AMD’s Ryzen Threadripper 2920X CPU reviewed

Throughout the year, AMD has been renewing its Ryzen CPUs with the Zen+ architecture, a combination of minor tweaks and process changes that promises better memory latency, more granular control over clocks in mixed workloads, and higher peak clock speeds overall. The Ryzen Threadripper 2950X was the first AMD high-end desktop CPU to benefit from this set of changes,  and now the company is bringing them to its new entry-level Threadripper: the 2920X. For $649, the 2920X offers a Ryzen 7 2700X-like 4.3-GHz peak clock speed plus a 3.5-GHz base clock.

Cores/

threads

Base

clock (GHz)

Peak boost

clock (GHz)

L2

cache (MB)

L3

cache (MB)

TDP Suggested

price

Threadripper 2990WX 32/64 3.0 4.2 16 64 250 W $1799
Threadripper 2970WX 24/48 12 64 $1299
Threadripper 2950X 16/32 3.5 4.4 8 32 180 W $899
Threadripper 1950X 16/32 3.4 4.2 8 32 $999
Threadripper 2920X 12/24 3.5 4.3 6 32 $649
Threadripper 1920X 12/24 3.5 4.2 6 32 $799
Threadripper 1900X 8/16 3.8 4.2 4 16 $549

While those figures may seem little changed from those of the Ryzen Threadripper 1920X, AMD’s Precision Boost 2 technology promises a more graceful descent to that base clock as cores and threads become loaded down. We’ve seen modest benefits from Precision Boost 2 and Threadrippers in our past testing, since many of our benchmarks load either one thread or all threads. For workloads that don’t run flat-out, like compiling, Precision Boost 2 could prove an asset to the 2920X. Let’s see if the combination of Zen+ tweaks and a lower price can help the 2920X stand out in an increasingly competitive high-end desktop space.

Our testing methods

As always, we did our best to deliver clean benchmarking numbers. We ran each benchmark at least three times and took the median of those results. Our test systems were configured as follows:

Processor Intel Core i7-8700K Intel Core i7-9700K Intel Core i9-9900K
CPU cooler Corsair H100i Pro 240-mm closed-loop liquid cooler
Motherboard Gigabyte Z390 Aorus Master
Chipset Intel Z390
Memory size 16 GB
Memory type G.Skill Flare X 16 GB (2x 8 GB) DDR4 SDRAM
Memory speed 3200 MT/s (actual)
Memory timings 14-14-14-34 2T
System drive Samsung 960 Pro 512 GB NVMe SSD
Processor AMD Ryzen 7 2700X AMD Ryzen 5 2600X
CPU cooler EK Predator 240-mm closed-loop liquid cooler
Motherboard Gigabyte X470 Aorus Gaming 7 Wifi
Chipset AMD X470
Memory size 16 GB
Memory type G.Skill Flare X 16 GB (2x 8 GB) DDR4 SDRAM
Memory speed 3200 MT/s (actual)
Memory timings 14-14-14-34 2T
System drive Samsung 960 EVO 500 GB NVMe SSD
Processor Threadripper 2950X Threadripper 1920X Threadripper 2920X Threadripper 2970WX Threadripper 2990WX
CPU cooler Enermax Liqtech TR4 240-mm closed-loop liquid cooler
Motherboard Gigabyte X399 Aorus Xtreme
Chipset AMD X399
Memory size 32 GB
Memory type G.Skill Flare X 32 GB (4x 8 GB) DDR4 SDRAM
Memory speed 3200 MT/s (actual)
Memory timings 14-14-14-34 1T
System drive Samsung 970 EVO 500 GB NVMe SSD
Processor Core i7-7820X Core i9-7900X Core i9-7960X Core i9-7980XE
CPU cooler Corsair H100i Pro 240-mm closed-loop liquid cooler
Motherboard Gigabyte X299 Designare EX
Chipset Intel X299
Memory size 32 GB
Memory type G.Skill Flare X 32 GB (4x 8 GB) DDR4 SDRAM
Memory speed 3200 MT/s (actual)
Memory timings 14-14-14-34 1T
System drive Intel 750 Series 400 GB NVMe SSD

Our test systems shared the following components:

Graphics card Nvidia GeForce RTX 2080 Ti Founders Edition
Graphics driver GeForce 411.63
Power supply Thermaltake Grand Gold 1200 W (AMD)

Seasonic Prime Platinum 1000 W (Intel)

Some other notes on our testing methods:

  • All test systems were updated with the latest firmware, graphics drivers, and Windows updates before we began collecting data, including patches for the Spectre and Meltdown vulnerabilities where applicable. As a result, test data from this review should not be compared with results collected in past TR reviews. Similarly, all applications used in the course of data collection were the most current versions available as of press time and cannot be used to cross-compare with older data.
  • Our test systems were all configured using the Windows Balanced power plan, including AMD systems that previously would have used the Ryzen Balanced plan. AMD’s suggested configuration for its CPUs no longer includes the Ryzen Balanced power plan as of Windows’ Fall Creators Update, also known as “RS3” or Redstone 3.
  • Unless otherwise noted, all productivity tests were conducted with a display resolution of 2560×1440 at 60 Hz. Gaming tests were conducted at 1920×1080 and 144 Hz.

