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.

Mem
Hierarchy
AMD EPYC 7742
DDR4-3200

(ns @ 3.4GHz)
AMD EPYC 7601
DDR4-2400

(ns @ 3.2GHz)
Intel Xeon 8280
DDR-2666

(ns @ 2.7GHz)
L1 Cache 32KB

4 cycles
1.18ns
32KB

4 cycles
1.25ns
32KB

4 cycles
1.48ns
L2 Cache 512KB

13 cycles
3.86ns
512KB

12 cycles
3.76ns
1024KB

14 cycles
5.18ns
L3 Cache 16MB / CCX (4C)
256MB Total

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

 
38.5MB / (28C)
Shared

~46 cycles (avg)
~17.5ns
DRAM

128MB Full Random
~122ns (NPS1)

~113ns (NPS4)

~116ns

~89ns
DRAM

512MB Full Random
~134ns (NPS1)

~125ns (NPS4)
 
~109ns

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
Comments Locked

180 Comments

View All Comments

  • Zoolook - Saturday, August 10, 2019 - link

    It's been a pretty good investment for me, bought at 8$ two years ago, seems like I'll keep it for a while longer.
  • CheapSushi - Wednesday, August 7, 2019 - link

    It's glorious...one might say.... even EPYC.
  • abufrejoval - Wednesday, August 7, 2019 - link

    Hard to believe a 64 core CPU can be had for the price of a used middle class car or the price of four GTX 2080ti.

    Of course once you add 2TB of RAM and as many PCIe 4 SSDs as those lanes will feed, it no longer feels that affordable.

    There is a lot of clouds still running ancient Sandy/Ivy Bridge and Haswell CPUs: I guess replacing those will eat quite a lot of chips.

    And to think that it's the very same 8-core part that powers the engire range: That stroke of simplicity and genius took so many years of planning ahead and staying on track during times when AMD was really not doing well. Almost makes you believe that corporations owned by share holders can actually sometimes actually execute a strategy, without Facebook type voting rights.

    Raising my coffee mug in a salute!
  • schujj07 - Thursday, August 8, 2019 - link

    Sandy Bridge maxed out at 8c/16t.
    Ivy Bridge maxed out at 15c/30t.
    Haswell maxed out at 18c/36t.
    That means that a single socket Epyc 64c/128t can give you more CPU cores than a quad socket Sandy Bridge (32c/64t) or Ivy Bridge (60c/120t) and only a few less cores that a quad socket Haswell (72c/144t).
  • Eris_Floralia - Wednesday, August 7, 2019 - link

    This is what we've all been waiting for!
  • Eris_Floralia - Wednesday, August 7, 2019 - link

    Thank you for all the work!
  • quorm - Wednesday, August 7, 2019 - link

    Given the range of configurations and prices here, I don't see much room for threadripper. Maybe 16 - 32 cores with higher clock speeds? Really wondering what a new threadripper can bring to the table.
  • willis936 - Wednesday, August 7, 2019 - link

    A reduced feature set and lower prices, namely.
  • quorm - Wednesday, August 7, 2019 - link

    Reduced in what way, though? I'm assuming threadripper will be 4 chiplets, 64 pcie lanes, single socket only. All ryzen support ecc.

    So, what can it offer? At 32 cores, 8 channel memory becomes useful for a lot of workloads. Seems like a lot of professionals would just choose epyc this time. On the other end, I don't think any gamers need more than a 3900x/3950x. Is threadripper just going to be for bragging rights?
  • quorm - Wednesday, August 7, 2019 - link

    Sorry, forgot to add, 3950x is $750, epyc 7302p is $825. Where is threadripper going to fit?

Log in

Don't have an account? Sign up now