AMD Threadripper 1950X Trounces Core I9-7900X In Multithreading Benchmark (pcper.com)

dryriver writes: The Cinebench R15 benchmark is a popular tool for measuring how well CPUs cope with multithreaded compute loads. AMD's Threadripper 1950X 16 core CPU, priced at $999 according to AMD, benchmarks 41% faster in Cinebench R15 than Intel's also $999 10 core Core i9-7900X CPU. While Intel's Core i9-7900X scores 2186 points on Cinebench, AMD's Threadripper 1950X scores 3046 points. Even the cheaper 12 core $799 Threadripper 1920X is over 200 points faster in Cinebench R15 than Intel's Core i9-7900X. Intel has its own 16 core Core i9-7960X in the works, performance yet unknown, priced at $1,699, but AMD's 16 core part currently appears to be a full $700 cheaper than Intel's MSRP. It remaines to be seen who is faster in single-threaded performance -- Intel may take that crown --and what the power consumption of a fully loaded Threadripper looks like compared to its Core i9 counterpart.

Comments

What is the target for these?

[duke_cheetah2003]

I surely hope it's servers. These processors would be silly in a desktop computer. We're not even fully loading down 2-8 core machines now. Gaming performance has and still is a single core endeavor, and even now, most of my stuff has trouble pegging any cores to 100% for any length of time.

About the only thing I do that consumes a lot of cpu time is compiling. Not very many computer users compile stuff.

Again, it's ultimate more of the same lackluster improvements. Throwing more threads/core at stuff, when it's still who's got the FASTER single core that matters at the end of the day. At least in my opinion.

For servers however, running virtualization stuff, these CPU's should be great, squeeze even more out every physical server unit.

Re:What is the target for these?

[corychristison]

High end workstations, that traditionally are built with server CPUs.

I suspect the next generation of AMD's CPUs will be based on these.

Re:What is the target for these?

[willy_me]

I surely hope it's servers. These processors would be silly in a desktop computer.

The quad-channel memory could help in a lot of situations. There will be plenty of applications that can benefit. But the 12/24 core model could be the better choice for many.

Re:What is the target for these?

[Anonymous Coward]

Some games do indeed use many cores. Some RTS games can use them (usually the types with lots of units) and even Battlefield 1 correctly loads all 12 of my threads - 6 cores 80-90% and 6 threads ~30% - just about the ideal load for a 6/12 CPU to achieve max throughput without starving any threads. (BF1 was maxing out all 8 threads on my old i7-2600 so it does like CPU, current CPU is an i7-6850k @ 4.3Ghz for reference).

Most of my games are not single threaded and those that are are not CPU bound (often not even GPU bound), granted most of my games also don't need more than 4 cores (most become GPU limited with 4 threads under heavy load - most really just have 1 medium loaded core and a 2-6 lightly loaded threads which is well within the capacity of any 4 core 4 thread CPU). Most game only use multiple cores because it has them not because it needs them - CPU can run at lower clocks if it only needs 20% of 8 core instead 40% of 4 cores.

I really don't see why people seem to want single threaded performance - single thread has been fast enough for a long time now. Anything that really needs more cycles than a single thread can give has long since been made into multi-threaded apps. All the remaining single threaded apps tend to not be CPU bound anyway (poor implementations excluded - you can always burn cycles if you really want to) so who cares if the app now only needs 20% instead of 25% of one core (unless you are talking about power constrained devices like laptops but then the argument isn't so much about raw speed but efficiency).

So while I agree, a home server with this would be awesome (actually thinking about it once the tech stabilizes a bit) I also would love a nice gaming rig with this (this is actually what I was waiting for when my 2600 died a while back - I needed a new PC and couldn't wait long enough for this come out so the 6850k it was).

Re: What is the target for these?

[koomba]

The quad channel memory most definitely will help, even more so than than on Intels new processors.

There are many reasons, but one crucial one is the very architecture of these new Threadripper CPUs. These higher core count processors are literally multiple lower core Ryzen chips in one die.

I won't get into any pros or cons of that aspect, but just mention it to explain the significance of the quad channel DDR4. The way AMD has designed these smaller "packages" to work together as one CPU is, to put it very simply, to have them communicate through the DDR4 bus.

This is significantly different than Intels so-called ring bus, or uncore. So it's a pretty big change for HEDT users who have essentially been exclusively using Intel since around 2007.

I don't claim to know every technical detail of TR/Ryzen, but I do know the end result of this is that your DDR4 memory speed with TR can have a large impact on performance. In particular, running higher than the official platform speed memory, or just overclocking above standard gives very nice increases in many scenarios.

I'm on mobile so I can't look it up right now, so don't hold be to this, but I remember what I saw being something like maybe 20-30%ish(?) improvement in some gaming benchmarks I believe. And there were a couple others that I only remember the numbers right now, but it was around 30% and 40% even in one case.

So I think that couple persuade a decent part of the HEDT community who isn't super hardcore, and probably games more or as much anyways as they utilize the massive multithreaded advantages of a HEDT platform. So I think that's pretty exciting, and gives TR at least a decent ranking in parts of the HEDT user base.

Re:What is the target for these?

[ckatko]

> We're not even fully loading down 2-8 core machines now. Gaming performance has and still is a single core endeavor, and even now, most of my stuff has trouble pegging any cores to 100% for any length of time.

