AMD’s Ryzen Threadripper 1950X, Threadripper 1920X, and Threadripper 1900X CPUs revealed

I’ve been spending the weekend in sunny Los Angeles with AMD to learn about two of the company’s most important product launches this year. Ryzen Threadripper is that launch for AMD’s CPU engineers, but the company has also been sharing details of its Vega architecture and products with us over the past couple days. If you’d rather read about what’s happening with Vega first, feel free to go check out my separate article discussing AMD’s plans for its Radeon RX Vega 56 and RX Vega 64 graphics cards.

All Threadripper chips start with an Epyc-esque multi-chip module that uses two of AMD’s eight-core Zen dies connected over the company’s Infinity Fabric. To achieve its quad-channel memory architecture, Threadripper yokes the dual-channel memory controller on each die into the fabric, and it also enables all 32 of the PCIe lanes available from each die to create the 64 PCIe lanes available to all Threadrippers. All Threadripper CPUs will have support for ECC RAM, as well, and AMD said its motherboard partners have all included support for the feature in their accompanying motherboards.

  Cores Threads Base clock Four-core

boost clock

XFR

boost range

PCIe 3.0 lanes

from CPU

Memory

channels

Price
Ryzen Threadripper

1950X

16 32 3.4 GHz 4.0 GHz 200 MHz 64 4 $999
Ryzen Threadripper

1920X

12 24 3.5 GHz $799
Ryzen Threadripper

1900X

    3.8 GHz $549

Four of those lanes are dedicated to the X399 chipset, leaving 60 PCIe lanes available to motherboard makers for PCIe or M.2 slots. The two active dies on a Threadripper MCM are arranged diagonally on the package, and they’re flanked by two dummy dies that provide stability to the large integrated heat spreader that caps off the whole affair. That integrated heat spreader is soldered to the dies beneath for effective heat transfer.

AMD has already disclosed clock speeds and core counts for its two highest-end Threadrippers, but we now have more details about them. The Ryzen Threadripper 1950X will offer the full 16 cores and 32 threads of compute power that Threadripper promises. AMD says it’s selecting the best 2% of dies for use in Threadrippers. That means the 1950X will offer 200 MHz of XFR headroom for operation at speeds of up to 4.2 GHz in lightly-threaded workloads. The chip will also be able to boost up to 4 GHz on two cores of each eight-core die.

Overclockers may also find plenty of potential to be tapped within reasonable voltages as a result of that binning. AMD testing guru James Prior noted that he was able to overclock his own Threadripper 1950X to 4 GHz across all cores using just 1.325V, as opposed to 1.4V for his personal Ryzen 7 1800X.

AMD’s in-house team of extreme overclockers tried their hand at putting a Threadripper 1950X under liquid nitrogen last night, and the net result was a roughly 5.2 GHz all-core overclock. With 16 cores and 32 threads churning away at those speeds, the chip produced a Cinebench all-core score of 4188. For perspective, that’s 2000 points more than the Core i9-7900X turned in during our review.

The Ryzen Threadripper 1920X will offer 12 cores and 24 threads of compute power for $799, as well. This chip will run with the same 4.0 GHz four-core boost as the Threadripper 1950X, and it’ll have a slightly higher 3.5 GHz base clock than its 16-core sibling. Like the 1950X, the Threadripper 1920X will offer 200 MHz of XFR headroom in lightly-threaded workloads.

For those who’d rather have four channels of memory to pair with eight Ryzen cores—something Ryzen 7 chips can’t boast—AMD will offer the previously-undisclosed Threadripper 1900X. This chip will have a 3.8 GHz base clock, a 4 GHz boost clock, and the same 200 MHz of XFR headroom as its brethren. Compared to the Ryzen 7 1800X, the Threadripper 1900X will offer a 200-MHz-higher base clock and 100 MHz more XFR headroom to go with what will surely be much higher memory bandwidth. Like its counterparts, the Threadripper 1900X will still offer all 64 lanes of the platform’s PCIe 3.0 connectivity to host motherboards.

For users who want even higher stock-clocked performance or overclocking potential than the Ryzen 7 1800X, the Threadripper 1900X could prove an appealing point of entry to the X399 platform (unlike the bemusing Kaby Lake-X CPUs for Intel’s X299 motherboards), and it’ll start at $549—just $50 more than the nominal price of the Ryzen 7 1800X.

AMD didn’t officially disclose the cache amounts and core configurations for Threadripper dies at its tech day, so that information will need to wait until our full review. Each full eight-core Zen die offers 4MB of L2 cache and 16MB of L3 cache, though, so the numbers should be prodigious. I’m especially curious to know how AMD is configuring the eight-core Ryzen 7 1900X, but the company is remaining tight-lipped for now.