Our testing methods are generally publicly available and reproducible. If you have any questions regarding our testing methods, feel free to leave a comment on this article or join us in the forums to discuss them.

 

Memory subsystem performance

The AIDA64 utility includes some basic tests of memory bandwidth and latency that will let us peer into the differences in behavior among the memory subsystems of the processors on the bench today, if there are any.

Some quick synthetic math tests

AIDA64 also includes some useful micro-benchmarks that we can use to flush out broad differences among CPUs on our bench. The PhotoWorxx test uses AVX2 instructions on all of these chips. The CPU Hash integer benchmark uses AVX and Ryzen CPUs’ Intel SHA Extensions support, while the single-precision FPU Julia and double-precision Mandel tests use AVX2 with FMA.

 

Javascript

The usefulness of Javascript microbenchmarks for comparing browser performance may be on the wane, but these tests still allow us to tease out some single-threaded performance differences among CPUs. As part of our transition to using the Mechanical TuRk to benchmark our chips, we’ve had to switch to Google’s Chrome browser so that we can automate these tests. Chrome does perform differently on these benchmarks than Microsoft Edge, our previous browser of choice, so it’s vitally important not to cross-compare these results with older TR reviews.

WebXPRT 3

The WebXPRT 3 benchmark is meant to simulate some realistic workloads one might encounter in web browsing. It’s here primarily as a counterweight to the more synthetic microbenchmarking tools above.

WebXPRT isn’t entirely single-threaded—it uses web workers to perform asynchronous execution of Javascript in some of its tests.

 

Compiling code with GCC

Our resident code monkey, Bruno Ferreira, helped us put together this code-compiling test. Qtbench records the time needed to compile the Qt SDK using the GCC compiler. The number of jobs dispatched by the Qtbench script is configurable, and we set the number of threads to match the hardware thread count for each CPU.

File compression with 7-Zip

The free and open-source 7-Zip archiving utility has a built-in benchmark that occupies every core and thread of the host system.

Disk encryption with Veracrypt

 

Cinebench

The evergreen Cinebench benchmark is powered by Maxon’s Cinema 4D rendering engine. It’s multithreaded and comes with a 64-bit executable. The test runs with a single thread and then with as many threads as possible.

Blender

Blender is a widely-used, open-source 3D modeling and rendering application. The app can take advantage of AVX2 instructions on compatible CPUs. We chose the “bmw27” test file from Blender’s selection of benchmark scenes to put our CPUs through their paces.

Corona

Corona, as its developers put it, is a “high-performance (un)biased photorealistic renderer, available for Autodesk 3ds Max and as a standalone CLI application, and in development for Maxon Cinema 4D.”

The company has made a standalone benchmark with its rendering engine inside, so it’s a no-brainer to give it a spin on these CPUs.

Indigo

Indigo Bench is a standalone application based on the Indigo rendering engine, which creates photo-realistic images using what its developers call “unbiased rendering technologies.”

V-Ray

 

Handbrake

Handbrake is a popular video-transcoding app that recently hit version 1.1.1. To see how it performs on these chips, we converted a roughly two-minute 4K source file from an iPhone 6S into a 1920×1080, 30 FPS MKV using the HEVC algorithm implemented in the x265 open-source encoder. We otherwise left the preset at its default settings.

SPECwpc WPCcfd

Computational fluid dynamics is an interesting and CPU-intensive benchmark. For years and years, we’ve used the Euler3D benchmark from Oklahoma State University’s CASElab, but that benchmark has become more and more difficult to continue justifying in today’s newly-competitive CPU landscape thanks to its compilation with Intel tools (and the resulting baked-in vendor advantage).

We set out to find a more vendor-neutral and up-to-date computational fluid dynamics benchmark than the wizened Euler3D. As it happens, the SPECwpc benchmark includes a CFD test constructed with Microsoft’s HPC Pack, the OpenFOAM toolkit, and the XiFoam solver. More information on XiFoam is available here. SPECwpc allows us to yoke every core and thread of our test systems for this benchmark.

SPECwpc NAMD

The SPECwpc benchmark also includes a Windows-ready implementation of NAMD. As its developers describe it, NAMD “is a parallel molecular dynamics code designed for high-performance simulation of large biomolecular systems. Based on Charm++ parallel objects, NAMD scales to hundreds of cores for typical simulations and beyond 500,000 cores for the largest simulations.” Our ambitions are considerably more modest, but NAMD seems an ideal benchmark for our many-core single-socket CPUs.

