Rumor: Ryzen Threadripper 2990X makes its first showing at e-tail

AMD's second-generation Ryzen Threadripper CPUs are coming, and recent rumors have suggested the top-end part in that family will be called the Ryzen Threadripper 2990X. The first retail listing for such a chip just added some fire to that smoke. German retailer Cyberport has put up a retail listing for the 2990X (as spotted by VideoCardz).

While the specs Cyberport lists for the purported 2990X appear preliminary, the site seems a lot surer in its pricing of this as-yet-unannounced part. With VAT, the 2990X could ring in at €1509. Lop the 19% German VAT off that price, and we get €1222.29, or around $1420 USD at today's exchange rates.

Cyberport does list some specifications for this purported chip, but most appear to have been copied and pasted from past Ryzen listings. The headline for the listing does suggest the chip will have 32 cores running at 3.4 GHz, although the site doesn't say whether that's a base speed or an all-core Turbo speed. Cyberport further says the chip will have a 4-GHz boost clock. The site's item description says the 2990X will be built on GlobalFoundries' 12-nm process and will carry a 250-W TDP.

The 3.4-GHz all-core speed and 4-GHz boost speed are consistent with past reports from Hong Kong hardware site HKEPC. We'd expect any second-generation Ryzen part to be built on GloFo's 12LP process, and the 250-W TDP is also corroborated by HKEPC's early information. Little else about the chip would seem to be confirmed by Cyberport's listing.

Although it's difficult to make reliable predictions through the lens of exchange rates or VAT, a 32-core Threadripper 2990X for $1399 or $1499 would—at least by sheer core count—seem poised to rip apart Intel's Core i9 family. It'd be priced well below the halo Core i9-7980XE while potentially delivering much better performance in some workloads. We'll still need to see how the 2990X performs across the board given the potential challenges of its NUMA arrangements, but signs point to another aggressive move by AMD in the contested HEDT space.

Comments closed
    • ronch
    • 1 year ago

    Good grief this thing has enough threads to make a sweater.

    • ronch
    • 1 year ago

    Newbie here. I’ve been using an FX-8350 for a while now. Will this be a good upgrade for me? Will it fit into my existing motherboard?

      • chuckula
      • 1 year ago

      [quote<]Will this be a good upgrade for me?[/quote<] THIS IS THE ULTIMATE UPGRADE! [quote<]Will it fit into my existing motherboard?[/quote<] IT WILL TOTALLY FIT IN THAT THREADRIPPER MOTHERBOARD THAT'S BEEN SITTING UNUSED NEXT TO YOUR FX-8350 SYSTEM! AMD NEVER MAKES YOU BUY A NEW MOTHERBOARD!

        • ronch
        • 1 year ago

        Can I use my existing RAM with this? Sorry, I’m a fry cook and this is the first time I’m doing this.

          • K-L-Waster
          • 1 year ago

          If you RAM it hard enough I’m sure it will go in. Try a meat tenderizer for extra force.

            • ronch
            • 1 year ago

            I think only a DIMMwit would do that.

            • K-L-Waster
            • 1 year ago

            Would a rolling pin be better?

            • Redocbew
            • 1 year ago

            A sock filled with spare change is best. CPUs always do better when given more cash.

    • jokinin
    • 1 year ago

    It is expensive, but sure that is an impressive CPU, I wonder what kind of uses can benefit from having so many threads available.

      • blastdoor
      • 1 year ago

      I sometimes do big Monte Carlo simulations to investigate the properties of different statistical estimation procedures under various scenarios. A simulation involves at least 10,000 replications of applying the estimation procedure to a randomly generated data set. So, that’s up to 10,000 entirely independent threads. That falls under the heading “embarrassingly parallel.”

      • K-L-Waster
      • 1 year ago

      3D Rendering
      Databases
      Scientific & Mathematic Simulations

      That’s by no means exhaustive but it’s a good starting point.

        • Beelzebubba9
        • 1 year ago

        Chrome. You forgot about Google Chrome.

