Intel shows off its Xeon roadmap through 2020

Intel is holding its Data-Centric Innovation Summit at its headquarters in Santa Clara, CA today, and the company took the opportunity to talk about its plans for its future server processors at the event.

This year, the company will be launching Cascade Lake Xeon Scalable processors. The Cascade Lake microarchitecture will include a new integrated memory controller with support for Intel's Optane Persistent Memory DIMMs and a new instruction set extension called DLBoost that tailors the AVX-512 instruction set for certain deep learning operations. The architecture will also have hardware-level mitigations for the Spectre and Meltdown vulnerabilities.

Cooper Lake will follow Cascade Lake in 2019 with a new platform that will also be shared with Intel's first 10-nm server architecture, called Ice Lake. Cooper Lake will still be built on a revision of Intel's 14-nm process technology, but it'll include support for the bfloat16 floating-point number format that's useful for AI-related workloads. Ice Lake will follow Cooper Lake in 2020, but aside from the new platform details, Intel didn't have any further information to share about the microarchitecture today.

Comments closed
    • ronch
    • 1 year ago

    I’m more interested to see Sapphire Rapids and Zen 2/3.

      • jts888
      • 1 year ago

      That’s going to be coming in 2022 or so. If things actually stay on track this time.

    • Wirko
    • 1 year ago

    Intel, please take note that gold is worth more than platinum per ounce.

      • chuckula
      • 1 year ago

      Don’t worry, our next lineup has “Early 2017 Bitcoin” at the high end and “Late 2018 Bitcoin” at the low end!

    • jts888
    • 1 year ago

    Some relevant data points left off these slides (as opposed to the leaked roadmap ones a few week ago):[list<][*<]Cascade Lake is still a 6ch DDR4 platform with max of 28c [/*<][*<]Cooper Lake is the last-minute Plan B after 10nm slid into its 4th year of delays and is still officially "In Planning" as opposed to "In Development" for Cascade/Ice Lake, and is scheduled for launch on December 30-something of 2019. My money is on 2 Cascade Lake respins on an MCM with 2*4 channel DRAM, which could at least get them to into 48c-56c territory with XCCs or 36c with HCCs. [/*<][*<]Intel has to date shipped maybe 100k 10nm laptop parts with 2c and all the iGPUs disabled with allegations that this took 6 months of pilot production to amass.[/*<][/list<] To drive the point home, Intel has nothing on the table for >28c/socket parts until the beginning of 2020, and AMD will be selling 48c and potentially 64c MCMs late next Summer. But hey, at least Meltdown and maybe and handful of the Specter variants will be gone around the start of next year.

      • Anonymous Coward
      • 1 year ago

      Not a lot on that roadmap to keep a person’s interest, 16-bit float on a CPU seems pretty niche. Good plan with the MCM of course.

      • the
      • 1 year ago

      At this point I’d expect a chiplet design with EMIB than a more traditional MCM part. Minor quibble on the difference but they could scale the design far higher and control yields better at the price of packaging complexity. This could get them into >64 core/socket terroritory. Main issue moves to power and thermals. Intel is spec’ing out 230W for high end Cooper Lake it seems. I’d almost expect that to change to 250W or higher depending on that they’d need to do to remain competitive with AMD. If AMD moves Epyc to 250 W or more and isn’t shunned by the market (i.e. It delivers performance) Intel will follow suit.

        • jts888
        • 1 year ago

        I think our thinking is largely in line regarding thermals/power, but you are waaaaaaaay more optimistic on EMIB than I am. Even if Intel has all the packaging/mechanical issues ironed out already (and that’s a long way from a HBM bridge to a Radeon chip), that line of attack requires adding interconnect switching, buffering, and protocol bridging that would seem to be without precedent for them.

        If they had at least another year, I could see it happening, but remember that “In Planning” means that there is not even a solidified plan yet, much less real execution towards a finished product. I feel like their time window leave them room to do little more than a Core 2 Quad-style slapdash design with power-tuned same-package UPI and the rest of the signal/thermal/mechanical validation. IMO, successful launch of a EMIB-based Xeon series in Q1’20 would require perfect execution, at least a moderate level of divine intervention, and having already started a year ago.

          • the
          • 1 year ago

          I’ll admit I am optimistic about EMIB as it is a solution for the lack of node updates of late by throwing more silicon at the problem.

          Kaby Lake-G isn’t the first chip Intel has used EMIB on, just the first mainstream product. There are several [url=https://www.anandtech.com/show/12477/intel-launches-stratix-10-tx-leveraging-emib-with-58g-transceivers-<]Altera FPGA parts[/url<] that leverage it for transceivers which makes sense as the transceivers contain a bit of analog logic. Similarly EMIB is also used for HBM with some Altera products too. Intel coincidently has [url=https://www.eetimes.com/document.asp?doc_id=1333512<]announced a month ago AIB[/url<] which covers the issues of switching, buffering, and protocol bridging between IP of different vendors. In the context of Xeons though, they moved to a 2D tile based design for the cores. EMIB would simply be extending the tile designs between cores as if it were one large piece of silicon. Granted there is far more engineering than that (power, clock routing etc.) but Intel's design appears outwardly to lend itself to that level of scaling. In other words, Sky Lake-SP seems to have been designed with this concept in mind.