 

Digital audio workstation performance

After an extended hiatus, the duo of DAWBench project files—DSP 2017 and VI 2017—return to make our CPUs sweat. The DSP benchmark tests the raw number of VST plugins a system can handle, while the complex VI project simulates a virtual instrument and sampling workload.

A very special thanks is in order here for Native Instruments, who kindly provided us with the Kontakt licenses necessary to run the DAWBench VI project file. We greatly appreciate NI’s support—this benchmark would not have been possible without the help of the folks there. Be sure to check out their many fine digital audio products.

A very special thanks also to RME Audio, who cut us a deal on one of its Babyface Pro audio interfaces to assist us with our testing. RME’s hardware and software is legendary for its low latency and high quality, and the Babyface Pro has exemplified those virtues over the course of our time with it.

We used the latest version of the Reaper DAW for Windows as the platform for our tests. To simulate a demanding workload, we tested each CPU with a 24-bit depth and 96-KHz sampling rate, and at two ASIO buffer depths: 96, the lowest our interface will allow at a 96 KHz sampling rate, and 128. In response to popular demand, we’re also testing two buffer depths at a sampling rate of 48 KHz: 64 and 128. We added VSTs or notes of polyphony to each session until we started hearing popping or other audio artifacts.

Apologies for the lack of results at 96 KHz and a buffer depth of 96 here. Thanks to something in the chain of Reaper, Windows 10, and our ASIO driver, our many-core CPUs couldn’t run the 96-96 test at all—we got popping and crackling from the get-go.




 

Crysis 3

Even as it passes six years of age, Crysis 3 remains one of the most punishing games one can run. With an appetite for CPU performance and graphics power alike, this title remains a great way to put the performance of any gaming system in perspective.



 

Assassin’s Creed Odyssey

Ubisoft’s most recent Assassin’s Creed games have developed reputations as CPU hogs, so we grabbed Odyssey and put it to the test on our systems using a 1920×1080 resolution and the Ultra High preset.



 

Deus Ex: Mankind Divided

Thanks to its richly detailed environments and copious graphics settings, Deus Ex: Mankind Divided can punish graphics cards at high resolutions and make CPUs sweat at high refresh rates.



 

Grand Theft Auto V

Grand Theft Auto V‘s lavish simulation of Los Santos and surrounding locales can really put the hurt on a CPU, and we’re putting that characteristic to good use here.



 

Hitman

After an extended absence from our test suite thanks to a frame rate cap, Hitman is back. This game tends to max out a couple of threads but not every core on a chip, so it’s a good test of the intermediate parts of each processor’s frequency-scaling curve. We cranked the game’s graphics settings at 1920×1080 and got to testing.



 

Far Cry 5



 

Conclusions

AMD’s $649 Ryzen Threadripper 2920X launches during a turbulent time on the bridge between mainstream and high-end desktop PCs. Just a couple weeks ago, the 2920X would have been competing directly against Intel’s Core i7-7820X, and our results would have shown a favorable outcome for the entry-level second-generation Threadripper in that comparison. The problem for AMD is that the i7-7820X has been well and truly put out of a job by the $500-ish Core i9-9900K, a chip that is ostensibly quite a bit cheaper than both the 7820X and 2920X. We certainly wouldn’t recommend an i7-7820X today, but that doesn’t make the 2920X a winner by default.

That frightening boost in performance per dollar from the Intel corner means the i9-9900K and Threadripper 2920X end up trading blows throughout our productivity benchmarks. The i9-9900K again comes out looking like a superb chip for people who need copious multithreaded grunt alongside the best high-refresh-rate gaming, single-PC streaming, and single-threaded performance around. If it ever hits store shelves in volume and at its suggested price, the i9-9900K could be the clear pick in this range for people who don’t need the memory bandwidth and PCIe lanes of a high-end desktop platform, yet still want HEDT-class performance when they need to lean on a system. That stellar performance leaves the Threadripper 2920X less breathing room than past entry-level high-end desktop chips might have enjoyed.

There are still some things that only four channels of memory with ECC support and gobs of PCIe lanes can do, of course. If you have specific workloads that the 2920X rips through and the i9-9900K doesn’t, like rendering and scientific computing, you should obviously spring for the high-end desktop part. If you need every PCIe lane you can get for storage devices or other expansion cards, AMD’s no-limits X399 platform could also be a winner. If you only occasionally game and don’t chase high refresh rates in games at 1920×1080, or you have a 2560×1440 or 4K monitor instead, the i9-9900K’s advantages over the 2920X won’t be as obvious for gaming or same-PC streaming.

If I still sound a bit ambivalent about entry-level high-end desktop CPUs in the wake of the i9-9900K’s arrival, though, it’s because I am. Few enthusiasts are memory-bandwidth-bound or memory-capacity-limited, and even fewer are storage-bandwidth-constrained or short on PCIe lanes for doubling up on graphics cards. Folks straddling the line between mainstream and high-end desktop platforms need to be darn sure their workload needs what Threadrippers and X399 motherboards offer, because at the end of the day, it’s hard to look past the Core i9-9900K’s value in this price range. Even if volume shipments of i9-9900Ks don’t happen for some time, the Threadripper 2920X is hardly a direct substitute. PC builders need to choose the right tool for the job they’re trying to do, and that’s down to you, dear reader.

