Need help comparing/contrasting high end CPUs

January 27, 2017 7:46 AM

Building a high end desktop, starting with CPUs in the $1000 range. Here's three that several tech websites rate highly and thoroughly review, albeit in a language of (to me) inscrutable abbreviations, acronyms, and statistics.

1) Intel i7-980X Extreme

6 cores
3.33 GHz
12 MB cache

2) Intel Xeon E5-2650V3

10 cores
2.30 GHZ
25 MB Cache

3) AMD Opteron 6378 Abu Dhabi

(link doesn't work. Not sure why)

16 Cores
2.4 GHz
16 MB L2 Cache
16 MB L3 Cache

I sort know what cores, clock speed, and cache mean by themselves, but don't know how to use them together to draw conclusions about the absolute and relative performance of individual CPU's.

I ask myself but can't answer

--What's "faster": fewer cores at a higher clock speed, or more cores at a lower clock speed?

--At a given clock speed and given number of cores, how does having a larger/smaller cache affect performance?

--Do I need to really need to go into all this minutiae, or can I assume that any chip that costs $1000 will kick ass, and that it really doesn't which particular chip I pick?

I hope these questions give the hive an idea of what sort of info and advice I'm looking for.

Thanks

posted by BadgerDoctor to Computers & Internet (27 answers total)

Two things:

1) This is what sites like "Tom's Hardware" are for--for comparing different procs of different cores and speeds via controlled bench-marking. The results are laid out in pretty easy to understand terms.

2) Depends what are you going to do with your PC: Would your main applications make use of all 10 cores of the Xenon, or will they only make use of 2 or 4 of the cores, for example. (Real life example, albeit quite dated: I remember way back in the Flight Sim X days when it was still single-core only. My single core P4 out-performed the "much faster" dual cores that were being introduced, as each core of the dual core proc was far slower than my one single core. Extrapolate that into today's applications, etc.)
posted by TinWhistle at 7:53 AM on January 27, 2017 [3 favorites]

You really shouldn't be spending $1000 on a CPU if you don't already know the answers, to be frank.

It also (very much) depends on what you're doing. If I were building a desktop for general use, or games, I would buy a ~$400 i7. If I were building a workstation for my (virtualization heavy) professional workload, I would buy a high-core-count, recent Xeon. Throwing $1000 at the wall and hoping something good sticks isn't the best idea.
posted by so fucking future at 7:54 AM on January 27, 2017 [16 favorites]

Unless you have a very specific workload in mind for this computer, $1,000 is almost certainly too much to spend on the CPU itself.

If you *do* have a specific workload in mind, you will need to tell us what that is - what is it that you want the computer to be fast at doing - before anyone can give you a reasonable answer to your question.
posted by Juffo-Wup at 8:43 AM on January 27, 2017 [1 favorite]

That i7-980x is six years old. It's a terrible deal. You want to look at something like this:

http://www.tomshardware.com/reviews/best-cpus,3986.html

Take a good look at the i7-7700k. It is much, much faster than the relic you posted, and only costs $350ish.
posted by BrunoLatourFanclub at 9:28 AM on January 27, 2017 [6 favorites]

Seconding Bruno here. If you absolutely must have more than 4 cores (for media encoding, video editing, etc.) Intel has a one-generation-old part, the i7-6850k for $650 that should do what you need. If it's just games though, get the 7700k and put the extra money into a bigger/better GPU.
posted by Oktober at 9:41 AM on January 27, 2017

I lost a long-ish post all about the absence of corrolation between dollars and processor cycles. Please, BadgerDoctor, go ahead and threadsit a little bit, and tell us about the use-case for your desktop. I suspect that you are trying to future-proof your machine by spending lots of money. There are ways to do that, but no $1k processor is going to do that for you, no matter when it was released.

(I basically endorse Oktober's post, but I don't really game, and have been subsisting on integrated graphics for years. I would spend lots of extra money on a bleeding-edge motherboard if I was to try to future-proof my machine with money. Both processor and GPU could be upgraded in the future, if you bought the latest-and-greatest mobo with the latest-and-greatest chipset. But the latest-and-greatest is always overpriced, because the manufacturers are milking the enthusiast market.)
posted by BrunoLatourFanclub at 9:55 AM on January 27, 2017

I built a gaming box from scratch with a $150 CPU, a decent SSD, and a top of the line graphics card for about 50% more what you are budgeting for just the CPU, and it can play the most 3D intensive games at a gazillion frames per second.

Whether you need to drop a grand on a CPU depends very much on what you're doing; with no description of the rest of the box or the application, I would not recommend blowing your budget on what sounds like a server-class CPU.
posted by zippy at 9:55 AM on January 27, 2017

If it's just games, you will be fine with a 7700 or any of the Kaby Lake i5s. The i7s carry a premium because they support hyperthreading and the 7700k carries an additional premium because its multiplier is unlocked, but you haven't mentioned overclocking.