    • Unknown-Error
    • 1 year ago

    So will Ice-Lake be the new Conroe?

      • chuckula
      • 1 year ago

      More like the new Prescott!

        • ronch
        • 1 year ago

        Now wouldn’t that be the greatest irony in CPU codenames?

      • NTMBK
      • 1 year ago

      Or the new Barcelona…

        • chuckula
        • 1 year ago

        Stop being a pro-Intel shill!

        You should have said “Or the new Bulldozer”

      • freebird
      • 1 year ago

      I(ce)tanium2.

    • ronch
    • 1 year ago

    I must say, whether they fully foresaw something like this happening or not, AMD’s strategy to be flexible with the manufacturing partners seems to be really paying off now. And at a time when Intel is having trouble with their manufacturing, no less.

      • Shobai
      • 1 year ago

      At the same time, though, check the news for TSMC + virus for some giggles… (I don’t recall seeing a story about that here, yet, but I may have missed it).

        • ronch
        • 1 year ago

        Really? So has that already become.. viral?

        • JustAnEngineer
        • 1 year ago

        Chuckula did it.
        [url<]https://in.reuters.com/article/taiwan-tsmc-virus/apple-chip-supplier-tsmc-resumes-production-after-wannacry-attack-idINKBN1KR0B9[/url<]

          • chuckula
          • 1 year ago

          That’s insulting.

          I would have done it a hell of a lot better than just throwing wannacry on a PC.

    • anotherengineer
    • 1 year ago

    Does it come in Titanium?

      • the
      • 1 year ago

      It used to come with one less T.

    • Kougar
    • 1 year ago

    AMD vs Intel, Round 2. FIGHT!

    This time around Intel won’t have much if any foundry advantage left to rely on. Gonna be a fun match to enjoy!

      • Gadoran
      • 1 year ago

      Trust me AMD Rome will not be exciting at all and it will not change the market. IMO it will be a many cores SKU clocked very low for applications of nice.

    • techguy
    • 1 year ago

    I must admit that no 10nm server parts until 2020 is disappointing, and Intel’s execution must be seen as lacking, as of late. I wouldn’t mind seeing AMD pick up some much-needed market share so we can return to the glorious days of strong competition!

      • Krogoth
      • 1 year ago

      The 10nm process in its current form is DOA. Intel already spend billions of capital and countless manhours on it. They still cannot make anything large and complicated on it. It would be a miracle if they managed to make server-tier SKUs on it tangible by 2020.

      • thx1138r
      • 1 year ago

      The next question is [b<]when[/b<] exactly in 2020 is 10 going to make an appearance, early or late. I too am thoroughly enjoying the increased CPU competition we've been seeing of late, hopefully AMD will pick up some enterprise market share so that they will be able to afford to keep the competition going for some time to come.

        • freebird
        • 1 year ago

        Intel stated Holiday 2019 for 10nm production rollout on their Q2 2018 Earnings call, most likely a laptop/desktop (small die) CPU (they didn’t specify what would be made on it first. As for Server 10nm, 2H 2020 seems likely, since Q1 2020 is the release date slated for Cooper Lake.

        [url<]https://wccftech.com/intel-xeon-scalable-family-roadmap-cooper-lake-14nm-2019-ice-lake-10nm-2020/[/url<] When looking at the Intel HPC Focused Roadmap - June 2018 Take note of the note...for Xeon Scalable Performance 1. Leading Edge of box indicates Lead Silicon Die Factory Production ramp begins.... * Add 1 or 2 Quarters for End-user availability.

        • TheRazorsEdge
        • 1 year ago

        There is no guarantee 10nm will make an appearance. It probably won’t ship in significant volume unless there are big problems with 7nm as well.

        Intel has already been working on 7nm for a while, and Fab 42 in Arizona is slated to go online in 2020-2021. They’ve delayed 10nm almost to the point where 7nm launches–assuming, of course, that 7nm isn’t also delayed.

        This wouldn’t be the first generation skip. Most foundries saw limited use of 20nm due to horrendous yields, so they jumped directly from 28nm to 16/14.

        That double-generation jump in process tech is part of the reason GeForce 1000-series and Radeon Polaris/Vega GPUs are so far ahead of their predecessors.

          • freebird
          • 1 year ago

          Yes, but we are currently working under the assumption that Intel hasn’t been lying to us the last 2 years…so 2H 2020. ; D

          Although, I agree with everything you wrote, but I get down voted for posting such things around here.

