Ryzen Threadripper 1900X hits the streets today for $549

AMD's Ryzen Threadripper 1950X and Ryzen Threadripper 1920X have been hogging the high-end desktop spotlight of late, but this morning, the Threadripper 1900X takes the stage. This entry-level ticket to the X399 platform pairs eight Zen cores with quad-channel memory and 64 PCIe lanes for expansion cards or storage for $549, and it's available today from Newegg, at Amazon, and from other online retailers.

  Cores Threads Base

clock

Cache

(L2/L3)

Four-core

boost clock

XFR

boost range

PCIe 3.0 lanes

from CPU

Memory

channels

TDP Price
Ryzen Threadripper

1950X

16 32 3.4 GHz 8MB/32MB 4.0 GHz 200 MHz 64 4 180W $999
Ryzen Threadripper

1920X

12 24 3.5 GHz 6MB/32MB $799
Ryzen Threadripper

1900X

8 16 3.8 GHz 4MB/16MB $549

Although the 1900X's basic specifications may sound similar to those of the Ryzen 7 1800X, the story isn't quite that simple. Like other Threadripper CPUs, the 1900X draws its processing power from two eight-core Zeppelin dies on a monster TR4 package. Each die has four cores and eight threads enabled, and they're connected across a diagonal Infinity Fabric topology. The 1900X has the same 4 MB of L2 cache and 16 MB of L3 cache as the Ryzen 7 1800X.

AMD also took advantage of the extra thermal headroom and top-5% die selection process for Threadripper to give the 1900X a slight speed boost over the Ryzen 7 1800X. The 1900X has a 3.8 GHz base clock, a 4 GHz boost speed, and a 4.2 GHz XFR clock for lightly-threaded workloads. Those are mild increases over the 3.6 GHz base clock, 4.0 GHz boost clock, and 4.1 GHz XFR speeds of the Ryzen 7 1800X, but every little bit counts.

AMD hasn't sent out Threadripper 1900X chips for review just yet, but the company did offer some of its own performance comparisons against the Ryzen 7 1800X for perspective. In 1920×1080 gaming (specifically, in the NUMA Game Mode), the 1900X trails the 1800X under Hitman and GTA V, but matches or beats it in Rise of the Tomb Raider and Deus Ex: Mankind Divided. No telling how the chip performs in its default UMA mode, but we'd expect fairly similar results.

Productivity performance is up a bit overall for the 1900X versus the 1800X, though, at least according to AMD. Using a Ryzen 7 1700 as a baseline, the company shows anywhere from a 4% to 10% performance boost over the 1800X from the eight-core Threadripper. We expect that quad-channel memory bandwidth and the extra PCIe lanes available on X399 will be the biggest draws relative to the Ryzen 7 1800X and the AM4 platform, but it is nice to get a little extra performance from the 1900X for the extra money, too.

In a separate bit of welcome news, AMD announced that it'll be bringing bootable NVMe RAID support to the X399 platform in a software update tentatively set for September 25. This free update will let builders boot RAID 0, RAID 1, or RAID 10 arrays of NVMe devices on X399 motherboards. As far as we can tell, this feature will be SSD-vendor-agnostic. Contrast AMD's free and vendor-agnostic approach with the Virtual RAID on CPU, or VROC, technology for Intel's X299 platform, and X399 could be all the more appealing for a high-end desktop.

From the details we can discern so far, VROC works best with a PCIe riser card that gets around the DMI 3.0 interface between Skylake-X CPUs and the X299 chipset. Most X299 boards draw their M.2 connectivity from the PCH, not the CPU. Use the four-M.2-slot card for VROC that Asus demonstrated to PCWorld, though, and you'd better have the 44-lane Core i9-7900X in the socket. 28-lane Skylake-X CPUs (that is to say, anything not a Core i9-7900X or better) generally won't be able to spare more than eight lanes of PCIe from the CPU while also running a graphics card with 16 lanes of PCIe 3.0. Even then, Intel only permits VROC operation in RAID 0. Other modes require an extra-cost dongle. Furthermore, bootable VROC RAIDs require Intel SSDs exclusively. Those are all minor annoyances, but they add up in light of what AMD is offering.

Even if Threadripper's NVMe RAID support isn't managed by the CPU like VROC, AMD's seemingly less fussy approach is another breath of fresh air for the X399 platform. Presuming builders need bootable NVMe RAID from their systems, X399 should be able to put another feather in its cap soon.

