel Xeon 8280

(ns @ 2.7GHz)
L1 Cache 32KB

4 cycles

4 cycles

4 cycles
L2 Cache 512KB

13 cycles

12 cycles

14 cycles
L3 Cache 16MB / CCX (4C)
256MB Total

~34 cycles (avg)
~10.27 ns
16MB / CCX (4C)
64MB Total

38.5MB / (28C)

~46 cycles (avg)

128MB Full Random
~122ns (NPS1)

~113ns (NPS4)



512MB Full Random
~134ns (NPS1)

~125ns (NPS4)

Update 2019/10/1: We've discovered inaccuracies with our originally published latency numbers, and have subsequently updated the article with more representative figures with a new testing tool.

Things get really interesting when starting to look at cache depths beyond the L2. Naturally Intel here this happens at 1MB while for AMD this is after 512KB, however AMD’s L2 has a speed advantage over Intel’s larger cache.

Where AMD has an ever more clearer speed advantage is in the L3 caches that are clearly significantly faster than Intel’s chips. The big difference here is that AMD’s L3’s here are only local to a CCX of 4 cores – for the EPYC 7742 this is now doubled to 16MB up from 8MB on the 7601.

Currently this is a two-edged sword for the AMD platforms: On one hand, the EPYC processors have significantly more total cache, coming in at a whopping 256MB for the 7742, quadruple the amount over the 64MB of the 7601, and a lot more than Intel’s platforms, which come in at 38.5MB for the Xeon 8180, 8176, 8280, and a larger 55MB for the Xeon E5-2699 v4.

The disadvantage for AMD is that while they have more cache, the EPYC 7742 rather consist of 16 CCX which all have a very fast 16 MB L3. Although the 64 cores are one big NUMA node now, the 64-core chip is basically 16x 4 cores, each with 16 MB L3-caches. Once you get beyond that 16 MB cache, the prefetchers can soften the blow, but you will be accessing the main DRAM.

A little bit weird is the fact that accessing data that resides at the same die (CCD) but is not within the same CCX is just as slow as accessing data is on a totally different die. This is because regardless of where the other CCX is, whether it is nearby on the same die or on the other side of the chip, the data access still has to go through the IF to the IO die and back again.

Is that necessarily a bad thing? The answer: most of the time it is not. First of all, in most applications only a low percentage of accesses must be answered by the L3 cache. Secondly, each core on the CCX has no less than 4 MB of L3 available, which is far more than the Intel cores have at their disposal (1.375 MB). The prefetchers have a lot more space to make sure that the data is there before it is needed.

But database performance might still suffer somewhat. For example, keeping a large part of the index in the cache improve performance, and especially OLTP accesses tend to quite random. Secondly the relatively slow communication over a central hub slow down synchronization communication. That is a real thing is shown by the fact that Intel states that the OLTP hammerDB runs 60% faster on a 28-core Intel Xeon 8280 than on EPYC 7601. We were not able to check it before the deadline, but it seems reasonable.

But for the vast majority of these high-end CPUs, they will be running many parallel applications, like running microservices, docker containers, virtual machines, map/reducing smaller chunks of data and parallel HPC Jobs. In almost all cases 16 MB L3 for 4 cores is more than enough.

Although come to think of it, when running an 8-core virtual machine there might be small corner cases where performance suffers a (little) bit.

In short, AMD leaves still a bit of performance on table by not using a larger 8-core CCX. We await to see what happens in future platforms.

Memory Subsystem: Bandwidth Latency Part Two: Beating The Prefetchers


View All Comments

  • Dug - Friday, August 9, 2019 - link

    You still have to pay for cores on datacenter. Each datacenter license covers 2 cores with a minimum purchase of 8. So over 8 cores and you are buying more licenses. 64 cores is about $25k Reply
  • MDD1963 - Friday, August 9, 2019 - link

    Windows license (Standard or Datacenter) covers 2 *sockets* for, a total of 16 cores....; if you have more than 2 sockets, you need more licenses...; if you have 2 sockets, filled with 8 core CPUs, you are good with one standard license... If you have 20 total cores, you need a standard license, and a pair of '2 core' add ons... If you have 32 cores, you need 2 full standard licenses.... Reply
  • MDD1963 - Friday, August 9, 2019 - link

    Datacenter is still licensed for 16 cores, with little 2 pack increments available, or, in the case of a 64 core CPU, effectively 4 Datacenter licenses would be required...($6k per 16 cores, or, roughly $24k) Reply
  • deltaFx2 - Friday, August 9, 2019 - link

    @schujj07: Of course I get that. The OP @Pancakes implied that Rome was going to hurt the wallets of buyers using windows server. The implication being this would not happen if they bought Intel. I was questioning those assumptions. How can Rome cost more money for windows licenses unless rome needs more cores to get the same job done or enterprises overprovision Rome (in terms of total cores) vs. Intel. That would make sense if the per-thread performance is worse but it's not. Reply
  • schujj07 - Friday, August 9, 2019 - link

    The problem is Microsoft went to the Oracle model of licensing for Server 2016/19. That means that you have to license EVERY CPU core it can be run on. Even if you create a VM with only 8 cores, those 8 cores won't always be running on the same cores of the CPU. That is where Rome hurts the pockets of people. You would pay $10k/instance of Server Standard on a single dual 64 core host or $65k/host for Server DataCenter on a dual 64 core host. Reply
  • browned - Saturday, August 10, 2019 - link

    We are currently a small MS shop, VMWare with 8 sockets licensed, Windows Datacenter License. 4 Hosts, 2 x 8 core due to Windows Licensing limits. But we are running 120+ majority Windows systems on those hosts.

    I see our future with 4 x 16 core systems, unless our CPU requirements grow, in which case we could look at 3 x 48 or 2 x 64 core or 4 x 24 core and buy another lot of datacenter licenses. Because we already have 64 cores licensed the uplift to 96 or 128 is not something we would worry about.

    We would also get a benefit from only using 2, 3, or 4 of our 8 VMWare socket licenses. We could them implement a better DR system, or use those licenses at another site that currently use Robo licenses.
  • jgraham11 - Thursday, August 8, 2019 - link

    so how does it work with hyper threaded CPUs? And what if the server owner decides to not run Intel Hyperthreading because it is so prone to CPU exploits (most 10 yrs+ old). Does Google still pay for those cores?? Reply
  • ianisiam - Thursday, August 8, 2019 - link

    You only pay for physical cores, not logical. Reply
  • twotwotwo - Thursday, August 8, 2019 - link

    Sort of a fun thing there is that in the past you've had to buy more cores than you need sometimes: lower-end parts that had enough CPU oomph may not support all the RAM or I/O you want, or maybe some feature you wanted was absent or disabled. These seem to let you load up on RAM and I/O at even 8C or 16C (min. 1P or 2P configs).

    Of course, some CPU-bound apps can't take advantage of that, but in the right situation being able to build as lopsided a machine as you want might even help out the folks who pay by the core.
  • azfacea - Wednesday, August 7, 2019 - link

    F Reply

Log in

Don't have an account? Sign up now