Cache and Infinity Fabric

If it hasn’t been hammered in already,  the big change in the cache is the L1 instruction cache which has been reduced from 64 KB to 32 KB, but the associativity has increased from 4-way to 8-way. This change enabled AMD to increase the size of the micro-op cache from 2K entry to 4K entry, and AMD felt that this gave a better performance balance with how modern workloads are evolving.

The L1-D cache is still 32KB 8-way, while the L2 cache is still 512KB 8-way. The L3 cache, which is a non-inclusive cache (compared to the L2 inclusive cache), has now doubled in size to 16 MB per core complex, up from 8 MB. AMD manages its L3 by sharing a 16MB block per CCX, rather than enabling access to any L3 from any core.

Because of the increase in size of the L3, latency has increased slightly. L1 is still 4-cycle, L2 is still 12-cycle, but L3 has increased from ~35 cycle to ~40 cycle (this is a characteristic of larger caches, they end up being slightly slower latency; it’s an interesting trade off to measure). AMD has stated that it has increased the size of the queues handling L1 and L2 misses, although hasn’t elaborated as to how big they now are.

Infinity Fabric

With the move to Zen 2, we also move to the second generation of Infinity Fabric. One of the major updates with IF2 is the support of PCIe 4.0, and thus the increase of the bus width from 256-bit to 512-bit.

Overall efficiency of IF2 has improved 27% according to AMD, leading to a lower power per bit. As we move to more IF links in EPYC, this will become very important as data is transferred from chiplet to IO die.

One of the features of IF2 is that the clock has been decoupled from the main DRAM clock. In Zen and Zen+, the IF frequency was coupled to the DRAM frequency, which led to some interesting scenarios where the memory could go a lot faster but the limitations in the IF meant that they were both limited by the lock-step nature of the clock. For Zen 2, AMD has introduced ratios to the IF2, enabling a 1:1 normal ratio or a 2:1 ratio that reduces the IF2 clock in half.

This ratio should automatically come into play around DDR4-3600 or DDR4-3800, but it does mean that IF2 clock does reduce in half, which has a knock on effect with respect to bandwidth. It should be noted that even if the DRAM frequency is high, having a slower IF frequency will likely limit the raw performance gain from that faster memory. AMD recommends keeping the ratio at a 1:1 around DDR4-3600, and instead optimizing sub-timings at that speed.

Integer Units, Load and Store Conclusions: Platform, SoC, Core
POST A COMMENT

217 Comments

View All Comments

  • Smell This - Sunday, June 16, 2019 - link


    AND ...
    it might be 12- to 16 IF links or, another substrate ?
    Reply
  • Targon - Thursday, June 13, 2019 - link

    Epyc and Ryzen CCX units are TSMC, the true CPU cores. The I/O unit is the only part that comes from Global Foundries, and is probably at TSMC just to satisfy the contracts currently in place. Reply
  • YukaKun - Monday, June 10, 2019 - link

    "Users focused on performance will love the new 16-core Ryzen 9 3950X, while the processor seems nice an efficient at 65W, so it will be interesting so see what happens at lower power."

    Shouldn't that be 105W?

    And great read as usual.

    Cheers!
    Reply
  • jjj - Monday, June 10, 2019 - link

    The big problem with this platform is that ST perf per dollar gains are from zero to minimal, depending on SKU.
    They give us around 20% ST gains (IPC+clocks) but at a cost. Would rather have 10-15% gains for free than to pay for 20%. Pretty much all SKUs need a price drop to become exciting, some about 50$, some a bit less and the 16 cores a lot more.

    Got to wonder about memory BW with the 16 cores. 2 channels with 8 cores is one thing but at 16 cores, it might become a limiting factor here and there.
    Reply
  • Threska - Tuesday, June 11, 2019 - link

    That could be said of any processor. "Yeah, drop the price of whatever it is and we'll love you for it." Improvements cost, just like DVD's costed more than VHS. Reply
  • jjj - Tuesday, June 11, 2019 - link

    In the semi business the entire point is to offer significantly more perf per dollar every year. That's what Moore's Law was, 2x the perf at same price every 2 years. Now progress is slower but consumers aren't getting anything anymore.

    And in pretty much all tech driven areas, products become better every year, even cars. When there is no innovation, it means that the market is dysfunctional. AMD certainly does not innovate here, except on the balance sheet. Innovation means that you get better value and that is missing here. TSMC gives them more perf per dollar, they have additional gains from packaging but those gains do not trickle down to us. At the end of the day even Intel tries to offer 10-15% perf per dollar gains every cycle.
    Reply
  • AlyxSharkBite - Tuesday, June 11, 2019 - link

    That’s not Moore’s Law at all. It stated that the number of transistors would double. Also it’s been dead a while

    Sandy bridge 4c 1.16b
    Coffee lake 4c is 2.1b (can’t compare the 6c or 8c)

    And that’s a lot more than 2 years.
    Reply
  • mode_13h - Tuesday, June 11, 2019 - link

    Yeah, but those two chips occupy different market segments. So, you should compare Sandybridge i7 vs. Coffelake i7. Reply
  • Teutorix - Tuesday, June 11, 2019 - link

    The number of transistors in an IC, not the number of transistors per CPU core. This is an important distinction since a CPU core in Moore's day had very little in it besides registers and an ALU. They didn't integrate FPUs until relatively recently.

    It's about overall transistor density, nothing more. You absolutely can compare an 8c to a 4c chip, because they are both a single IC.

    An 8 core coffee lake chip is 20% smaller than a quad core sandy bridge chip. That's double the CPU cores, double the GPU cores, with probably a massive increase in the transistors/core also.

    Moore's law had a minor slowdown with intel stuck at 14nm but its not dead.
    Reply
  • Wilco1 - Tuesday, June 11, 2019 - link

    Moore's Law is actually accelerating. Just not at Intel. See https://en.wikipedia.org/wiki/Transistor_count - the largest chips now have ~20 Billion transistors, and with 7nm and 5nm it looks like we're getting some more doublings soon. Reply

Log in

Don't have an account? Sign up now