        My Ryzen 1800X system is the only PC I’ve ever used where I don’t have to close Chrome tabs to keep the system happy.

          • chuckula
          • 1 year ago

          What kind of coin mining are you doing in those Chrome tabs?!?!?

            • Redocbew
            • 1 year ago

            I predict gonzo-tab usage.

            “I have only 178 tabs open. I don’t understand why this is happening.”

      • ronch
      • 1 year ago

      I run Space Quest 3. I’m sure I’ll benefit.

      • Concupiscence
      • 1 year ago

      Seismic data processing is embarrassingly parallel – the more cores and RAM you can throw at the job, the better off you’ll be. The RAM would probably cost as much as the CPU – 32 gig sticks of ECC DDR4 ain’t cheap, and large seismic datasets run into the terabytes routinely – but this could be highly cost-effective if you have a big job and GPGPU computing wouldn’t do everything you needed.

    • Unknown-Error
    • 1 year ago

    I still can’t wrap my head-around the thread-count.

      • ronch
      • 1 year ago

      Neither can we. The skull is pretty rigid.

      • caconym
      • 1 year ago

      it’s got a higher thread-count than my bed sheets

    • ronch
    • 1 year ago

    Basically around 65w per die. Not too bad considering the decent clocks and core swarm it brings to the hive.

    Edit – a downthumb. I wonder why…

      • Srsly_Bro
      • 1 year ago

      Who knows. But two are better than one.

      Sup?

        • ronch
        • 1 year ago

        Two are better than one?

          • Srsly_Bro
          • 1 year ago

          You had one down-thumb when I made that post.

    • chuckula
    • 1 year ago

    The following is [url=https://www.youtube.com/watch?v=SBe1hwg7V-M&t=130s<] a dramatic re-enactment of AMD's Board of Directors[/url<] after they learned of the price cut.

      • Amiga500+
      • 1 year ago

      Sweet!

      Where can I get one of those flying jet-skis?

      It’d let me sail right over the rat-race in their rush hour traffic jams.

      • ronch
      • 1 year ago

      That bearded guy looks like Captain [s<]Hassock[/s<] [s<]Padlock[/s<] Haddock.

    • brucethemoose
    • 1 year ago

    Not bad. For reference, the Epyc 7551P has a $2100 MSRP, with a slower clockspeed.

    I wonder if AMD will give us a 24C TR for $1k.

      • Goty
      • 1 year ago

      Lower clockspeeds, but each die has direct memory access. I’d love to see some benchmarks between the two platforms to see the performance differences.

      • jts888
      • 1 year ago

      24c TR2 is pretty likely IMO, but the question is whether the core config would be 4d*2ccx*3c or 3d*2ccx*4c. My gut feel is that AMD might try a $1100-ish price if the $1500 price for 32c is true, but they are also on record as saying that the highest-end TR (1950X) dwarfed sales for the other models last time around, so they might not even bother.

      I’m sure there is some demand in the market for fully-populated Epycs with higher TDPs and closer to 4 GHz boost clocks, which the socket clearly physically supports, but all the enterprise validation for that might be prohibitive given the potentially moderate sales.

        • brucethemoose
        • 1 year ago

        24C $1k TR would also massively devalue the 1950X. Then again, it would make most of Intel’s HEDT lineup obsolete… so it depends on how much old TR inventory is left, I think.

        And yeah. I bet AMD won’t touch Epyc until the 7nm update, as it’s just around the corner IIRC.

          • jts888
          • 1 year ago

          I would guess that in AMD’s supply chain, TR-capable Zeppelin dies aren’t committed to an Epyc or TR substrate until the last possible moment, so there may not even be many 1950Xs sitting around in inventory. And you are correct in that Epyc will be skipping 12nm altogether.

            • freebird
            • 1 year ago

            Epyc skipping 12nm is a given, considering Lisa Su was holding 7nm (Zen 2) Epyc at Computex… never heard/seen a demo of 12nm Eypc in public or Quarter conference call.

            That and it is too short of a time frame and too little bang for the buck to try and stick it into the Enterprise market, especially given AMD stating 7nm Epyc is sampling late 2018 shipping 2019.