 

Laying claim to the high-end desktop performance crown

We’re always skeptical of manufacturer-provided performance numbers, but it’s hard to argue with AMD’s proclamations of dominance for Threadripper performance against the Intel competition. The only Core i9 CPU available so far is the 10-core Core i9-7900X, and it sells for the same $999 suggested price as the Ryzen Threadripper 1950X. Given that the AMD part boasts a whopping six more cores and 12 more threads, even the single-threaded performance advantage and AVX-512 support the i9-7900X offers isn’t likely to be enough to let it overcome the Threadripper 1950X.

AMD’s numbers for the Threadripper 1950X handily beat the i9-7900X. At least in heavily-multithreaded workloads, the chip seems ready to lay claim to the title of highest-performing chip around for its price. Intel’s Core i9-7980XE promises 18 cores and 36 threads of Skylake-X, but that chip will sell for $2000, last we heard. For the moment, AMD seems ready to claim the absolute performance crown.

The Ryzen Threadripper 1920X, on the other hand, promises near-parity with the $999 i9-7900X across the same range of workloads. Presuming one’s workload doesn’t involve heavy use of AVX instructions, the 1920X’s $799 price tag could prove quite compelling for an entry-level workstation or heavy-duty streaming machine.

Threadripper could also offer impressive performance-per-watt figures to go with its raw performance. AMD claims that even though the Threadripper 1950X is 24% faster than the i9-7900X in its Blender testing, the AMD CPU requires 2% less power at the wall to get there—a claimed 29% performance-per-watt advantage. I’m eager to see how that potential plays out in my own testing, but it does suggest Threadripper will be quite efficient.

AMD also showed (but didn’t provide us with) gaming performance numbers for Threadripper CPUs that promise near-parity with Core i9 CPUs in both 1920×1080 and 2560×1440 scenarios. I’d question the wisdom of building around any of these high-end desktop chips for gaming and gaming alone, given their high price tags and likely platform costs. Still, if one needs to game and do the many other things that Threadripper promises to be capable of, the 1920X and 1950X should be willing companions.

 

Packaging and platform considerations

Each Threadripper chip will come in a massive box without a cooler. Instead, Threadrippers will come with a bracket for any recent Asetek-powered liquid cooler. That bracket will allow Threadripper builders to take advantage of a wide range of compatible liquid coolers from the word go. The Threadripper package will also include a torque wrench that builders will need to unlock and secure the CPU retention mechanism in the massive TR4 socket.

AMD expects that most builders will want to liquid-cool such a large and high-octane chip, but some of its partners will offer air coolers for those who want them. The company says options from Arctic Cooling and Cooler Master will be available at Threadripper’s launch, and options from Noctua will follow later this year.

AMD will have X399 motherboards from all of the big four motherboard makers at launch: Asus, Gigabyte, ASRock, and MSI. We don’t have full details of the X399 chipset and its resources yet, but we’ll know soon enough.

The Threadripper 1950X and Threadripper 1920X will be available August 10, followed by the Threadripper 1900X on August 31. We expect to have review hardware soon, and we’ll have as thorough a performance picture as we can get for those parts by the August 10 launch. Stay tuned.

Comments closed
    • Mr Bill
    • 2 years ago

    Its exciting to see that this design actually has headroom given sufficient cooling. This bodes well for future die shrinks and process improvements.
    [quote<]AMD's in-house team of extreme overclockers tried their hand at putting a Threadripper 1950X under liquid nitrogen last night, and the net result was a roughly 5.2 GHz all-core overclock. With 16 cores and 32 threads churning away at those speeds, the chip produced a Cinebench all-core score of 4188. [/quote<]

    • Shouefref
    • 2 years ago

    Very very interesting those big doodles.
    I just wonder: if you buy such a big one now, will it make your rig longer lasting than a ‘normal’ cpu? Because a few thousand dollars for a cpu is a lot of money, and that would only be interesting if it makes your system last longer than a cheaper alternative.
    Or could you take them to your next rig, changing the other elements but keeping the processor?

      • Shouefref
      • 2 years ago

      Oh, nice, I’m 4x downvoted …
      But that doesn’t give me an answer to my question, does it? If you’re only capable of downvoting somebody who asks a question, then, please, don’t bother and just f*k off.

    • albundy
    • 2 years ago

    did amd fix the memory support issues on the new chipset that has plagued the x370/b350 chipset since day 1?

    • Brainsan
    • 2 years ago

    “include a torque wrench”

    Seriously? An actual torque wrench? Wouldn’t that be expensive? Is it calibrated?

    Or do you (or AMD) mean something entirely different from torque wrench?

      • jts888
      • 2 years ago

      Torx(tm) wrench I think they mean, or at least a similar enough star wrench.
      Note: I don’t actually know how one should pronounce Torx, but I think it’s close enough to “torque” that confusion is understandable.

        • Brainsan
        • 2 years ago

        Yep, Torx wrench it is.

        [url<]http://images.anandtech.com/doci/11689/15_-_dismantled.png[/url<] From: [url<]http://www.anandtech.com/show/11689/amd-threadripper-unboxing[/url<]

      • Anonymous Coward
      • 2 years ago

      I think it would be easy and cheap to produce a torque wrench which could be used just a few times at exactly one torque setting. You just need your metal to flex predictably and some mechanism for the user to know when sufficient flex has occurred. It could be a plastic sheath over a hex key, the end of the key would extend beyond the plastic sheath, and the plastic would be attached in such a way that the end of the key would contact the plastic when it has flexed enough.