Comments closed
    • HisDivineOrder
    • 2 years ago

    For my purposes, I just don’t see the purpose of this product at the pricing it’s listed at. Based on the incredible pricing of the best Threadripper and the cheapest octacore Ryzen, I don’t see much value in any of the lower versions of Threadripper. Either you need cores or you don’t.

    So for me, it’s like Threadripper 16-core at $999 or Ryzen 8-core power efficient for $279 or performance focused for $389.

    If you do happen to need cores, you want more cores for a couple hundred over the hexacore, which is a couple hundred more than the strange TR octacore. If you don’t, you also probably don’t need more memory bandwidth. So…

    Yeah. I sense price drops for the lower versions of Threadripper in the not incredibly distant future.

      • frenchy2k1
      • 2 years ago

      This thread ripper gains massive memory bandwidth and PCIe lanes.
      This is a more specific case, but Threadripper is not *only* about number of cores.

      AMD is doing similar segmentation in their server line, offering EPYC processors with 4x2cores, offering less processing, but more memory and storage.

      • squngy
      • 2 years ago

      8 channel memory and 64 pcie lanes, yea there are definitely going to be a few people taking a close look at this.

      Some workloads don’t need that much CPU, but use a lot of memory and some people (crypto miners?) are just looking for the best way to shove as many pcie cards into one system as possible.

      • Anonymous Coward
      • 2 years ago

      IO, ECC, RAM capacity come to mind.

      Those are probably professional workloads and the difference in CPU price is going to be a small part of the whole deal, but it seems sensible enough to offer that savings.

    • NeelyCam
    • 2 years ago

    Can you unlock the disabled cores…?

      • chuckula
      • 2 years ago

      Unlocking disabled cores is old & busted.

      The new hotness is disabling cores that were turned on when you got the chip for moar performance. It’s called game mode.

        • ptsant
        • 2 years ago

        That wouldn’t be necessary if OS and game developers were better able to predict and manage core affinity. I understand that this is not the highest priority for game developers, but it’s doable in software.

          • smilingcrow
          • 2 years ago

          It’s not just about cores but also about memory latency.

    • AnotherReader
    • 2 years ago

    The 1900X is very tempting as an entry point for the X399 platform. Though the TDP is 180W, I would wager that its actual power consumption would be significantly lower than its bigger brethren.

      • Anonymous Coward
      • 2 years ago

      AMD may have needed the watts for those last few mhz. 🙂

      Agreed that it is important as an entry-level part, if nothing else. Maybe can push some volume to the platform.

    • DPete27
    • 2 years ago

    The 1900X…for those who are okay trading performance for PCIe lanes.

      • Wildchild
      • 2 years ago

      Guaranteed ECC memory support alone is worth it to a lot of people who would be looking at this. I say “guaranteed” because I know Ryzen unofficially supports it, but, from my understanding, that’s up to the motherboard manufacturer on whether or not they choose to support it.

      I have a hard time seeing individuals wanting to do serious work gambling on whether or not they chose the right motherboard and still missing out on the other features threadripper has to offer.

      • blahsaysblah
      • 2 years ago

      What? Who pays $500 for CPU to game at 1080p? If they do, they will be more than happy with those numbers as it wont be the primary reason for the build.

      • ptsant
      • 2 years ago

      The 1900X is faster than the 1800X in several workloads, including almost everything memory-bound, anything that benefits from a slightly higher base frequency and anything that benefits from the better XFR. In fact, the only situation that is penalized is when the process has to leap from module to module, which on Threadripper is bit costlier in memory latency (diagonal access, 4 modules on the chip etc) than on the 1800X. This will probably improve with more intelligent memory allocation and OS support.

      Memory capacity is also a key feature (128GB vs 64GB), as is ECC support.

      Finally, the price difference vs the 1800X is very small. The biggest differences are in motherboard prices (TR4 vs AM4).

        • Anonymous Coward
        • 2 years ago

        I very much want to see this thing reviewed. Lots of people complained at first about Ryzen having dual-channel memory, now its time to test that feature out, albeit with a latency penalty.

    • psuedonymous
    • 2 years ago

    [quote<]28-lane Skylake-X CPUs (that is to say, anything not a Core i9-7900X or better) generally won't be able to spare more than eight lanes of PCIe from the CPU while also running a graphics card with 16 lanes of PCIe 3.0. [/quote<] The who in the what now? 28 lanes is sufficient for an x16 slot plus a full three x4 m.2 slots all from the CPU. Indeed, that's exactly how the ASRock X299 ITX board is set up. If you motherboard has hooked up m.2 slots to the PCH on X299, then go blame the motherboard vendor for doing something silly.