          • the
          • 1 year ago

          The other foundries doing 7 nm soon are closer to Intel’s 10 nm process than what Intel is expecting to ship for their own 7 nm node. However, the other foundries are tinkering with EUV for their 7 nm which in context of Intel’s contrary decision, the correct choice.

          Similarly the other foundries had issues at 20 nm as they were still leveraging planar transistors where at Intel moved to tri-gate. For Intel, this was a victory as moving a generation before everyone else was a clear strategic advantage.

            • Gadoran
            • 1 year ago

            The real question, what is Foundries 7nm?? It is feasible for a server or a fast Pc cpu?

            AMD Rome tell us nothing, it is nebulous and likely will be clocked very low IF it will yield enough to ship in quantities.
            Ryzen on TSMC 7nm is missing in action.

            For now we know IBM refused 7nm and POWER 10 will be out on a customized GF 10nm (more like a relaxed GF 7nm).
            This tell us it is impossible to integrate a fast cpu of actual Foundry 7nm processes.

            • the
            • 1 year ago

            Rome and Ryzen 2 have been on the road map for a 2019 release for awhile so they’re not MIA yet. First big chip on 7 nm to my knowledge is going to be Vega 20 which AMD has been demoing this year. nVidia could also be using 7nm from TSMC for their new GPUs later this year too. Apple is expected to move to Samsung’s or TSMC’s 7 nm node as soon as they can too. So there are plenty of 7 nm chips on the horizon that implies the other foundries have figured out how to get things to work for volume production. Give it a few more months and 7 nm chips will be in end user hands.

            IBM wants to stick with SOI so they are going with that GF process. Wild card would be if IBM brought POWER9 to that same process for a hypothetical POWER9+ refresh.

    • chuckula
    • 1 year ago

    “BFLOAT16”????

    So while AMD is shipping 64 core Epycs in January of next year Intel is just rebranding Super Nintendos?!?!!?!?

    Give up Intel, AMD owns the Enterprise market.

      • Krogoth
      • 1 year ago

      AMD isn’t going to be owning the Enterprise market anytime soon. They’ll just regain marketshare back and return to the Opteron/K8 days.

      It is going to take more than 10nm process woes for Intel to completely lose the SMB/Enterprise market.

        • the
        • 1 year ago

        From Intel’s roadmap, they’re vulnerable until 2020 when Ice Lake-SP arrives. Cascade Lake does appear to be a fixed Sky Lake-SP with Optane DIMM support. New on the roadmaps for Cascade Lake are seemingly Spectre + Meltdown fixes which will certainly help in the server market.

        The wild card is Cooper Lake. Cooper Lake could simply be Cascade Lake + on package Nervana coprocessor. Alternatively, Cooper Lake could be Cascade Lake with the same AVX-512 changes that occurred between Knights Landing and Knights Mill. Both help the emerging machine learning market but likely won’t mean much for legacy server software.

        So far nothing really jumps out in terms of major performance improvements for general server workloads. Niche areas are getting attention (HPC, AI, DB with Optane DIMM). Clock speeds will likely get a boost as with every recent respin.

        All AMD has to do is deliver on Zen 2 on top and with its targeted 10% to 15% IPC increases, and more cores per die and the server market starts shift over rapidly. It will be like the Opteron/K8 days as AMD will likely be able to sell every chip they can deliver.

      • Kretschmer
      • 1 year ago

      The man speaks truth to power.

        • chuckula
        • 1 year ago

        Somebody has to stop the constant pro-Intel bias that’s all over the place around here!

        I mean, just the other day some paid shill had the gall to suggest that ThreadRipper 2 isn’t the ultimate HTPC chip. It’s just disgusting to see the level of bias.

          • Kretschmer
          • 1 year ago

          I think I’m in love. Don’t hurt me like AMD hurt fanboys with Phenom I.

            • chuckula
            • 1 year ago

            Don’t worry, no matter what happens Intel won’t be able to beat a 16 core RyZen 2 in Cinebench!

      • derFunkenstein
      • 1 year ago

      Super NES turned out to be better at AI compute than Intel, so they had no choice but to start building 16-bit CPUs again.

        • chuckula
        • 1 year ago

        As usual Intel’s graphics solutions can’t keep up with the competition.
        Even when the competition is the SNES.

          • Shobai
          • 1 year ago

          There’s no denying it remains a spritely solution, to this day.

            • jihadjoe
            • 1 year ago

            It could also do polygons with the [b<]Super FX[/b<] chip. Hmmm

            • chuckula
            • 1 year ago

            The inclusion of the SuperFX chip in Starfox cartridges clearly marked the beginning of the end for discrete GPUs.

            • NTMBK
            • 1 year ago

            I wish modern GPUs came with a built in Starfox game.

            • Shobai
            • 1 year ago

            Or would that be the start, as an AIB?

Pin It on Pinterest

Share This