    • Thresher
    • 2 years ago

    If this lights a fire underneath intel, then it’s good for everyone.

    Intel has stagnated because there has been no real reason to do much better than they’ve been doing. Prices have stagnated as well. Real competition from AMD could make all CPUs a better value proposition than they have been in 10 years or so.

    • Mr Bill
    • 2 years ago

    Using that spatial symmetry, YEAH!.
    [quote<]The chip will also be able to boost up to 4 GHz on two cores of each eight-core die.[/quote<]

    • willyolioleo
    • 2 years ago

    1900x here i come!

    • Srsly_Bro
    • 2 years ago

    Hello Jeff. The price of the 1800x is no longer $499.99, the original MSRP. The current prices are much lower.

    [url<]https://techreport.com/news/32030/amd-and-newegg-drop-prices-on-ryzen-7-cpus[/url<] The 1800x officially dropped to $460 according to the link and often sells for less.

      • f0d
      • 2 years ago

      that seems to be a newegg only pricedrop

      • chuckula
      • 2 years ago

      At no point has AMD officially dropped the price of the 1800X.
      Stop trying to speak over Lisa Su heathen.

        • Srsly_Bro
        • 2 years ago

        My apologies to the great Dr. Su. She really is amazing.

    • Shobai
    • 2 years ago

    Your handy first page table is still missing entries for the 1900X.

    • NeoForever
    • 2 years ago

    “Threadripper” is starting to sound like a more appropriate (cool?) name for these CPUs.

      • [+Duracell-]
      • 2 years ago

      Does TR mean Threadripper or Tech Report?

        • MOSFET
        • 2 years ago

        I don’t know but your name is cool.

    • Kretschmer
    • 2 years ago

    Have fun, HEDT guys! I can’t get excited about tons of threads, but right now more than a few lab coats are getting sweaty.

    • davidbowser
    • 2 years ago

    My desire for these is tempered by the difference in TDP between the 1800X and 1920X. It goes from 95W to 180W, so until I see the benchmarks, I just can’t get as excited as the primal (MOAR THREADS!!!) part of my brain.

      • cygnus1
      • 2 years ago

      It is basically 2 1800X’s under one lid though right? So is that TDP really unexpected?

        • cygnus1
        • 2 years ago

        I mean, AMD has basically took all the guts needed for multiple sockets and stuck it into 1 socket. So when they drop multiple chips into a single socket, I think that power usage is fair in the first generation of it.

        • davidbowser
        • 2 years ago

        You’re right, of course, but I guess I was thinking that the slightly lower clocks would have the TDP in the 130-150 range (like combining two R5 1600). I will probably get over it once I see the benchmarks.

        I am a quiet PC person, so I am constantly trying to find the perfect balance of performance and heat/noise. The reality is that I am leaning towards an R5 1600 anyway because it is the sweet spot of cores/TDP.

          • cygnus1
          • 2 years ago

          I’m with you on the quiet PC preference. the TR parts are interesting, but not for my average desktop. I’m definitely keeping an eye out for boards that are more geared toward being a home server. Right now my lab gear is old Xeon’s, but only quads and maxed at 32GB of RAM. Kind of hoping SuperMicro comes up with something that’ll work with these. Throw an IPMI module and ECC ram compatibility in with a TR CPU and that’s compelling home server material right there.

            • ermo
            • 2 years ago

            And the R7 1700/1700X/1800X aren’t?

            8 cores/16 threads and 4×16 GB of ECC RAM on the right motherboard?

            What am I missing?

            • cygnus1
            • 2 years ago

            For starters, quad channel memory and 64 PCIe channels… I also haven’t seen any AM4 motherboards with IPMI but I think it’s more likely someone might release a TR4 board with that and a slew of PCIe x8 slots. That would make for a server rung just under the AMD Epyc processors.

            Edit: also I don’t think the AM4 chips support RDIMMs or LRDIMMs. I don’t know if TR4 does or not, but it’s still a possibility.

            • jts888
            • 2 years ago

            TR already got implicitly promised for RDIMM/LRDIMM support (“up to 1TB memory”) by one of their PR guys recently, but no mention about which actual mobos would do so (necessary trace impedances etc.).

            I’m still waiting for AMD’s 1d/2d BGA platform (Snowy Owl?), since Zeppelin’s closest match has always been 8c Broadwell-D.

            • cygnus1
            • 2 years ago

            Yep, that’s what I was thinking I remembered about the RDIMM/LRDIMM situation. It was the only way to get one of the systems up to the supposed memory limit. So my fingers are crossed that somebody (Supermicro is who I’m thinking, but maybe Tyan or Asrock) drops a motherboard with that memory support and IPMI onboard. I wouldn’t be mad if it had 10gigE onboard too 😉

    • jihadjoe
    • 2 years ago

    AMD marketing is totally bipolar looking from both the Ryzen and the Vega side.