This latest chapter in high-end desktop performance is still being written, too. Intel is about to refresh its Skylake-X CPUs with solder TIM, the same 44 PCIe lanes from every CPU, higher clocks, and more cache. AMD has reignited the competitive fire in high-end CPU performance, and builders can only benefit by waiting and seeing whose chips are best-positioned to take the heat. If you want to step into Threadripper right now, the 2920X is a fine CPU at a fair price, but we’d advise waiting until we have the full story on all the latest high-end desktop chips before making any moves.

Comments closed
    • guyr
    • 11 months ago

    Many people here are discussing Threadripper as a gaming CPU. I’m not interested in gaming, but think Threadripper is a good professional workstation chip. I do professional software development. For that, I need lots of cores and threads and fast memory access. A typical backend Java or C++ project employs a database (say 8 threads for a local DBMS like Oracle or PostgreSQL), an IDE for writing the software (say 12 threads for parallel compiles) and a runtime environment (can be very different, but let’s just say 8 threads of a typical 3-tier architecture.) For normal business or scientific apps, 32 GB or RAM is good, and a single mid-tier GPU is good enough.

    There are a *lot* of software development jobs for which this setup will work well. As many have said, Threadripper is overkill for gaming; many chips at half the price will outperform TR.

      • K-L-Waster
      • 11 months ago

      I don’t think anyone is disputing those points.

      The question really is, does the 2920X make sense when for $250 more you can get 4 more cores and higher boost clocks from the 2950X.

        • guyr
        • 11 months ago

        @K-L-Waster, I agree. If I were upgrading at this point, I’d be looking at the 2950X, not the 2920X. Especially for professional work, no one is going to sweat the $250. I’ve got a still adequate dual Opteron 4234 system, so I’m going to wait until the 3000 series to see what that offers.

    • Fonbu
    • 12 months ago

    Needs a price cut!

    • ronch
    • 12 months ago

    This chip is like being stuck between the Civic and the Accord.

    • Unknown-Error
    • 12 months ago

    Yeah, not a very interesting CPU. 2950X and 2700X are AMD’s best offerings.

    • DancinJack
    • 12 months ago

    I’ve virtually zero interest in this CPU, but I just wanted to say this is a great review.

    • Jeff Kampman
    • 12 months ago

    I know it’s not the primary mission of these CPUs, but I just added full gaming data for all the Threadrippers we had under test today. Enjoy.

      • Anonymous Coward
      • 12 months ago

      I’m pretty impressed how a dual-die 12-core product like 2920X often enough beats its single-die 8-core brethren. Seems like the complexity (latency) and over-abundance of threads should ruin it.

        • Krogoth
        • 12 months ago

        It is the larger L3 cache pool that is saving 2920X in gaming benches albeit by a small margin. It barely makes up the difference in turbo and base speed that has 2700X though.

          • Anonymous Coward
          • 12 months ago

          I want to see a Ryzen die where they replace one CCX with a hunk of L3.

          • jihadjoe
          • 12 months ago

          Don’t they have the exact same 4.3GHz turbo speed? Plus the TR chip has two dies so it can run two cores at the peak turbo clock.

            • Anonymous Coward
            • 12 months ago

            I would have thought the trouble of asking for stuff from the other die’s RAM would kill whatever small edge there is in turbo speeds.

            I remember on-die memory controllers being a big deal back in the day. I have here a C2D E8500 (3.16ghz, 6MB L2, 1.3ghz FSB) and it doesn’t seem to fare too well against a dual-cored i5 with a meager 3MB L3 and turbo capped at 3.2ghz. Even web browsing on the C2D lacks the full smoothness of the i5, [i<]and[/i<] it has a better GPU installed, along with an SSD, both Win10. So long story short, I remain impressed that AMD's dual-die penalty is so low in latency-critical CPU-intensive things like gaming.

    • BooTs
    • 12 months ago

    Not sure if anyone else notices, but the number of mentions of “Intel” and ” i9-9900K” in the conclusions section for this review compared to the single mention of “AMD” in the conclusion section of the i9 review give off a bad impression.

    When you review the AMD part, you are comparing it against Intel.

    When you review the Intel part, you are comparing it against Intel.

    I’m sure it’s unintentional but I believe TR is still a place that can accept this small criticism.

      • derFunkenstein
      • 12 months ago

      Not sure how you came to this conclusion since the word “Threadripper” appears a dozen times and “AMD” appears 4x as often as you gave Jeff credit. “2920X” shows up another dozen. If you enjoy keeping score, please keep a complete score.