      • Jeff Kampman
      • 2 years ago

      It’s sufficient if you don’t take the way most X299 motherboards split PCIe lanes into account. On ATX boards, most board makers have devoted CPU PCIe lanes from 28-lane chips to feeding a full PCIe 3.0 x16 slot in single-GPU configs, or an x16/x8/x4 slot config with multiple cards. If you’re trying to hang four M.2 devices off a 28-lane CPU, you need to use the primary PCIe x16 slot, which forces your graphics card into the secondary x8 slot. That may not be a real compromise for performance with many of today’s graphics cards, but it’s still a compromise.

      Given how tight this allotment gets for multi-GPU or expansion cards on its own, it’s no shock that board makers are tapping the chipset’s consistent offering of PCIe 3.0 resources for M.2 slots on X299.

      The ASRock X299-ITX/ac is the one exception (or one of the very few exceptions) to this rule because its designers didn’t have to consider the inconveniences of splitting different provisions of CPU PCIe lanes among multiple PCIe x16 slots. Designers of ATX boards don’t have that luxury, so chipset PCIe 3.0 lanes are the way it’s gotta be.

        • blahsaysblah
        • 2 years ago

        I think you need a little more sleep 🙂 AMD is listing 64 lanes which includes the their 4 PCH lanes. Intel doesnt. Too much math in too little time.

        He is pointing out that Intel math is actually 28 – 16 = 12 === 3 x 4

        edit: i know you know this, get some rest.

          • Jeff Kampman
          • 2 years ago

          My original analysis is correct, and I’m doing fine on sleep, thanks. While there are, in theory, 12 lanes remaining from a 28-lane CPU after you account for a PCIe 3.0 x16 graphics card, the way they’re wired/split on most motherboards cannot feed a three- or four-drive M.2 riser like the Asus one in the PCWorld article I linked with a full four lanes to each of those devices. Again, most motherboard makers are turning to the X299 chipset to get PCIe 3.0 lanes for M.2 devices.

          Asus’ Prime X299-Deluxe, the X299 board I have at hand, is specced to run with the following allocations with a 28-lane CPU on board:

          28-Lane CPU-
          2 x PCIe 3.0/2.0 x16 (x16, x16/x8)
          2 x PCIe 3.0/2.0 x16 (max at x4 mode)
          2 x PCIe 3.0/2.0 x1

          The M.2 slots on the board are specifically tied to the X299 chipset:

          Intel® X299 Chipset :
          1 x M.2 x4 Socket 3, with M key, type 2242/2260/2280 storage devices support (SATA & PCIE 3.0 x 4 mode)
          1 x M.2 x4 Socket 3, with M key, type 2242/2260/2280/22110 storage devices support (PCIE 3.0 x 4 mode)*3
          7 x SATA 6Gb/s port(s)*4
          1 x U.2 port

          Look at any ATX X299 board and you’ll see similar wiring.

            • ddarko
            • 2 years ago

            To add to Jeff’s observation, the X299 platform is forcing mobo makers to share/disable PCIe lanes or run them through the PCH bottleneck even if you have a CPU that gives you 44 lanes from the CPU.

            The Asus Rampage VI Extreme, the top of the line model that retails for $650, has four PCIe graphics slots and supports three M.2 devices. Great, right? Well, hmm, the third M.2 slot comes off the PCH and shares its lanes with the U.2 port so using one disables the other. The other two M.2 ports have their PCIe lanes from the CPU but only the first one has dedicated lanes in all scenarios. The second M.2 slot shares its lanes with the fourth PCIe graphics slot so that slot drops down to x4 mode if the second M.2 slot is used.

            Builds where the third M.2 and U.2 slots are both occupied or the two CPU-connected M.2 slots and four graphics cards are installed are obviously going to be extremely rare but none of these limitations are present on the X399 platform. It’s not even an issue on Intel’s X99 platform. Intel has really done a disservice to users of their HEDT platform with the X299, which is basically a rebranded Z170/270/370 chipset that costs more.

            • blahsaysblah
            • 2 years ago

            Oh i see, Arc doesnt list PCI-e root configurations in the new website format, you’re saying the Skylake-X root hub on the CPU only has 4 controllers(so maximum of 4 distinct devices), so it cant do like: 8, 8, 4, 4, 4.