    On Ryzen they are objective, honest, and open.

    From the Vega side they resort to all sorts of perception-based A-B blind tests, deceptive measurements (Only minimum frame rate range, really? That’s usually down to how much the game bogs down when assets load), and are evasive when pressed for facts.

      • davidbowser
      • 2 years ago

      That would tend to lead us to the conclusion many (non-fanatic) TR readers would make: Ryzen is great (competitive bang for buck) and Vega will likely be competitively good, but not great.

      Either way, we are all better off waiting for the TR and other benchmarks to come in on Vega.

        • Kretschmer
        • 2 years ago

        Ryzen is competitive in some use cases, but that price:performance ratio craters if you can’t use a bunch of additional threads.

      • bwoodring
      • 2 years ago

      My experience is that marketing people tend to be open, transparent, and straightforward – when they have a good story to tell. They tend to be the opposite when they are trying to sell a shit sandwich.

      • JustAnEngineer
      • 2 years ago

      Planet Earth is bipolar.

      • Gastec
      • 2 years ago

      Bah, the jihadi obviously just came out from under the proverbial desert rock .
      AMD’s Vega GPU will NOT do 120fps @4K resolution on a constant basis. It is known. That’s why Marketing didn’t spam us with false advertising.

    • muffindell
    • 2 years ago

    Wondering if this will be any good to replace my aging Sandy 2700k running@4.6mhz on air for sole use with Photoshop CS6? BTW I’m a pro-photographer, so large files are the norm. Any ideas?

      • brucethemoose
      • 2 years ago

      Fortunately, that’s exactly what reviews are designed to answer.

      • titan
      • 2 years ago

      Image manipulation programs are getting more threaded as time goes on., and memory intensive.

      The 4 channels to RAM will probably net you the largest gain.

      • cygnus1
      • 2 years ago

      Most likely this will be a very good upgrade from what you currently have for much less than an Intel solution. You’ll be able to put in more RAM, and that RAM will be much faster. But… wait for reviews with PS benchmarks to confirm because I don’t know that the crazy number of threads will be super useful in your day to day. So you might could go with the 1900X and put the money saved over the 1920X or 1950X into more RAM or NVMe SSD capacity.

    • DPete27
    • 2 years ago

    Just a few thoughts:
    1) I hope AMD isn’t cutting their margins too thin to gain market traction.
    2) Gotta disable half your $1000 CPU to be compatible with games
    3) [quote<]That integrated heat spreader is soldered to the dies beneath for effective heat transfer.[/quote<] Oooh! Burn!!!

    • guilmon14
    • 2 years ago

    I wonder how much quad channel memory will effect 1900x’s performance (compared to the 1800x)

      • Freon
      • 2 years ago

      There’s probably that one weird corner case where it helps a fair bit, but I’d guess zero for most, possibly worse off for games.

      • Goty
      • 2 years ago

      The gains Ryzen gets from “better” memory come from the fact that higher memory speeds are tied to an “overclock” of the infinity fabric that ties the CCXs together within the die. Doubling the number of memory channels will bring performance gains in memory limited scenarios, but you’ll still need to run higher clocked memory to see the general performance gains on Threadripper that you get on Ryzen under the same circumstances.

        • Brother Michigan
        • 2 years ago

        Not sure why this is getting downvoted since it’s simply a statement of fact…

    • chuckula
    • 2 years ago

    [quote<]Each Threadripper chip will come in a massive box without a cooler. Instead, Threadrippers will come with a bracket for any recent Asetek-powered liquid cooler. That bracket will allow Threadripper builders to take advantage of a wide range of compatible liquid coolers from the word go. [/quote<] Wait so there isn't a cooler in that giant case? Bummer. As for the selected benchmarks, even AMD isn't calling for a 60% performance advantage in heavily-threaded software even given the 1950X's higher base clocks.

    • kalelovil
    • 2 years ago

    “AMD didn’t officially disclose the cache amounts and core configurations for Threadripper dies at its tech day, so that information will need to wait until our full review.”

    You can work it out from the slide directly above that paragraph.

    i7-7820X has 8×1+11 = 19MB L2&3. Threadripper 1900x has 5% more, so 8x.5+16=20MB L2&3.
    Probably half the cores and half the L3 disabled in each CCX, like the Ryzen 5 1400.

      • kalelovil
      • 2 years ago

      Why the down-votes? If there is a factual inaccuracy I will correct it.

        • ClickClick5
        • 2 years ago

        Because this guy has been going through the posts lately and down voting everyone.

        [url<]https://i.ytimg.com/vi/OKnGFuthtFM/hqdefault.jpg[/url<]

      • jts888
      • 2 years ago

      I’m guessing the opposite, that there is just 1 fully enabled CCX and 1 disabled CCX.

      Same amount of L3 per core, slightly more physically concentrated heat production, but much lower average latency access to all same-die L3 and no unnecessary duplication of lines between same-die L3s.