            [url<]https://wccftech.com/amd-zen-2-design-complete-7nm-epyc-2018/[/url<]

          • dragontamer5788
          • 1 year ago

          [quote<]24C $1k TR would also massively devalue the 1950X.[/quote<] The 1950x is already severely devalued: [url<]http://www.microcenter.com/product/483132/ryzen_threadripper_1950x_34_ghz_16_core_tr4_boxed_processor[/url<] $629 for the 1950x at Microcenter. I doubt they're selling "below costs", but maybe they're close to $0 margin. EDIT: Even Amazon / Newegg is selling 1950x at well under $800 these days.

            • synthtel2
            • 1 year ago

            If they can put dies that size in nominally-$200 GPUs along with power delivery and some pricey RAM, they can definitely turn a profit (probably a big one) on two of those dies in a $600 part without so much other hardware included.

            • Beahmont
            • 1 year ago

            CPU dies and cores are orders of magnitude more complex than GPU dies to design and manufacture. The price increases are not linear for the added complexity. And that’s before AMD had to pay the profit margins of the fabs they use, which is also likely to increase more than linearly to give a higher return for the fab on the more complex parts to create. As anyone at Apple can tell you, vertical integration has it’s benefits.

            So yes, while I don’t think AMD is selling these chips at a loss, equivalent die size GPU’s from a different fab, TSMC vs GloFlo, are not a good comparison at all.

            • synthtel2
            • 1 year ago

            Most of the stuff that doesn’t compare is about the costs to get a particular chip into production. Given the same process (it is actually the same fab and process) and good yields for both, the marginal costs per die should be pretty directly comparable, and while the CPU is a more complex chip (but certainly not by orders of magnitude), they’re also moving more volume on the CPU side to hide the increased fixed costs.

            I wouldn’t take this stance if we were talking about factor-of-2 differences, but other AMD products give us plenty of reference points to show that these 200-250 mm[super<]2[/super<] dies are cheap. Look at the Zen APUs too, if you'd like.

            • techguy
            • 1 year ago

            Micro Center has sold the 1950x as low as $699 over the past year, so $629 isn’t exactly a fire sale.

    • techguy
    • 1 year ago

    Going to be hard to resist replacing my 7900x at that price… I have workloads that need single-threaded and AVX512 performance occasionally, but 64 threads and PCI-e lanes for $1500 is utterly amazing.

      • DragonDaddyBear
      • 1 year ago

      I don’t think Ryzen has AVX512. Even the AVX256 I think is just 2 AVX units working together.

        • techguy
        • 1 year ago

        The workloads I run that use AVX512 have fallback paths. AVX and AVX2 should suffice.

        • dragontamer5788
        • 1 year ago

        Note that Intel currently implements AVX 512 as two 256-bit units working together. AMD, as you already mention, is two 128-bit units work together to emulate 256-bits.

        But if AMD can offer 2x the cores for the same price, then that’s strictly better. AVX512-based code will always scale well (otherwise, it couldn’t be coded as AVX512 in the first place!!)

          • chuckula
          • 1 year ago

          One of the AVX-512 units uses the same hardware as two AVX-256 units, but in the HEDT lineup each core also includes a dedicated AVX-512 unit as well [although the dedicated AVX-512 unit does not appear to have a shifter, it’s fully dedicated to FMA/FMUL and other ALU operations].

          [url<]https://techreport.com/news/33251/official-intel-docs-say-all-skylake-x-cpus-are-equally-avx-512-capable[/url<]

            • dragontamer5788
            • 1 year ago

            Interesting.

            So Intel has 3x 256-bit units, and 2x-AVX512 units (albeit with some degree of restrictions).

          • synthtel2
          • 1 year ago

          Not strictly better – SKL-X cores are good for 64 SP FLOP/clock (2x 512-bit FMAs) compared to 16 from Zen (2x 128-bit or 1x 256-bit FMAs) (half that in both cases if your workload can’t use FMAs).