            I would have thought its like any old CPU from forever:
            16 – 3 controllers, so 16 or 8 + 8 or 8 + 4 + 4 (not changed forever)
            8 – 2 controllers, 8 or 4 + 4
            4 – 1 controller

            edit: sorry, i still dont see why the PCI-E bifurcation would be so exotic. The root port configuration above makes most sense.
            I guess there is a pin limit, maybe
            16 – 2 controllers, 16 or 8 + 8
            8 – 2 controllers, 8 or 4 + 4
            4 – 1 controller

            16 – 3 controller, 16 or 8 + 8 or 8 + 4 + 4
            8 – 1 controller
            4 – 1 controller

            No matter how you dice it, there that last 4 has to have its own controller, and no matter what you do, you will add a minimum of two 4s or no fours. Not only one more four lane port. So its either one four lane or at least three 4x off the CPU.

            • MOSFET
            • 2 years ago

            [url=http://b2b.gigabyte.com/Server-Motherboard/MW51-HP0-rev-10#ov<]The block diagram on Gigabyte's site[/url<] for the C422 board announced yesterday is a decent illustrator of the lane-sharing and lane/bandwidth distribution. I know it's a slightly different chipset-CPU combo, but still...

            • blahsaysblah
            • 2 years ago

            Not really, its showing what giga-byte did for that board, which includes an add-on multi-plexor chip. Been trying to google the Intel datasheet for the PCI-e layout in Skylake-X but no such luck. AMDs Zen was easier

            edit: found it [url=https://www.intel.com/content/dam/www/public/us/en/documents/datasheets/6th-gen-x-series-datasheet-vol-1.pdf<]Intel Core™ X-Series Processor[/url<] Families

            • blahsaysblah
            • 2 years ago

            And according to that, they can support many more devices than i thought.

            • derFunkenstein
            • 2 years ago

            that’s really sucky for such a dearly expensive platform. You’re making the DMI 3.0 interface a bottleneck anytime you want to get data to main memory. Isn’t that just equivalent to four lanes of PCIe 3.0?

        • psuedonymous
        • 2 years ago

        I hadn’t looked at the ATX boards (too big), but they do appear to either not connect the m.2 slots to the CPU at all, or connect only two to the CPU. It’s a shame that the physically larger boards have less functionality.

    • Goty
    • 2 years ago

    I wish this part and the X399 platform had been available back in March when I built my current system. I don’t have any workstreams that would benefit from the massive number of threads of the 1920X and 1950X, but X399 is certainly appealing. I might have gone with a X399/1900X setup given the opportunity.

    • spiketheaardvark
    • 2 years ago

    These will provide an interesting test case comparing the increased latency cache and heterogeneous memory but with greater memory speed compared to the 1800x.

    • jts888
    • 2 years ago

    What happened to the rumored non-X 1920?

    $250 between the 1900X and 1920X seems like a big enough hole to drop a slightly lower clocked 12c chip, which would be more tempting in the $650 neighborhood than any of the other released models.

      • the
      • 2 years ago

      I would fathom that OEMs get the lower clocked 12 core chip first to help fill out their product line up. As disappointing as it is, end consumers are not AMD’s biggest customers but the big OEMs are. 🙁

      • jarder
      • 2 years ago

      Toms had a article on the non-x models a few weeks ago:
      [url<]http://www.tomshardware.com/news/amd-threadripper-1920-1950-1900,35165.html[/url<] It seems that they are in the pipeline, but with no official release date yet. As a patient tightwad who wants to run a lot of VM's on the cheap, I'm holding out for these.

        • jts888
        • 2 years ago

        Considering how well the non-X 1600 and 1700 have done, I do expect other people to be looking for their parallels in the Threadripper line.

        My personal line of reasoning is that while I’d like ~12c workstation now, I know that I’d be highly likely to upgrade to a Zen2/3 in a few years, so paying for something on the painful end of the curve for that short a useful life makes less sense for me.

      • ptsant
      • 2 years ago

      I don’t see the point of making a non-X chip at these prices. I do see the interest of a slightly slower 12-core chip at a lower price, but I expect the 1920X to gently come down in price.
      My predictions are 1920X eventually at $650-700 and the 1950X at $850. Not right now, but in a few months.

    • kuraegomon
    • 2 years ago

    The NVMe RAID news is heartening, for that _very_ small subset of folks with use cases that might demand it. The general philosophy that AMD’s displaying this CPU generation is even more heartening:

    “Hey, we know we’re not the incumbent, and we know we’re behind on IPC… so have every other possible feature as standard, and product segmentation be damned!”