      The bigger challenge for AMD was probably deciding between 2d*2ccx*(2c+8MB) design or a 2d*1ccx*(4c+8MB) one, since lower local L3 latency might still overcome halved gross L3 capacity.

        • Anonymous Coward
        • 2 years ago

        Why would they choose that arrangement for a ThreadRipper but not any Ryzen?

        Unfortunate that the went with halved L3, I think.

          • jts888
          • 2 years ago

          Because on TR each CCX would need to maintain coherency with 3 others in a 2*2*2 configuration instead of just 1 in a 2*1*4 one.

          R5 1400 and lower don’t have as many concerns about heavy thread interactions and could have (in theory) been chosen to use 2*2 as the designated dumping grounds for dies with defects in both CCXs (particularly in L3), just so dies with defects in 1 CCX only could be used in the TR 1900 style SKUs.

            • Anonymous Coward
            • 2 years ago

            So they are free to use only one CCX, they just preferred to use both on low-end Ryzen and presumably will choose to use only one on low-end ThreadRipper. That would put TR-1900 in a very similar situation to R-1800, I guess identical behavior except for a tiny bit higher latency between the 2 CCX’s that each of them will have. And that IO.

            This does sound like a good strategy to extract optimal value from a range of different defect scenarios.

    • chuckula
    • 2 years ago

    Bring back the turbo button!

      • Gadoran
      • 2 years ago

      Since Zen doesn’t overclock at acceptable voltages by default, AMD is doing a factory overclock at 4.2GHz “if all goes well” on one core.
      I don’t like these SKU pushed to the limits just to do a “hot” show.

      I think the only AMD accaptable Zen line is actually Ryzen PRO, the cpus have decent clock speeds inside the good portion of their power/frequence curve. Moreover they can survive ten years like all Intel SKUs.
      What will be the operative life of a standard Ryzen?? with the cores pushed at over than 1.4/1.45V under turbo operations??

        • chuckula
        • 2 years ago

        I was talking about the low-core mode they were advertising.

        • synthtel2
        • 2 years ago

        I’m not quite as worried about that as I am about that Skylake chip I saw doing ~1.38V all-core from the factory, because at least AMD is only running a core or two at once like that (for better thermals as well as the obvious wear distribution), but either way it’s disconcerting seeing factory voltages 100+ mV higher than what I’ve been willing to use in overclocking.

        I’m so tired of random hardware (including CPU) failures that I’ll probably run this 1700 I just bought at 945mV and ~3.0 GHz full-time (basically just disallowing turbo). Maybe it’s pointless resisting the era of CPUs that don’t last forever, but I’m going to try.

    • SkyWarrior
    • 2 years ago

    [quote<] AMD will also offer another compatibility mode of sorts for Threadripper regarding its active core count. The "Creator Mode" will run every core and thread on the chip normally, and it'll be the out-of-the-box configuration. However, AMD has discovered that not every application will work properly with 32 threads available. For compatibility with those applications, a "Game Mode" will leave both dies active but only eight cores and 16 threads processing work. Memory access in Game mode is automatically switched to the local mode we just discussed, and the active chip will have its local memory pool filled up first before the remote die begins using its memory pool. AMD says it doesn't expect more than a 5% performance delta between games on average in Game Mode, and that most workstation applications won't see any performance benefit from disabling cores at all. [/quote<] Am I the only one who did not like this idea?

      • Anonymous Coward
      • 2 years ago

      Makes me wonder how it is implemented. Windows drivers?

      Hopefully they can drop the strange workarounds once more people have a chance to test and fix their software with large core counts.

        • SkyWarrior
        • 2 years ago

        I bet the implementation will be best on windows (which will be MEH) and probably a mess on initial linux kernels.

        When I buy a 16 core processor with a quad channel memory I don’t want to use only half of it during a gaming session.

        Consider a scenario where you have 16gb memory on half of the quad channel and 16 on the other half. During gaming mode you will have full speed access to only half of your total memory and the rest will have latency. Similarly I wonder if SLI CF implementation will be held back due to dormant core group ?. Which gpu will be connected to which core group?

        I bet 64 PCI express lanes are only connected 32 + 32 to each 8 core groups.

        These may only be my paranoia about this paragraph. May not reflect anything about the regular performance but just my 2 cents about Ryzen TR.

          • NTMBK
          • 2 years ago

          Who on earth uses Linux for gaming anyway? Just install Windows and get much better performance and larger game selection.

            • SkyWarrior
            • 2 years ago

            I don’t mean that people use linux for gaming but even for data analysis and extensive scientific computing memory compartmentalization and latency may be an issue for linux world. For running numerous small VMs with a few cores (2-4) and few gbs(4-8) of ram won’t be affected much (Unless linux kernel will be aware of filling empty whole cores not empty SMTs first to run instances. Thats why I think initial implementation will be a mess.) which I believe is the main usage scenario for TR and EPYC foreseen by AMD.