          I’d be happy to trade away a lot of raw FPU for more threads, but it’ll depend on the workload.

            • jts888
            • 1 year ago

            Even in pure SIMD FP loads, it is very hard to do apples to apple in any sort of broad sense. Skylake-E has 4x the per-clock FP throughput of Zen, and its datapath to L1D is actually 4x wider (2x64B reads + 1x64B write vs. 16B for each in Zen), but the datapath to L2 is only 2x wider, and the path to L3 is not much at all if any wider depending on how you look at CCX crossbars vs. meshes with and without congestion, etc. If you add on top of this the various levels of progressively stronger downclocking for wider AVX, things get even more murky.

            Given all this though, heavy use of really wide SIMD on CPUs is questionable given the FLOPS/Watt advantages of SIMT GPUs, which are effectively 1024b or 2048b wide units with various more leverageable caches with more specialized coherency characteristics.

            • synthtel2
            • 1 year ago

            I’m 100% with you on all points. My point was just that there are workloads where a 16C SKL-X will be a great deal faster than a 32C Zen due to well-understood things about the design, though they may be rare.

        • Sahrin
        • 1 year ago

        Zen has 4x 128-bit FPU’s.

        Also of note, Skylake-SP has 4x 256-bit FPU’s.

        A vector unit is a processor that performs the same operator on several operands – when they say “512-bit” what they mean is it can do 8 64-bit floats (or any combination of precision adding up to 512 bits). It is not an execution unit that can operate on a 512-bit operand.

      • ronch
      • 1 year ago

      Hard to resist replacing your 7900X? Just wait for this month’s bills, bud.

        • kuraegomon
        • 1 year ago

        I’m guessing he understands his cash flow situation better than you do 😉

    • ptsant
    • 1 year ago

    At $50 per core, it doesn’t even have a significant price premium compared with smaller CPUs.

    In other news, RAM prices seem to decline slightly. Time for a quad-channel build?

    • jts888
    • 1 year ago

    Let us not forget that the 10c Broadwell-based Core i7-6950X was released two years ago for $1700, nor that we will in all likelihood see 48c consumer MCMs next year too.

    What a difference a little competition can make. 🙂

      • DPete27
      • 1 year ago

      It’s just a thread count race nowadays. Same as all the other silly spec-races out there. Everyone gets carried away an loses track of reality. This thread count race is good in the sense that the few professionals that actually NEED this many [b<][u<]CPU[/u<][/b<] cores/threads can get them much cheaper today than ever before. But again, its just AMDs effort to win Enterprise market share from Intel.

        • nanoflower
        • 1 year ago

        It’s only silly if you convince yourself to buy one when you don’t need it. Otherwise it’s just great for people that want more popular CPUs as the competition at the high end looks to be going to drive down their prices which should mean a decline in prices down the entire lineup. That’s good for every customer.

          • DPete27
          • 1 year ago

          The thing I worry about is how this “distraction” plays out for the “average” consumer/gamer. People who are still unlikely to ever use more than 8 threads into the forseeable future.

          Granted, back in the day when AMD helped pull the CPU arena out of single-core, there was much to be gained from moar coarz. Today, it seems we’ve hit a wall with multi-threading many consumer level software. And with the IPC/frequency race drying up as well, AMD definitely capitalized on the low-hanging fruit of core count. But if we don’t see another “revolution” of multi-threading, I fear the R&D is moving away from the advancements that affect consumer desktop performance the most.

          Between this “Enterprise” core race and the mobile-side power-efficiency race, the already small slot for consumer/gaming desktop CPUs continues to shrink.

            • K-L-Waster
            • 1 year ago

            These are kinda like the CPU equivalent of full-size pickup trucks.

            Most people would be better off with something smaller, but there are some people who really do need them.

            And then there are those who don’t need them but *think* they do… mostly for bragging rights…

            • synthtel2
            • 1 year ago

            We’re already seeing games going to 12 plenty well. 😉

            Two years ago, there wasn’t much incentive to go beyond 8 in consumer software. I don’t think consumer software not going much beyond 8 is because we’re running into any particular wall, but because it takes a bit longer than this for software to change course to reflect changes in new hardware. This stuff doesn’t get rewritten every year, and the existing code carries a lot of assumptions (explicit and purposeful or otherwise) about the hardware it’s supposed to run on.