    I’m still waiting to see reviews of the new functionality – and purported performance improvements 🙂 – when the AGESA updates arrive, but I really am leaning towards the 1950X now. My use case (microservices architecture with high memory and threadcount requirements, deployed in one or more VMs) is a very good fit.

      • jts888
      • 2 years ago

      I’m curious about who is really clamoring for bootable NVMe volumes.

      I suspect most folks could get away with a SATA SSD boot volume and a secondary NVMe RAID data volume, never perceive any difference in speed, and have a lot less install and driver headaches overall.

        • the
        • 2 years ago

        Bootable RAID1 does have its use-cases for people who have a need for data protection and uptime. Going hardware RAID1 [i<]should[/i<] provide a performance boost vs. a software RAID1. The more precise question is who would need RAID0 NVMe bootable. Typically RAID0 is done to enhance write speeds for things like video recording or high speed data capture. Ditto for RAID10.

        • ptsant
        • 2 years ago

        Booting from RAID1 is complicated. I use RAID1 for most of my media and I know that if something happens I will have to remake the system drive. For some people that would be unacceptable (downtime etc). So, having everything on RAID1, including the OS, can be useful.

        Having everything on transparent RAID1 is simply more elegant. I would do it, if I had the option but I wouldn’t pay into the platform for that alone. Still, when you start adding up the ECC, the 8 DIMM slots, the PCIe expandability and this it does make the transition appealing.

      • JustAnEngineer
      • 2 years ago

      Simple curiosity prompts me to ask:
      What is it about your use case that makes [url=http://www.amd.com/en/products/cpu/amd-ryzen-threadripper-1950x<]Ryzen Threadripper 1950X[/url<] a better fit than an [url=http://www.amd.com/system/files/2017-06/AMD-EPYC-Data-Sheet.pdf<]EPYC[/url<] CPU with up to twice as many threads?

        • kuraegomon
        • 2 years ago

        Honestly, budget and availability. I might eventually be able to justify either Epyc or a Skylake-X Xeon, but that isn’t the case today. Absolute single-threaded performance matters as well for certain microservices – which is why I haven’t completely ruled out an i9-7960X (once it’s available). That 2-core 4.2 GHz turbo looks quite promising.

      • Mr Bill
      • 2 years ago

      Quite appealing, if only I had a computational need to match it.
      [quote<]"Hey, we know we're not the incumbent, and we know we're behind on IPC... so have every other possible feature as standard, and product segmentation be damned!"[/quote<]

    • Alexko
    • 2 years ago

    That’s great, but what about reviews? I take it TR didn’t receive a sample, but even AnandTech doesn’t seem to have anything up.

      • chuckula
      • 2 years ago

      It does not appear that AMD sent out review samples since nobody seems to have a review of these chips. At least they were consistent though.

        • kuraegomon
        • 2 years ago

        Yay consistency? Seriously though, I did scratch my head a bit at this. My thinking is that they want to move as many units as possible up front to gullib–excuse me, [i<]eager[/i<] early adopters before the reviews hit and point out how relatively small (or non-existent, or negative) a delta exists between the TR 1900X and the R7 1800X.

          • Wildchild
          • 2 years ago

          It’s obvious that anyone who would choose this over the 1800X is looking at the extra features rather than the difference in performance.

          I can see this being popular for individuals wanting a workstation for personal use and don’t want to fork out a ton of money.

            • DrDominodog51
            • 2 years ago

            Allegedly, Threadripper has better binned dies than the 1800X as well, and can often reach 4.2 GHz on liquid cooling.

            • blahsaysblah
            • 2 years ago

            Anyone who would pay the premium for 1800X isnt looking for OC, they just want best no hassles experience. And the 1900X would now surpass that because you get all those extra PCIE and memory lanes to have no hassle adding whatever you want.

      • Jeff Kampman
      • 2 years ago

      As I noted in the piece, AMD didn’t sample anybody on one of these, to my knowledge. Not even Youtube stars got any to play with.

        • Mr Bill
        • 2 years ago

        Maybe we could crowdsource one for review. I’d pitch in $25 to see it reviewed by TR.

          • kuraegomon
          • 2 years ago

          Hmmm, I suspect a review sample will eventually make its way to TR. However, the crowdsource-for-rapid-review model (followed by eventual raffling off to the donor pool?) could be an interesting option – assuming actual retail availability.

Pin It on Pinterest

Share This