        • BooTs
        • 12 months ago

        Maybe I was not clear enough. What I was trying to point out is that in the 2920X review, conclusions section, Intel is mentioned heavily, but in the i9-9900k review, conclusions section, AMD is mentioned once.

        I’m not trying to troll, just expressing what appears to me to be a very small amount of bias.
        I’ve obviously been a TR fan for a long time. I just though it might be something to consider for a writing improvement – the quality of the TR reviews and data collections has always been above reproach.

          • derFunkenstein
          • 12 months ago

          OK I understand what you mean now. I think the answer to your question lies in the preceding pages of the i9-9900K review. At no point does any Ryzen desktop CPU pose even a remote threat to the 9900K. And it doesn’t have to, because total platform costs are 30% less than Intel’s. It’s comparing very fast cars to even faster, more expensive ones.

          OTOH if TR does publish a 9700K review, I would expect that Ryzen will come up a lot, because the Ryzen CPUs are much closer.

      • enixenigma
      • 12 months ago

      Perhaps because Intel is largely competing against itself with these new 9th-gen processors, especially so in gaming. When AMD brings its Zen 2 processors to bear, I think (hope?) you’ll see a lot more talk about AMD.

        • drfish
        • 12 months ago

        [quote<]If you're gonna compare a Hanzo sword.. ...you compare it to every other sword ever made - wasn't made - by Hattori Hanzo.[/quote<]

          • Krogoth
          • 12 months ago

          [url<]https://www.youtube.com/watch?v=RFV3JGLf7xg[/url<]

      • Beahmont
      • 12 months ago

      It’s fairly simple.

      There’s nothing that really competes with the 9900K. You are either over it in cost to get better in performance, or well under it in performance for most of the same price if it was going for MSRP. The only real question is if you need that much performance in a mainstream platform. If you do, then the choice is clear and AMD doesn’t really compete. There really isn’t that much to talk about on AMD’s part.

      The Threadripper 2920X has direct competition because of it’s twice as expensive and about the same performance as Intel’s mainstream 9900K and platform. There’s a real choice and competition to be had with each side having pluses and minuses. That means there’s a lot to talk about explaining how and why there is competition and a choice to be had based on the builder’s needs.

        • Srsly_Bro
        • 12 months ago

        The 9900k is sitting around $600 online if you can find it. Stop the fairy tale MSRP arguments. Still, at $600, I’d rather have that that the 2920x. it’s not twice as expensive. The CPU prices are about even right now and only the board will cost more but that depends if you get a Halo board. At MSRP, the 9900k wins, so lmk, bro, when we get it.

    • Chrispy_
    • 12 months ago

    The 2920X is competing against DDR4 prices more than anything else.

    You don’t buy into an HEDT platform to run it with just 16GB RAM. Most of the entry-level jobs that a Threadripper, Xeon, or X299-based i7 are purchased for are going to need 64GB, probably as a bare minimum. If you don’t need RAM, then you don’t need an HEDT platform in the first place.

    With the RAM costing $800-1000 for a typical 64GB of Samsung B-die 2933, and some configurations likely to require a $1600 memory kit, it seems pretty short-sighted and pointless to give up so much CPU performance over the 2950X for just $250, and if you can schedule your workload across threads effectively, the WX models are easy to justify.

    The real thorn in the side of the 2920X is the abundance of 1950X inventory. We just purchased 16 of them for less than the asking price of 2920X and until those stocks dry up, the non-WX Threadrippers don’t make any sense at all.

      • blastdoor
      • 12 months ago

      Agreed.

      The 2950X probably makes the most sense for the most people, with the 2970WX coming in second place. The 2990WX is super niche (not because of the core count but rather because of the RAM issues… the 2970WX also has RAM issues, but it’s cheaper); the 2920X doesn’t make sense for the reasons you describe.

      • dragontamer5788
      • 12 months ago

      I dunno, I think 32GB of RAM is reasonable for a Blender rig. There are certainly larger scenes which require more RAM, but a lot of 3d modeling can be done on 32GB of RAM. or even 16GB.

      If you’re primarily doing rigging / animations on scaled down models, you probably can make due with 16GB… but still benefit from higher-core counts whenever you preview the animation. That said: 64GB of RAM is super useful for rendering animations with multiple copies of Blender to maximize the CPU utilization of your rig.