            5-10 years ago, it was counting on tomorrow’s single-thread performance being better than today’s, and that assumption resulted in a lot of games that still won’t do a solid 144 fps on 2018’s best hardware. 0-5 years ago, the hardware targets were pretty impressively stable, so we’ve got games that very cleanly meet pretty good performance targets on that manner of hardware, but don’t scale beyond it so well. Maybe in 2023 we’ll be lamenting how games focus on threading too much at expense of something else (relying on too much frame pipelining and too many hacks to bring latency back down seems like a decent bet).

            Heavy threading does have a big advantage over focusing on some other aspect of performance: extra cores stay out of the way nicely when not in use. It’s tougher to say that about clocks, SIMD, or even instruction-level parallelism.

        • jts888
        • 1 year ago

        I agree that it is difficult for me to personally see much immediate usefulness in a 4 die/2+2+0+0 channel/32 core workstation chip, but I can at least appreciate that it can only help bracket into a less stratospheric price range next year’s 7nm/2d/24c TR3, which feels like a more broadly useful design to me.

        It would be nice to be able to spawn off a 6c or 8c Windows 10 sandbox and barely notice the loss from my host’s usable CPU pool.

        • freebird
        • 1 year ago

        Yeah, and that “silly” Enterprise market that revolves around threads, density and efficiency is only worth what $18 Billion dollars…a QUARTER now a days.

        [url<]https://www.idc.com/getdoc.jsp?containerId=prUS43860318[/url<] Also the reason, we'll see 7nm Epyc 2(Zen 2) long before the 7nm Ryzen 3 Desktop. AMD is going after the big money. So, it looks like we'll (PC community) have to wait for "sloppy seconds" just like with 7nm GPUs from AMD, because there is definitely not enough 7nm manufacturing capacity go around; I don't expect Ryzen 3 until 2H2019, late Q3 maybe Q4 of 2019.

          • jts888
          • 1 year ago

          I think you could be right about your forecast of late 7nm Ryzen 3 if fab output stays too low for too long, but your general principle is backwards in general. Enterprise platforms need substantially longer validation periods to let customers sleep easier at night, and consumer parts, especially enthusiast ones, play a lot faster and looser.

          Zeppelin got a respin between the Ryzen 1×00 series and TR/Eypc, which allowed an L2 bug and performance sapping workaround (17 cycle latency vs. 12) to be removed. And the “embedded” Epyc line that actually uses the on-die 10 GbE controllers didn’t get released until this year, although it is not know if there is also a silicon revision bump there as well.

          My expectation is fully that consumers will be the guinea pigs for Zen 2 next year, but we will have to see.

      • ColeLT1
      • 1 year ago

      Even more impressive that year and some change ago (Feb-17) AMD’s top end chip was the FX-9590, that came out at $920 before the deep price cuts.

      AMD went from high dollar meh pre-sandy speeds, to broadwell speeds, with huge core counts in a freakin year, then the next year doubled the core count again. That’s impressive.

    • blastdoor
    • 1 year ago

    If that price ends up being right then Wowzers

      • Waco
      • 1 year ago

      No kidding, that’s a great price given the potential performance.

        • auxy
        • 1 year ago

        Yah woo [url=https://www.phoronix.com/scan.php?page=article&item=threadripper-linux-ddr4&num=1<]the potential performance.[/url<] ('ω')

          • ronch
          • 1 year ago

          Your comment can potentially attract downthumbs, you know.

            • Mr Bill
            • 1 year ago

            I fixed it.

      • anotherengineer
      • 1 year ago

      [url<]https://www.youtube.com/watch?v=9i7kQDne2Ns[/url<]

        • derFunkenstein
        • 1 year ago

        wowzers, that’s exactly what I was thinking.

Pin It on Pinterest

Share This