            • Anonymous Coward
            • 2 years ago

            If I had a big fat Linux Threadripper box with a good GPU, for work mind you, I’d be quite tempted to use it as a gaming box, and rebooting is bad. I don’t have much of a hardware budget for toys.

          • Anonymous Coward
          • 2 years ago

          I can’t imagine Linux kernel devs adding anything thats a mess here. I wonder if the NUMA vs fake-non-NUMA setting is just something they can flip on and off in the OS. NUMA has been on the desktop since the original Opteron so probably nothing new is happening here, except for some crap AMD is providing to coordinate it with games. I hope.

            • davidbowser
            • 2 years ago

            Agreed. I can’t fathom a Linux app that would barf with too many cores/threads.

            If I’m being picky, I wonder how a modern windows app would choke if there are too many threads. That seems like a pretty big fail on the software dev side.

        • exilon
        • 2 years ago

        NUMA vs fake-UMA mode is implemented in firmware so that will need an OS restart, I think.

        If game-mode is enabled through some extreme form of core parking, then it might not need a restart, but that doesn’t hide the 16 extra threads.

        Does Windows 10 have CPU hot-plugging support now?

    • Gadoran
    • 2 years ago

    A little boring this ThreadNeverEndingStory.
    What will be the end of this saga? Nothing. Not many will buy these things, not many are interested in HEDT. After all spending a lot less in cpu and in motherboard a poor guy can have the same performance on its loved games.

    So only a lot of Hype but where is the real disruptive SKU here?? Where is the SKU for laptops ?
    We all know desktop is dead definitively, it is only a niche segment today. All the volume is on mobile now.

      • Anonymous Coward
      • 2 years ago

      The survival of desktops must surely depend on processors like these, which would allow more interesting simulations to be run, but don’t fit in laptops.

        • Gadoran
        • 2 years ago

        Sorry but i have a lot of doubts about this.
        HEDT cpus are available from a lot of time at low prices, the best example were FX-8350 and FX-9590, eight cores, the best available in standard multithreading without strange AVX tricks and ‘intel’ optimizations. So the multicore simulations were available from years still nothing changed and the Pc market goes down slowly.
        We must to face the fact nearly all the people around know the Pc in the laptop form and few good guys with good ideas doesn’t help to save the boat. I don’t like this at all but we must to be realistic.

          • NTMBK
          • 2 years ago

          The 8350 had a slight lead in [i<]some[/i<] multithreaded benchmarks, but it absolutely sucked at per thread performance compared to the opposition.

            • Gadoran
            • 2 years ago

            The point is about 8 core simulations, this was available since years ago on 8350 with a good performance (higher than competition without optimizations) still this does not changed the market at all.
            Do you really believe the average user care something of ThreadTripper??

            • Anonymous Coward
            • 2 years ago

            The dozers had terrible performance and are therefore irrelevant to this discussion.

        • maxxcool
        • 2 years ago

        I as well will say these are dog and pony chips. Neat, but utterly non-important to AMD beyond press material. the cost they are in the SAME BOAT as intel with their old 999$ cpus.

        a TINY TINY few enthusiasts will buy one. Dell will advertise a bunch of them … but their overall sales will NOT move AMD’s needle.

          • Anonymous Coward
          • 2 years ago

          These are for workstations and for the foolish consumer, but thats only because no game exists that can employ 16 cores. If the software was there, AMD would ship these in volume, and not at $1k either.

          What comes to my mind is the experience I had with Minecraft. Kinda interesting but it doesn’t take too long to notice the world is paused anywhere you walk away from. Or that enemies appear at random despite your best efforts. Or that even the biggest castle is empty and pointless. A dead world of vast size.

          I imagine a game like that, but covered in systems of interacting beings, mining and building, all there to be organized into kingdoms and whatever one player or a small group of players want. Its easy to visualize the architecture for that. Hordes of threads or separate processes in which AI routines guide thousands of pigs and cows and zombies and witches and knights and whatever allover the landscape. Its a lot like a multiplayer server, except full of little AI agents doing their own things. I’d love to go on a quest to kill that stupid witch somewhere in the forest a couple days walk away, who keeps spreading her evil creations. Self-generating adventures. Dungeons being created by monsters even as you play. You’d rapidly arrive at barter economies between villagers, trade routes.

          All those words, I must be excited.

          Anyway, bring on the threads.

      • flip-mode
      • 2 years ago

      I guess if the product isn’t just what you wanted for yourself then it’s just not a good product for anyone.

    • Anonymous Coward
    • 2 years ago

    Pretty smooth. First they have kept the best dies for this. Second, they offer the 1900X to give people those quad memory channels and PCI lanes just one small tempting step above regular Ryzen.

    • odizzido
    • 2 years ago

    I would be interested in seeing some gaming performance at the same clock rate as the 7/5 CPUs to see if the extra potential ram bandwidth helps at all.

    • Pancake
    • 2 years ago

    Now, if it doesn’t turn out to be a buggy unreliable pile of fail a 1900X order will have my name on it. ECC seals the deal. Here’s hoping to it not being rubbish.