      Ex: I’m pretty sure I measured the Gooseberry benchmark at 12GB of RAM last time I tried it.
      EDIT: Seems like [url=https://www.blendernation.com/2010/04/26/durian-render-farm-arrived-logo-design-comicon/<]Sintel / Durian was 6GB per instance[/url<] for their render farms (2010 numbers). (I heard on some blog it was 24GB but I guess I was mistaken before the edit) Another niche: GPU-based rendering productions will need the 60x PCIe lanes for 4x GPUs. But the actual CPU will barely be doing any work on something like [url=https://www.redshift3d.com/<]Redshift[/url<]. So you'd want as many PCIe lanes as possible, with as weak of a CPU as possible on those rendering farms. Since GPUs have very little RAM (ie: 1080 Ti farm will only have 12GB), you're naturally RAM limited in most scenes. I don't expect the CPU to require much more RAM than what is passed to the actual rendering engines in the GPU, so 32GB is probably acceptable (but if anyone actually uses Redshift and thinks otherwise, let me know!) EDIT: Just so that its clear: I just do Blender as a hobby thingy. I'm no pro. Nonetheless, I know Sintel / Durian is a 2010 era project which is probably why it uses less RAM, but it still demonstrates whats possible with under 6GB or so. I know its not fun to optimize models / vertex counts to get things to render... but its certainly doable... and probably still necessary to do on GPU-based rendering farms. [quote<]The real thorn in the side of the 2920X is the abundance of 1950X inventory.[/quote<] Agreed. As long as the 1950x exists at these prices, there is very little reason to buy the 2920x. -------- In any case, I think I agree with you that the 2950x (and 1950x, while it still is in stock) are the primary choices for most Threadripper builds. Still, the GPU-rendering situation really means that a "low end" CPU for the x399 chipset has its uses.

        • K-L-Waster
        • 12 months ago

        I think the conclusions still stand though. There are relatively few workflows that require quad channel RAM and loads of PCI lanes but also don’t require a lot of CPU. As you mention, some exist, but they’re edge cases for the most part.

        The 2920X doesn’t separate itself enough from the 9900K to make it a significant step up if thread count isn’t critical, and at the same time the 2950X is close enough in price that if you do need many threads you can get 25% more without breaking the bank. (And that’s before you consider clearance prices on 1950X’s.)

        By the time you’ve spent ~$1000 on RAM, ~$500 on a motherboard, plus whatever else on storage, add-in cards, and the beefy PSU you’re going to want, saving $250 on the CPU while giving up 25% of the threads compared to a 2950X seems like a false economy.

          • dragontamer5788
          • 12 months ago

          [quote<]By the time you've spent ~$1000 on RAM, ~$500 on a motherboard, plus whatever else on storage, add-in cards, and the beefy PSU you're going to want, saving $250 on the CPU while giving up 25% of the threads compared to a 2950X seems like a false economy. [/quote<] Wait. I think I was making a strong case for 32GB of RAM, with maybe 16GB for special cases. So $300 on 32GB of RAM, $300 motherboard. [quote<]The 2920X doesn't separate itself enough from the 9900K to make it a significant step up if thread count isn't critical[/quote<] 16-lanes GPU. 4-lanes for NVMe drive #1, 4-lanes for NVMe drive #2. And an expectation to use all three of those at once. IE: Video editing (read videos from NVMe drive #1, write output to NVMe drive #2, CUDA-acceleration on some effects). Bam, you're already beyond 16x PCIe + 4GB/s DMI that the 9900k offers. Strangely enough, video editing doesn't seem to be very core-count heavy. I think beyond 20-threads, video editing kinda caps off these days. I dunno what the bottleneck is, but my Threadripper 1950x only has 30% utilization when I'm video editing. EDIT: I think I should note that my video editing workflow has all my video in lossless UTVideo format at 1080p (and I'm considering a switch to lossless MagicYUV). H264 output [b<]may[/b<] be more CPU heavy, but I personally prefer lossless workflows. I'd imagine 4k video to be more I/O heavy, but I haven't really tested it. Not that RAM isn't useful: 32GB of RAM is useful for "render to RAM" previews. But you can easily just render to NVMe drives and stream it from I/O these days. I guess 8k video editing may require 4x memory channels and HUGE RAM to watch those previews. But your typical 1080p editing setup or even 4k editing setup should be fine with 16GB or 32GB of RAM... and instead push / pull things from multiple high-quality, speedy NVMe drives.

            • K-L-Waster
            • 12 months ago

            [quote<]16-lanes GPU. 4-lanes for NVMe drive #1, 4-lanes for NVMe drive #2. And an expectation to use all three of those at once. IE: Video editing (read videos from NVMe drive #1, write output to NVMe drive #2, CUDA-acceleration on some effects). Bam, you're already beyond 16x PCIe + 4GB/s DMI that the 9900k offers.[/quote<] Wait, why do you need to allocate 16 lanes to the GPU though? You've already established earlier that GPU performance isn't your primary focus in this example: why not put it in one of the 8 lane slots and free of the other 8 lanes for NVME? (Admittedly not all motherboards give you that flexibility, but if you can find one that does...)