As others have written, it will be much easier to give you advice with more information about what you plan to use this computer for.

Buying last-generation processors is sometimes justified, but buying super-old chips like the ones you posted is a terrible deal. I'm not sure you can even buy moderm motherboards with such old sockets.
posted by lozierj at 9:58 AM on January 27, 2017

Best answer: Because of the magic of pipelining and out-of-order execution and all the other tricks that CPU makers have come up with to make CPUs faster without increasing clock speeds, it's almost impossible to look at the specs and figure out which one will be faster.

Your best bet is to use the results of comprehensive benchmarks, like those at Anandtech, Tom's Hardware, or Passmark. They each have multiple benchmarks, because different workloads will be faster or slower on different CPUs. Try to find a benchmark that's most like the workload you expect to have. General office work? Media creation? Data analysis? When you're looking at the benchmarks, note that many of them give different results for single-core vs. multi-core applications. Find out: Can your applications actually use multiple cores?

And you didn't ask about this, but just in case it's something you haven't considered: For many workloads, the limit on performance is getting data to and from your CPU, not your CPU itself. There are some workloads that can fit inside a CPU's cache, but they're relatively rare. More often, your performance will be hurt by a) not enough RAM and b) slow disk.

Unless you have a very specific CPU-dependent workload, you'll probably be better off buying one of the ~$300 CPUs that others have recommended and spend the $700 difference on more RAM, a faster SSD, and - if you need graphical performance - your video card.
posted by clawsoon at 10:07 AM on January 27, 2017 [1 favorite]

Best answer: A question that might lead to answers that'll be more helpful for you, even if they're not exactly what you asked: When is your current desktop too slow? What are you doing when it's slow? How many (and what kinds) of applications do you have open? What does Task Manager tell you about your CPU, network, memory and disk usage when your computer is slow?
posted by clawsoon at 10:13 AM on January 27, 2017

A tiny bit of info that may be helpful for you -- if you're putting together a machine for gaming, I wouldn't suggest the Xeon. Although they're well-built and powerful, they're unusual enough that we've run into problems with ours in the context of gaming. Nothing insurmountable, but sometimes you have to troubleshoot something for ages to figure out what weird quirk of the processor is causing a problem, and do weird things w/ game config files to fix it.
posted by nosila at 10:17 AM on January 27, 2017

Bruno's on the nose as far as CPU recommendations go. You don't really need to go well above $400 for it to kick ass, thankfully! Any more than that, you top out, and get diminishing returns. Unless... you use some serious number-crunching applications (e.g., CAD design, 4K video, 3D rendering), you'll be at a loss to perceive a difference, if all your other components are high-end. Let us know if this is the desired usage for you, though; we might then advise a model of Xeon and speak to some of its benefits.

Save your money here. For perception of intense speed with a top i7, invest into hard drives, RAM, and possibly a graphics card. The new best thing is m.2 hard drives (try the new Samsung 960 Pro variant, or the 950, as either are fantastic choices), and be sure to top out your RAM with DDR4. No, it's not a huge improvement from DDR3, but if we're loading you out, might as well. :) Go for a great graphics card if you'd like (use that $600 CPU savings on a GTX1080! - the Ti version may come out Q2 2017.)
posted by a good beginning at 10:20 AM on January 27, 2017 [1 favorite]

Response by poster: Thank you for the answers so far. In reply

1) Context/task: I'm a medical doctor working with San Diego County to develop a new, comprehensive, educational website for both practicing EMT's and Paramedics, and students studying paramedicine and/or to become an EMT.

One goal of the website is to offer an exquisite suite of interactive animations, anatomical and physiological models, and 3D rendered images derived from real-word medical imaging--for example, creating a heart from a CT scan, converting high quality MRI images into 3d Mesh Models of soft tissues.

My current computer cannot run the imaging/modelling software needed to develop the visuals for the website

2) Price is not an issue; I make an excellent living. Plus, buying a high end CPU now means I won't need to buy another for at least a couple years
posted by BadgerDoctor at 10:32 AM on January 27, 2017 [2 favorites]

Go to Logical Increments, and scroll down to the row that says "Monstrous". Buy that.
posted by gregr at 10:48 AM on January 27, 2017 [5 favorites]

Definitely go for the Xeon if money is no object. It's among the very fastest CPUs available. You could get more cores, for a roughly linear increase in performance on the kind of computing you're doing, but it gets much more expensive -- a 24-core version costs about $8000.

Geekbench is a good source of performance data from hobbyists, and SPEC is the professional version.

