Decode

For the decode stage, the main uptick here is the micro-op cache. By doubling in size from 2K entry to 4K entry, it will hold more decoded operations than before, which means it should experience a lot of reuse. In order to facilitate that use, AMD has increased the dispatch rate from the micro-op cache into the buffers up to 8 fused instructions. Assuming that AMD can bypass its decoders often, this should be a very efficient block of silicon.

What makes the 4K entry more impressive is when we compare it to the competition. In Intel’s Skylake family, the micro-op cache in those cores are only 1.5K entry. Intel increased the size by 50% for Ice Lake to 2.25K, but that core is coming to mobile platforms later this year and perhaps to servers next year. By comparison AMD’s Zen 2 core will cover the gamut from consumer to enterprise. Also at this time we can compare it to Arm’s A77 CPU micro-op cache, which is 1.5K entry, however that cache is Arm’s first micro-op cache design for a core.

The decoders in Zen 2 stay the same, we still have access to four complex decoders (compared to Intel’s 1 complex + 4 simple decoders), and decoded instructions are cached into the micro-op cache as well as dispatched into the micro-op queue.

AMD has also stated that it has improved its micro-op fusion algorithm, although did not go into detail as to how this affects performance. Current micro-op fusion conversion is already pretty good, so it would be interesting to see what AMD have done here. Compared to Zen and Zen+, based on the support for AVX2, it does mean that the decoder doesn’t need to crack an AVX2 instruction into two micro-ops: AVX2 is now a single micro-op through the pipeline.

Going beyond the decoders, the micro-op queue and dispatch can feed six micro-ops per cycle into the schedulers. This is slightly imbalanced however, as AMD has independent integer and floating point schedulers: the integer scheduler can accept six micro-ops per cycle, whereas the floating point scheduler can only accept four. The dispatch can simultaneously send micro-ops to both at the same time however.

Fetch/Prefetch Floating Point
Comments Locked

216 Comments

View All Comments

  • Walkeer - Thursday, June 13, 2019 - link

    Superb analysis, thanks a lot @Ian! very excited to have the 3900x at home already
  • FreckledTrout - Thursday, June 13, 2019 - link

    Reading over the Zen2 microarchitecture article Im left wondering if the Windows scheduler improvements are making use of a new unmentioned RDPID feature in Zen2 to determine where threads are placed?
  • cooker358 - Thursday, June 13, 2019 - link

    感谢分享!
  • Gastec - Thursday, June 13, 2019 - link

    I too am curious about the latencies, particularly between the chiplets. With the clock selection down to 2 ns and Windows' 10 hopefully improved thread allocation (filling a CCX, then the next one before jumping to the 2nd chiplet) latencies should be lower. We'll just have to wait for honest extensive testing and reviews to be done. You were not planning on buying these CPUs on release day or even worse, pre-ordering them, were you? :)
  • jamescox - Sunday, June 16, 2019 - link

    I expect the CCX to CCX latencies to be very good. There is no memory clock on the cpu chiplet, so the two on die CCX almost certainly communicate at cpu clock rather than memory clock as in Zen 1. It isn’t the same as Intel’s mesh network, but AMD’s solution will have better L3 latency within the CCX compared to Intel. Intel’s mesh network seems to be terrible for power consumption. Intel’s ring bus didn’t scale to enough cores. For their 18 core chip (if I am remembering right), they actually had 3 separate ring buses. The mesh network is obviously not workable across multiple chiplets, so it will be interesting to see what Intel does.

    For the chiplet to chiplet latency, they have more than doubled the infinity fabric serdes clock with the higher than PCIe 4.0 speeds. It seems that the internal IF clock is also around doubled. It was operating at actual memory clock in Sen 1 which was half the DDR rate. They seem to be running the internal IF clock the same as the DDR rate with the option to drop back to half DDR rate. So if you are running DDR 3200, the IF clock may actually be 3200 instead of 1600 as it would be in Zen 1. If you re overclocking to DDR 4000 or something, then it may need to drop down to 2000 for the internal IF clock. If this is the way it is set up, then they may have an option to explicitly set the divider, but it is probably going to not be stable past 3.7 GHz or so. The IO die is 14 nm global foundries, so that seems like a reasonable limitation.