            • dragontamer5788
            • 12 months ago

            [quote<]why not put it in one of the 8 lane slots and free of the other 8 lanes for NVME?[/quote<] That's possible to do of course, but it gimps the bandwidth to the GPU. You've gone from 15.8 GB/s (16x) to 7.9GB/s (8x) bandwidth to the GPU. While video games aren't very much affected by dropping down to 8x PCIe, the hypothetical was for CUDA-accelerated Adobe Premier, or maybe Davinci Resolve. I admit that I haven't tested 8x PCIe vs 16x PCIe with regards to CUDA-acceleration, but my ultimate point is that you have to build a compromise system in this simple case already. EDIT: Basic research online makes this seem like an unanswered question, unfortunately. [url<]https://forum.blackmagicdesign.com/viewtopic.php?f=21&t=32590[/url<] So I can't say if it matters for sure (or "when it matters", maybe it only matters at 4k or 8k resolution??) EDIT2: [url<]https://www.pugetsystems.com/pic_disp.php?id=40779[/url<] Seems like it makes a minor difference in this test. But also note: this system is running dual-Titan X (2016 test: [url<]https://www.pugetsystems.com/labs/articles/Titan-X-Performance-PCI-E-3-0-x8-vs-x16-851/).[/url<] Showing that Video Editing rigs can certainly take advantage of 2x GPUs for some effects.

            • K-L-Waster
            • 12 months ago

            Ok, but if you have a video editing need of all that GPU processing capability, wouldn’t you also want more CPU cores?

            You seem to be describing two different video editing scenarios: one where you don’t even need all of the cores a 2920X provides, and then a second one where GPU acceleration is beneficial. I’m sure both are legitimate, but I’m not convinced the same user is in both camps.

            • dragontamer5788
            • 12 months ago

            [quote<]You seem to be describing two different video editing scenarios: one where you don't even need all of the cores a 2920X provides, and then a second one where GPU acceleration is beneficial. I'm sure both are legitimate, but I'm not convinced the same user is in both camps. [/quote<] Its really weird. Video editing [b<]should[/b<] scale well to many CPU cores. You'd think its a massively parallel problem across many frames, many "tiles" of pictures and stuff... but in practice, there's a single-threaded bottleneck somewhere in there. [url<]https://www.pugetsystems.com/pic_disp.php?id=50696[/url<] [url<]https://www.pugetsystems.com/pic_disp.php?id=50616[/url<] Adobe just is bad at multithreaded coding or something (I have more experience personally with Magix Vegas 15, which also has a single-threaded bottleneck somewhere). Sure, more cores help, but not nearly as much as you'd expect. You can see that a 6-core i7-8700k is actually [b<]really good[/b<] choice (aside from the PCIe issue), and the i9-9900k is the best in the bottom benchmark. Nevertheless, some effects accelerate greatly with a GPU, but I don't think that these CUDA-accelerated filters are coded to work on multithreaded / CPUs. Yeah, they have a software fallback mode, but they're not exactly optimized. My guess is that Adobe just simply spent more time working on the CUDA code for those filters. Yeah, video editing is weird. There are some bad software architects in the field for sure. But this is the reality of the programs available to video editors today. Maybe in a few years they'll learn to code better or something. --------- Puget Systems overall recommendations: [url<]https://www.pugetsystems.com/recommended/Recommended-Systems-for-Adobe-Premiere-Pro-CC-143/Hardware-Recommendations[/url<] You can see that a GPU is useful for CUDA-effects, as well as multiple SSDs for the heavy amount of I/O you're going to be doing (input SSD + output SSD is a common, useful, configuration for video editors) ----------- Anyway, my overall point is that video editing (at least with Magix Vegas 15 and Adobe Premier 12.1.2) does NOT scale to many cores. So a 12-core Threadripper is fine. You'll [b<]still[/b<] want the 4x RAM channels and many-many PCIe lanes for high-speed CUDA and NVMe. Its counterintuitive, but that's just how it works for some reason.

            • Chrispy_
            • 12 months ago

            How effective/efficient is Adobe at scaling CUDA workloads though?

            If they are single-core bottlenecked on the CPU side, my suspicion is that they’re doing a piss-poor job of parallelism on the GPU side too.

            • dragontamer5788
            • 12 months ago

            Note: this is leaving what I know about, so I’m relying on reviews as opposed to experience to answer your question. Gamersnexus had the following review which only briefly looked at this question.

