Sunday, September 29th 2019
Intel Sunny Cove Successor Significantly Bigger: Jim Keller
Sunny Cove is codename for Intel's first truly new performance CPU core design since "Skylake," and made its debut with the company's 10 nm "Ice Lake" processors, packing the first tangible IPC increase in years. VLSI guru Jim Keller is leading the effort to build Intel's future CPU core designs, and dropped a big hint on what to expect, speaking at a gathering in U.C. Berkeley. It's unclear which specific core Keller is referring to. The immediate successor to "Sunny Cove" is codenamed "Willow Cove," and Intel's own public sketch hints at an incremental upgrade over Sunny Cove, with faster caches and process-level optimization. It's only with "Golden Cove," slated for 2021, that Intel speaks of its next round of IPC increases (dubbed "ST perf"). It's plausible that Keller is referring to this core since a 2021 launch would fit better with a 2018-19 design phase.
In his talk, Keller describes Intel's next big CPU core as being "significantly bigger" than "Sunny Cove," with its 800-wide instruction window, and "massive" data- and branch-predictors, to put Intel back on a linear performance growth trajectory between generations. Keller also commented on this being a "mindset change" at Intel, which over the past decade, only delivered minor IPC increments between generations, and focused on other areas, such as efficiency. In stark contrast, through the 1990s and 2000s, Intel delivered IPC leaps between generations, such as the one between "Netburst" and "Conroe," and onwards to "Nehalem." These were in-part helped by rapid process advancements that slowed in the 2010s as Intel approached the sub-10 nm scale.The video presentation by U.C. Berkeley follows.
In his talk, Keller describes Intel's next big CPU core as being "significantly bigger" than "Sunny Cove," with its 800-wide instruction window, and "massive" data- and branch-predictors, to put Intel back on a linear performance growth trajectory between generations. Keller also commented on this being a "mindset change" at Intel, which over the past decade, only delivered minor IPC increments between generations, and focused on other areas, such as efficiency. In stark contrast, through the 1990s and 2000s, Intel delivered IPC leaps between generations, such as the one between "Netburst" and "Conroe," and onwards to "Nehalem." These were in-part helped by rapid process advancements that slowed in the 2010s as Intel approached the sub-10 nm scale.The video presentation by U.C. Berkeley follows.

59 Comments on Intel Sunny Cove Successor Significantly Bigger: Jim Keller
But no, memory latency haven't changed in the last 10+ years.I agree that consistency is much more important than average frame rate, no issue there.
DDR4-3200 CL14 is 4.37ns of CAS latency, and that's today's cheap stuff. You can buy 4600 CL17 if you want to pay the premium tax!
So yeah, we're probably looking at about a 50% improvement in latency alone if you take a mid-range kit from now and 10 years ago and compare the raw latency in nanoseconds. By saying memory latency hasn't changed in a decade is an insult to all the work and progress Samsung, Hynix, and Micron have made over the last few generations.
Ryzen loves faster RAM.
Intel loves faster RAM.
Even in 2017 on an i7-7700K, 3200MHz seemed to be the sweet spot with several mainstream titles doing up to 25% worse when dropping from 3200 to 2666. As far as most reviewers and youtubers are concerned, using less than DDR4-3200 is a bad idea these days because you're losing out on CPU performance. It doesn't matter if that's latency or bandwidth, the results are just worse with slower RAM.
You can read about memory latency here.
Access time of an arbitrary address in DRAM is about ~50ns, with memory controller overhead it's about ~70-90ns as you can see here. So when you compare memory with different CAS latency with is 1.5-2ns quicker, it's not 50% quicker, more like ~2-3% quicker (and that's assuming you are able to actually run it on the best case non-JEDEC speeds they put in the spec sheet).
As I said, DRAM latencies haven't changed much changed a lot since the old Pentiums, which was still around ~>80ns.Zen/Zen2 benefits from faster memory frequencies because other timings are tied to it internally, so that's only indirectly.
As you increase the memory speed, you generally only get more bandwidth, as CAS latency has to be increased as the clock increases due to the inherit latencies in DRAM.
And if bandwidth were a bottleneck for gaming, gamers would all buy HEDT CPUs with quad/hex/octa channel memory, that's the easiest way to get a lot of memory bandwidth.
Also, the preliminary specs for upcoming DDR5 expects higher latencies than DDR4.
In terms of memory improvements though, I don't think it even matters that overall memory latency hasn't changed much. People can argue hypothetical theory all day, but in the real world, in real applications and real games, a significant portion of performance can be attributed to faster RAM, with a simple litmus test of "put in slower RAM and watch the performance vanish".
DDR4-3200 CL14 is most definitely not cheap stuff. DDR4-3200 CL16 costs almost half of what CL14 costs.Gamers did buy HEDT CPUs for a while and there are clear benefits even in gaming from more memory bandwidth. However, today the HEDT CPUs are overkill in terms of core count on one hand and/or not the best organization of inter-core communication for latency and this has clear negative effect on gaming. Both Intel's mesh CPUs and Threadrippers have the same problem here. Zen2-based Threadrippers might bring a change to this, hopefully.As was the case with DDR2, DDR3 and DDR4 ;)
As for pricing, I suspect that's just regional variation. 3200 14-16-16 is guaranteed on Samsung B-die It shouldn't be expensive and I certainly haven't paid much of a premium for it in the last quarter. You do have to read between the lines a bit because JEDEC-spec 3200 (PC4-25600) ranges from CL22 to CL20 and it's practically impossible to even buy RAM that slow. Even the lowest-grade Micron A-die or Samsung B-die will likely meet AMD's recommendations of 3200 CL14. It's not as if the XMP info in the SPD is particularly helpful either because it's not optimised for Ryzen and even across different generation of Intel DDR4-compatible CPUs there's a lot of variation in their RAM tolerance and IMC quality.
What I tried to say in my previous post is that none of these minor latency or timing details really matter. If you take any modern processor on today's mainstream DDR4 (somewhere around 3200) you are going to see drastic performance loss if you swap in much slower RAM, which brings us back to the original point of this discussion - that significant IPC gains made in the last decade are at least partly attributable to the improvements in DRAM speed and not something Intel can claim as true IPC gains. We all know that Haswell benefitted from a RAM overclock, despite the spec only being officially capped at 1600MHz.
You are also talking about memory overclocking here. Yeah, B-dies do 3200 CL14 but out-of-box spec module prices are a bit different thing. Looking at prices in Europe, 2x8GB CL16 will cost about 70€ and CL14 about 130€. Same relative difference applies to other sizes as well. JEDEC spec is a bit of a different thing because it states 1.2V as working voltage and low voltage is what gives you 3200 CL18-22.
How much of the speedup is attributable to RAM speedup is not a single question. There are benchmarks that really don't care. Cinebench (especially R15) is a very good example.