1 - I load up my 8 core machine every day.

2 - Gaming is not single threaded unless you're an idiot or living in 1993. At the very least, physics can run separate from display, and every modern game on the planet runs at least 1 frame lag for that same reason.

3 - Vulkan is designed from the ground up to utilize ALL cpu cores as well as multiple GPUs.

Just because YOU can't use your computer while playing Farmville, doesn't mean everyone else on the planet is incapable of doing so. I literally just finished playing a game that uses 100% GPU, 6 cores, and the other 2 I used for encoding the video recording. But nah, fuck it. 640k is enough for everyone, amirite?

Re:What is the target for these?

[somenickname]

when it's still who's got the FASTER single core that matters at the end of the day.

I'm not sure how much single core performance even matters at this point. My work machine is a modern i7 clocked at some crazy high speed while my home machines are a ThinkPad X220 and an old school dual Xeon X5690 setup. The work machine is actually worse than my old dual Xeon setup on multi-threaded stuff and, for a single core, is indistinguishable for real life performance. Yes, it compiles a lot faster than my X220 but, if I weren't compiling stuff, I wouldn't know the difference between the machines.

Re:What is the target for these?

[Ramze]

They're also useful for video encoding, animation, multimedia production, simulation, and AI.

Have you ever tried to transcode MPEG2 video to x.265 or VP9 on a desktop PC? 2 hrs of VHS-quality video can turn into 10 hours of transcoding easily on a 4core/8thread PC. Transcoding 1080p or 4K from MPEG2 or MPEG4 to HEVC can take even longer. Lots of art school students use animation on their home laptops, plenty of people work with video encoding and online streaming at home, too.

Gaming is mostly a GPU-bound task, but these also have a lot of PCIe lanes to help with that, and lots of games are being compiled for multi-cpu now.

That's great if you can do everything you need with what you have. I'd say that's the case for most people. I know some who do everything they need at home on their cell phones and/or tablets, but other people have different use-cases.

Re: What is the target for these?

[red_dragon]

"I won't get into any pros or cons of that aspect, but just mention it to explain the significance of the quad channel DDR4. The way AMD has designed these smaller "packages" to work together as one CPU is, to put it very simply, to have them communicate through the DDR4 bus." Wait... no, that's not right. The cores talk to each other via Infinity Fabric. To talk to cores on a separate module or to access memory managed by a different memory controller (Threadripper has two, Epyc has four), Infinity Fabric uses PCIe inside the package. Epyc dedicated 64 lanes for this purpose, so I assume that Threadripper uses 32 lanes. The memory buses never come into it.

Video Transcoding is one possibility...

[CraigCruden]

I managed to peg my 8 core Xeon at nearly 100% CPU usage for about 6 months straight - 7 days a week, 24 hours a day doing video transcoding on a library. But yes, any computer with more than 2 cores is really a niche computer these days since 90+ of the people run computers with CPUs idling 90%+. The $100 Ryzen 3 will be more than enough power for the masses. The greatest "performance" boost for personal computers in the last few years -- for the masses -- has been flash based SSDs...

Re:Single threaded??

[WilliamGeorge]

Those ARM chips only cost a fraction of the AMD and Intel processors, though, and in any sort of real computation are only a fraction of the speed as well (and power consumption, to be fair). Intel and AMD should stick to the big, powerful, and expensive chips.

Moreover, there are plenty of single-threaded workloads. Most modeling applications, for example. True, you wouldn't really want a single core processor - but you are far better off with a 4 core at high clock speed than an 8, 12, or 16 core. It will both cost less and actually perform faster, thanks to better clock speeds.

These AMD 12-16 core chips will have their place, no doubt, as do ARM processors... but it really is important to pick the right processor for your workload. There is no single best processor for everything.

Re: so, will Intel finally become affordable?

[koomba]

Definitely still the more expensive option, but AMD has put some pressure on them and forced them to be a little more price competitive.

This new x299 platform of Intels HEDT gives us an 8-core/16-thread at the $600 mark, with the 7820x. On the previous x99 platform the $600 slot was only 6 core; you had to jump all the way to the $1000 tier for 8 cores in the form of the 5960x and then the 6900k. So I am definitely happy about that.

Although being Intel, they're incapable of giving us something without us paying for it somehow. In this case, to get the full 44 PCIe lanes you DO have to go to the $1k tier, whereas on x99 only the very cheapest CPUs had the gimped 28 lanes, but they were only $350 processors. The $600 level or above got you the full 28 lanes.
br I That part really irks me, and is making me not so certain I want that 7820x after all. I was 100% in for a 7820x on the nice new platform until I saw that. So for now I'm still considering my options. I really don't want or need 10 cores, and I'd much rather put the $400 difference towards multiple NVMe storage buys.

Compilation flags & vectorization

[modrzej]

The article is silent about vectorization, and Intel invests a lot in that lately. Do we know anything about the compilation flags of that copy of cinebench? If not, the assessment could be extremely unfair. A newer set of vectorization instructions corresponds to a longer vector size for arithmetic operations that can be carried out concurrently. For example, in HPC applications, enabling the highest available level of AVX can lead to 2x gains compared to code compiled for legacy systems.