            [url<]https://www.gamersnexus.net/guides/3310-adobe-premiere-benchmarks-rendering-8700k-gpu-vs-ryzen[/url<] [quote<]How effective/efficient is Adobe at scaling CUDA workloads though? [/quote<] [url<]https://www.gamersnexus.net/images/media/2018/cpus/premiere/4_m22-review.png[/url<] We're looking at CUDA-acceleration providing runtimes roughly 10x faster than non-CUDA. Yes, a speedup on the order of 1000%. [quote<]Speaking of CUDA acceleration, we decided to torture ourselves just to give you an idea of how much that still matters. With CUDA disabled in our NZXT M22 review render, it took the 8700K 350 minutes – that’s 5.8 hours – to render the same clip that was finished in 34 minutes with software rendering and CUDA.[/quote<] [quote<]If they are single-core bottlenecked on the CPU side, my suspicion is that they're doing a piss-poor job of parallelism on the GPU side too.[/quote<] Yeah... you'd think that. But... Adobe programmers just want to prove you wrong or something. Now, I personally have a lot of questions about this result (was it using NVidia transcoder? What were the full settings? Etc. etc.). Nonetheless, it shows how Adobe heavily favors their CUDA code-base, at least with this workload. I'd obviously like to know more about it. But the evidence I've seen suggests that Adobe's CUDA (maybe OpenCL??) codebase is well optimized, but its CPU-codebase is not. Intel IGP acceleration makes a huge difference: cutting the render time from 34-minutes (CUDA-only) to 23-minutes (CUDA + Intel IGP). [quote<]Using hardware rendering on the IGP, we brought this down further to 21 minutes, a reduction from the 34-minute prior time of 38%[/quote<] So it is clear that Adobe is using OpenCL / CUDA very effectively for their workload. Since they're able to get such a huge difference from the i7-8700k iGPU (yes, Intel's [b<]iGPU[/b<] improves performance by 34%). What the heck??!?!? ------ Anyway, for now, Adobe Premier is single-core bound and scales well with CUDA / GPUs. Its a combination that really shouldn't exist, but that's just how the software works... at least for now. IMO, it just shows that Adobe's CPU engineers are bad or something. They really should be able to take advantage of core-counts better.

      • Bauxite
      • 12 months ago

      FWIW ECC B-die is $200 per 16GB stick and literally every one I’ve touched does 2933, even in 128GB configs. Probably a wiser buy for a 24/7 compute farm than blingy gamer ram.

        • Chrispy_
        • 12 months ago

        Yeah, we’re not actually running ECC, but our use case doesn’t really need it, since most of our stuff is just raytracing. The simulations are usually solar/weather sims too, and it’s all handwavium rather than scientific compute where accuracy to ninety decimal places is the difference between life or death.

        In saying that, B-die is in demand, so there doesn’t seem to be much of a premium on it which is good to know.

      • Waco
      • 12 months ago

      I don’t disagree (1950X vs 2920X) but I don’t know that I’d say that all applications that are CPU-heavy require lots of RAM. I do know that many workloads purport to requires lots of RAM, but ~2 GB/core seems to be pretty generous from my perspective.

      I’m struggling to utilize the 32 GB in my TR build for anything except a bigger ARC or a random VM that I don’t want to think about memory utilization on.

    • blastdoor
    • 12 months ago

    I see there are benchmarks for the 2970WX in the graphs — was there an earlier review of the 2970WX? I think I must have missed it… I’ve been very curious to see how that one performs, so nice to see some results for it!

    • Krogoth
    • 12 months ago

    2920X is basically a slightly to somewhat better version of 2700X at workstation/server workloads but has access to TR4 platform for almost twice the price point.

    I give 2920X a single Krogoth. It only makes sense if you need bandwidth and I/O throughput but Intel options are too much for your budget. While for whatever reason you do not want first-generation Threadrippers. Otherwise, the 9700K spanks it for nearly 1/2 of the price point for low-threaded workloads and isn’t too far behind in hilariously parallel stuff. 2700X is a strong contender for its price point.

    • Jigar
    • 12 months ago

    [quote=”Jeff Kampman”<][b<]the i9-9900K could be the clear pick in this range for people who don't need the memory bandwidth and PCIe lanes of a high-end desktop platform[/b<][/quote<] It might be only me but isn't the main purpose of HEDT to have all the PCIe lanes and memory bandwidth for future business needs ?

      • Kretschmer
      • 12 months ago

      Usually HEDT are purchased because system builders don’t understand diminishing returns on budget spend. There are a LOT of Newegg wishlists pairing HEDT platforms, low-core-count chips, and mediocre GPUs.

        • jihadjoe
        • 12 months ago

        Those builds sound to me like customers who actually might know what they’re doing: i.e. Lots of PCIe lanes for peripherals, without wasting too much money on the CPU or GPU.

          • K-L-Waster
          • 12 months ago

          How many non-GPU peripherals do you need to justify jumping to an HEDT platform? If you’re not using NVME storage or a platoon of NICs, you’d be hard pressed to do it. (Onboard USB sure the heck won’t…)

            • jihadjoe
            • 12 months ago

            Basically anything that captures a lot of data to dump onto discs (or rather NVMe). There’s a ton of possible scenarios for these.

            6k/8k video capture? A workstation capturing off a RED Rocket would be a good use of a low core count HEDT station since it will just need to facilitate DMA between capture cards and storage.

            And then the NICs+NVMe scenario you mentioned for servers handling incoming traffic.

            Very similar to #1 would be frame-grabber devices for stuff like medical and scientific imaging.

        • JustAnEngineer
        • 12 months ago

        You may have forgotten my grumbling over the September 2017 system guide’s selection of a Core i7-7740X and a super-expensive X299 motherboard for the “Sweet Spot”.

Pin It on Pinterest

Share This