    The CCX to CCX latency should be less important as the OS and software is better optimized for the architecture. There was quite a few cases on Zen 1 of applications performing significantly better on Linux compared to windows due to the scheduler. Most applications can be optimized a bit for this architecture also. The problem is fine grained shared memory between threads on different CCX. It generally a good idea to reduce that anyway since locking can be detrimental to performance. With Zen 2, I think application level optimizations are probably going to be a lot less necessary anyway, but a lot of the early issues were probably caused by bad multi-threaded programming. This type of architecture isn’t going away. Intel can’t compete with Epyc 2 with a monolithic die. Epyc 2 will be around 1000 square mm of silicon total. Intel can’t scale core count without moving to something similar.
  • frshi - Friday, June 14, 2019 - link

    @Ian Cutress What about 2x16GB sticks compared to 4x8GB? I remember Zen and Zen+ were kinda picky when using 4 sticks. Any change to that on Zen 2?
  • RAINFIRE - Saturday, June 15, 2019 - link

    Yeah - I'm curious. Can anyone speak to the (4 x 32GB) memory that Ryzen 3000 and x570 boards are supposed to support?
  • Holliday75 - Wednesday, June 19, 2019 - link

    IF reviewers have samples at this time they are under an NDA until July 7th. Only unconfirmed leaks can provide that kind of info and its super early. A lot of these types of issues won't be known until they go retail.
  • AdrianMel - Sunday, June 16, 2019 - link

    I would like these AMD chips to be used on laptops. Would be a breakthrough in terms of computing power, lower consumption. I think if a HBM2 or higher memory is integrated into the processor, I think it will double the computing power. Ar fi de studiat si o implementare a 2 porturi superiare thnic vechiului expresscard 54 in care sa putem introduce in laptopuri 2 placi video
  • jamescox - Sunday, June 16, 2019 - link

    Everyone keeps bringing up HBM for cpus as if it is magical in some manner. HBM can provide high bandwidth, but it is still DRAM. He latency isn’t that great, so it isn’t really that useful as a cpu cache. If you are trying to run AVX512 code across a bunch of CPU cores, then maybe you could use the bandwidth. If you have code that can use that level of parallelism, then it will almost certainly run much more efficiently on an actual gpu. I didn’t think that expanding AVX to 512-bits was a good idea. There isn’t too much difference from a cpu perspective between 1 512-bit instruction and 2 256-bit instructions. The registers are wider, but they can have many more smaller registers that are specified in the ISA by using existing register renaming techniques. At 14 nm, the 512-bit units seem to take too much space and consume too much power. They may be more easily doable in 7 nm or below eventually, but they may still have issues running at cpu core clocks. If you have to run it at half clock (which is about where gpus are vs. cpus) then you have lost the advantage of going double the width anyway. IMO, the AVX 512 instructions were Intel’s failed attempt (Xeon Phi seems to have been a disappointment) at making a cpu act like a gpu. They have basically given that up and are now designing an actual gpu.

    I went off in a bit of a tangent there, but HBM really isn’t that useful for a cpu cache. It isn’t going to be that low of latency; so it would not increase single thread performance much compared to stuff actually designed to be a low latency cache. The next generations form AMD May start using active silicon interposers, but I would doubt that they would use HBM. The interposer is most likely to be used in place of the IO die. They could place all of the large transistors needed for driving off die interfaces (reason why IO doesn’t scale well) in the active interposer. They could then stack 7 nm chips on top of the active interposer for the actual logic. Cache scales very well which is why AMD can do a $200 chip with 32 MB of L3 cache and a $500 chip with 64 MB of L3. Intel 14 nm chips top out at 38.5 MB, mostly for high priced Xeon chips. With an active interposer, they could, for example) make something like 4 or 8 memory controller chips with large SRAM caches on 7 nm while using the active interposer for the IO drivers. Many different configurations are possible with an active interposer, so it is hard to speculate. Placing HBM on the IO interposer, as the AdoredTV guy has speculated, doesn’t sound like a great idea. Two stacks of HBM deliver 512 GB/s, which would take around 10 IF links to transfer to the CPU chiplets. That would be a massive waste of power. If they do use HBM for cpu chiplets, you would want to connect it directly to the cpu chiplet; you would place the a cpu chiplet and HBM stack on the same interposer. That would have some latency advantage, but mostly for large systems like Epyc.

Log in

Don't have an account? Sign up now