AMD Zen 2 Microarchitecture Analysis: Ryzen 3000 and EPYC Rome
by Dr. Ian Cutress on June 10, 2019 7:22 PM EST- Posted in
- CPUs
- AMD
- Ryzen
- EPYC
- Infinity Fabric
- PCIe 4.0
- Zen 2
- Rome
- Ryzen 3000
- Ryzen 3rd Gen
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.
Facebook
Twitter
RSS
216 Comments
View All Comments
scineram - Wednesday, June 12, 2019 - link
No.Xyler94 - Thursday, June 13, 2019 - link
YesXyler94 - Thursday, June 13, 2019 - link
If he meant 2700x, of course. Darn misreading :Pnevcairiel - Monday, June 10, 2019 - link
A quick note. AVX2 is actually primarily Integer. AVX1 (or just AVX) is 256-bit floating point. The article often refers to "full AVX2 support", which isn't necessarily wrong, but Zen2 also adds full AVX support equally.NikosD - Saturday, June 15, 2019 - link
AVX256 is both integer and floating point because it includes AVX2 FMA which doubles floating point capability compared to AVX1NikosD - Saturday, June 15, 2019 - link
AVX256 was a typo, I meant AVX2 obviously.eastcoast_pete - Monday, June 10, 2019 - link
Thanks Ian? Two questions: what is the official memory bandwidth for the consumer chips? (Sounds like they remain dual channel) and: Any words on relative performance of AMD's AVX 2 implementation vs. Intel's AVX 512 with software that can use either?emn13 - Tuesday, June 11, 2019 - link
AVX-512 is a really misleading name; the interesting... bits... aren't the 512-bit width, but the dramatically increased flexibility. All kinds of operations are now maskable and better reshufflable, and where specific sub-segements of the vector were used, they're now sometimes usable at 1bit granularity (whereas previously that was greater).Assuming x86 sticks around for high-perf computing long enough for compilers to be able to automatically leverage it and then for most software to use it, AVX-512 is likely to be quite the game changer - but given intel's super-slow rollout so far, and AFAIK no AMD support... that's going to take a while.
Which is all a long-winded way to say that you might well expect AMDs AVX2 implementation to be not all that much slower than intel's 512 when executing code that's essentially AVX2-esque (because intel drops the frequency, so won't get the full factor 2 speedup), but AVX-512 has the potential to be *much* faster than that, because the win isn't actually in vector-width.
GreenReaper - Tuesday, June 11, 2019 - link
Intel's own product segmentation has caused it to lose its first-mover advantage here. System software aside, there's little point in most developers seeking to use instructions that most of their users will not have (and which they themselves may not have). By the time software does support it, AMD is likely to have it. And of course an increasing number of developers will be pouncing on Zen 2 thanks to fast, cheap cores that they can use to compile on...HStewart - Tuesday, June 11, 2019 - link
Intel only had AVX 512 versions in Xeon and Xeon derive chips, but the with Ice Lake ( don't really count Canon Lake test run ) AVX 512 will hit main stream starting with in a month and 2020 should be fully roll out.As for AMD AVX 2 is true 256 bit, the last I heard is that it actually like dual 128 bit unless they change it in Zen 2. I serious doubt AMD AVX 2 implement is going to any much different that Intel AVX 2 and AVX 512 is a total different beast.
It funny years ago we heard the same thing about 64 bit in x86 instructions, and now we here in 512 bit AVX.
As for as AMD support for AVX 512, that does not matter much since Intel is coming out with AVX 512 in full line over next year or so.
But keep in mind unlike normal x86 instruction, AVX is kind of specialize with vectorize processing, I know with Video processing like Power Director this was a deciding factor earlier for it.,