    • AMDisDEC
    • 2 years ago

    Great product roll out for AMD to enjoy a brief window of profitability before Intel responds with a superior solution.
    Hopefully, future AMD improvements on Zen will allow them to continue offering differentiated CPU solutions.
    Competition is a beautiful thing, and what makes America great!

    I hear Dr. Lisa Su will be ringside at the upcoming Mayweather- McGregor boxing match sitting alongside Warren Buffet.

      • NTMBK
      • 2 years ago

      … This is starting to get seriously creepy.

        • Pancake
        • 2 years ago

        I think it’s kinda sweet.

          • AMDisDEC
          • 2 years ago

          Finally, a brain with circulating fluid.

        • AMDisDEC
        • 2 years ago

        We hear this far too often;
        “I feared for my life”.

          • derFunkenstein
          • 2 years ago

          Should Lisa Su be fearing for hers?

            • cygnus1
            • 2 years ago

            Orrrr… is AMDisDEC actually Lisa Su doing some awesome trolling?

            • derFunkenstein
            • 2 years ago

            I think [url=https://s-media-cache-ak0.pinimg.com/736x/e4/0c/81/e40c814f113f4d605cef62c3d26f03fa–mind-blown-what-if.jpg<]this sums up my feelings[/url<].

            • cygnus1
            • 2 years ago

            Ha!

            • AMDisDEC
            • 2 years ago

            Trump 2020?

            • AMDisDEC
            • 2 years ago

            With the Donald Drumpf Capital Keystone Kops in the WH, everyone should.
            These incompetents are falling like flies. Like past AMD CEOs, until the board got smart and hired the talented Lisa Su.

        • Redocbew
        • 2 years ago

        The exact words seem to be unimportant. It’s like that Far Side cartoon about “what dogs hear”.

        What we say: Dude, what’s your deal with Lisa Su? Stop being such a creep already.
        What AMDisDEC hears: Blah blah blah, Lisa Su, blah blah blah blah.

          • AMDisDEC
          • 2 years ago

          You, the toothless kid in Stranger things.

    • Unknown-Error
    • 2 years ago

    Sounds promising. But word of caution to pre-orderers, DON’T. Wait for the reviews. Always wait for the reviews .

      • odizzido
      • 2 years ago

      words of wisdom. Works especially well for software preorders. You just don’t do it.

        • tacitust
        • 2 years ago

        But the hype…!

        • LostCat
        • 2 years ago

        *looks for Crackdown 3 preorder…can’t yet, aww*

      • tipoo
      • 2 years ago

      Preorders can always be returned, if you NEED a HEDT and think supply will be short, I don’t see why not.

    • DrDominodog51
    • 2 years ago

    I’m glad they learned from the cooler compatibility issues with Ryzen’s launch.

    On a completely unrelated note, does microsoft limit NUMA compatibility to more expensive versions of Windows?

      • juzz86
      • 2 years ago

      I’m still a little disappointed to not see a 120mm CLC in there by default – but the bracket is a neat idea as a working compromise. A few mates had Ryzen rigs laying around unfinished until OEMs had their AM4 brackets in-hand.

      Otherwise, seems promising. Looking forward to third-party reviews.

    • ronch
    • 2 years ago

    Looks like real trouble for Intel. All Intel can do at this point is thrash around like a crazed chicken whose eggs were poached. Just put the SKU tables from both houses side by side; Intel’s lineup looks like the product of a hastily called board meeting at 12am where everyone was screaming, pointing fingers, papers and staplers flying here and there. AMD’s lineup on the other hand obviously looks far far simpler and much easier to understand. It’s amazing how much in top form AMD seems to be these days, at least where products are concerned. Now they just need to keep the foot on the pedal.

    It’s also good to see that those extra dies are just dummy dies. It would be nuts for AMD to put actual working dies there just to be used as foot stools, even if they are crappy dies. Those dies can still be sold off as cheap OEM-only parts. I think anyone can still find use for even a 6-core or 4-core at <3GHz. Waste not want not.

      • Redocbew
      • 2 years ago

      [url=https://www.youtube.com/watch?v=lKie-vgUGdI<]There are dozens of us! Dozens![/url<]

      • tacitust
      • 2 years ago

      Way too early to call trouble for Intel. AMD is starting from way behind, and still has to prove it can deliver these processors in the volumes necessary to compete. We likely won’t know for another year at least whether AMD is going to succeed in cutting into Intel’s business in any meaningful way. It’s a good start, but it takes a lot more than a good product line to compete at this level.

      I hope AMD can prove itself to be a worthy long term competitor to Intel, since even a duopoly is a big step up from a near monopoly. No matter what the fanboys might want, the best result of this match up for us consumers would be a tie.

      • Klimax
      • 2 years ago

      No. And any surviving benchmarks will get devoured with second part of Intel’s lineup that’s still coming.

      And frankly, you are way overestimating impact of AMD’s offering on Intel. Those lineups are very rarely changed on such short notice short of far more defects or very large misjudgment of competitors offerings. Those lineups are classic Intel, nothing less or more.

      One more thing: According to discussion on Realworldtech, Skylake-X is still on original 14nm process not 14+ or 14++nm.

        • Zizy
        • 2 years ago

        Yeah I would hope that we get better performance with way more expensive parts. But if AMD’s claimed much higher efficiency is true vs this 10C, not sure if they will be really faster. With some assumptions that other Intel parts have same efficiency as calculated by cores*GHz/TDP, and that both TR 1950X and i9 7980XE run exactly at TDP, TR should win in performance, at slightly lower efficiency.
        So, quite many IFs here and I expect the 18C i9 to beat 16C TR in independent tests. But it better should at twice the price. You can get 32C Epyc for that kind of money.

        • synthtel2
        • 2 years ago

        Could I get a link for the RWT 14nm thing? That’s fascinating, if true. SKL-X still clocks a whole lot better than Broadwell, implying in that case that Skylake’s clock boost over Broadwell was architectural rather than a result of fixing problems with early 14nm.

      • ludi
      • 2 years ago

      Nah. AMD can pull rabits out of their hat and disrupt Intel’s product strategy for a generation or two, but it invariably takes them 5-10 years of marginal income (and a couple false starts) to design the next hat. Ryzen and Threadripper are incredible accomplishments, but they’ve revealed their whole hand. Intel has a weak hand for the next couple games but at the end of the day, they own the table, the dealer, and a majority stake in the house.

        • kuraegomon
        • 2 years ago

        A lot of that depends on just how many improvements they can make to Ryzen 2. There really appear to be some low-hanging fruit available that should let them close the IPC gap on Intel. It’s always easier to make optimizations to a newly-introduced architecture, as opposed to one that has been iterated on as many times as Intel’s (latest version of) Core.

        15 % IPC improvement doesn’t seem out of the realm of possibility for a Ryzen successor, and by that I mean 15% nearly across the board. If AMD can achieve that with the kinds of yields they already appear to be getting with Ryzen, and maintain or improve their power efficiency, then Intel will have to cede either some of their current market share or some of their profit margins. They don’t get to keep both – yay competition! 🙂

      • Freon
      • 2 years ago

      Intel seemed to rush out products as evidenced by X299 heat/power issues, but AMD’s solutions here are still clunky. They made one die and try to make it scale for every last solution. The 8 core game mode and latency penalties between cores/threads/dies is sort of a nightmare to deal with as a software engineer, and will require some hand tuning or special casing for AMD in some circumstances.

      I’m still fairly impressed with the value here, but I’m not sure I’d call it the pinnacle of smart engineering. It’s a small company throwing as much at the wall as they can, margins be damned. They’re still clearly way behind on R&D.

        • kuraegomon
        • 2 years ago

        Actually, it seems very likely that their margins on Ryzen are better than anything they’ve done in a while. Their design approach (and, presumably, tuning for their fabrication process) is resulting in excellent yields by all reports.

        As I said in my other post, lets see where a round or two of iterative optimizations on their initial design gets them, before passing judgement on the relative merits of the architecture. After all, word has it that the chief architect of the initial design knows what he’s doing. Apple and Intel certainly seem to think so, to name two 😉

        • cygnus1
        • 2 years ago

        It’s actually not that difficult. AMD is giving people the option of 1 or 2 NUMA nodes. I’m assuming based on how it was described, the 1 NUMA node mode most likely interleaves memory across the 2 die’s memory controllers so that all memory access has basically uniform performance/latency. In that mode there’s nothing for the software dev to do. For the better performing 2 NUMA node mode, a game process only needs to be NUMA aware so it can limit itself to the number of cores/threads from a single NUMA node. Obviously they’re going to have code in place to limit the number of threads so they don’t try to spawn too many, for instance more than a quad core can handle. Just need to throw in the test case for NUMA nodes to define that upper limit on threads…

          • willmore
          • 2 years ago

          You can already programatically query Windows about the CPU to find cores, threads, physical chips, etc. Any program sensitive to how it allocates CPU threads should already be doing this. If not, then that program is poorly written.

          I would shocked–shocked I tell you–to find out that games are poorly coded in this respect. </s>

            • cygnus1
            • 2 years ago

            Lol, I concur on the poorly codedv games part. I think a large part of that is how easy it is to build a game with existing engines. Inexperienced devs can easily bolt on crap code to a good engine and still have it mostly work ok. So I can definitely imagine there are plenty of games that query for hardware info but not necessarily being NUMA aware.

            Having more than 1 NUMA domain on a standard PC is pretty much unheard of, that’s server land. Consequently, game devs have never needed to worry about it before really. I don’t know the details of the exact calls they would use, but I would not be shocked if Windows can be queried for physical core/thread count and just spit back total for the system irrespective of the # of NUMA domains. I’m sure the code has to specifically look for that scenario and then configure the game process accordingly.

Pin It on Pinterest

Share This