Like everyone else said, your GPU, amount of RAM, SSD, and other factors make a way bigger difference than the CPU itself.
posted by miyabo at 10:49 AM on January 27, 2017

Oh and to compare CPUs just jump on google and search for "cpubenchmark i7-6950x" and click on the link that goes to 'www.cpubenchmark.net'. Look at the big red number 'Average CPU Mark'.
posted by gregr at 10:50 AM on January 27, 2017

Buy a computer from Falcon Northwest.
posted by oceanjesse at 10:57 AM on January 27, 2017

Best answer: Given your suggested workload (mostly rendering), you should see if your software supports a GPU. If it can, buy any server-quality CPU, and invest your money on the GPU. (Nvidia is the current leader in professional GPUs).

An easy (and low-commitment way) to investigate, is to open an account with AWS and test your software on an instance that has a GPU. Their P2 and G2 family have Nvidia GPUs (and Xeon E5-2686v4 or Xeon E5-2670 CPUs, respectively). Their C4 family has no GPU, but has a high quality (Xeon E5-2666v3) CPU.

AWS isn't the cheapest path forward, long-term, but they provide a lot of value. And if you're unhappy with the results, you can stop immediately, because they bill by the hour.
posted by whisk(e)y neat at 11:05 AM on January 27, 2017 [3 favorites]

Best answer: Most of the advice here--including my other post--is for video games; scientific rendering is a completely different ballgame where it is completely justified to spend thousands on a server CPU (but not the one you posted: it's from 2012.)

Many fewer people run rigs like this, so advice might be a little harder to come by (though Metafilter never ceases to amaze).

I assume you're using niche software. I'd suggest getting your exorbitant license fee's worth by calling up the vendors and asking their advice. Some software may by CPU bound, some might utilize the GPU, some might work best on server architectures, some might benefit from a multi-CPU solution, some might work best on a farm of several cheap computers.
posted by lozierj at 11:06 AM on January 27, 2017 [6 favorites]

Best answer: Absolutely agree with lozierj.

Having worked with medical imaging... call you vendor, ask if (1) they have a preferred vendor/configuration for their systems and if not (2) what are their recommendations, especially (3) which GPU configurations have they tested their package with.
posted by Oktober at 11:34 AM on January 27, 2017 [2 favorites]

When you say "My current computer cannot run the imaging/modelling software," do you mean "my computer is unacceptably slow running the software" or do you mean "the software doesn't run at all"?

If it's the latter, you are likely missing a critical feature like ECC RAM or discrete graphics. Contacting the vendor is the best route to figuring this out.
posted by lozierj at 11:57 AM on January 27, 2017 [1 favorite]

Thanks for your follow-up answer. It definitely helps!

I'll third lozierj and Oktober; talk to your software vendor first. They may have some limitation which means that you can't use a certain class of motherboard which means you can't use a certain class of CPU even if it is the latest and greatest. Best to find that out before you spend your money.

And then I'll ramble a bit:

I've done a fair bit of work with 3D rendering for film and animation, so some of what I know might be helpful. First: Finding out if your software really supports parallel processing will make a big difference. I supported users of Autodesk Maya, for example, and versions up to and including 2015 loaded scenes using a single core. Artists would have to sit for 20 minutes with a single core of the CPU pegged at 100% waiting for large scenes to load. (Something to do with calculation of the DAG, IIRC.) Given the software, the most that we could've sped them up, with the fastest processors available, would've been 10-20% - hardly anything - because all the work was happening on a single core. We could've spent $3000 on a CPU, and it would've made very little difference. Version 2016, I hear, fixed that problem, making the extra money for more cores worthwhile.

The rendering software, on the other hand, was mostly built to use multiple cores. However, this helped more for some tasks than others; for some image-processing tasks, there's a lot of data and relatively little processing, so the computer spends most of its time waiting for data from the server and not much time processing. The one place where CPU made the biggest difference was VFX simulations, where the VFX software was often written to take advantage of the latest SIMD instructions. That's one reason to go for newer CPUs instead of older ones that won't show up in the basic GHz, core count and cache size specs.

What I've learned over 15 or so years of hearing artists complain about their computers is that figuring out where to spend your money isn't simple. However, the biggest speedups usually come from spending more money on data transfer speeds, and the major factor in data transfer speeds is disk IOPS. As a very approximate rule of thumb, spend as much on making your data go faster as you spend on everything else put together, and you'll be happy.

If money really is no object, you could build a machine I thought about but never had the money for:

- ...and your software supports GPU rendering: Look into Nvidia's Tesla cards, which are massive GPUs with no video card output.

- Get a FusionIO card for your data.

- Get the newest, fastest Xeon you can find, say an E7-8893v4.

- Put 512MB-1TB of RAM in it.

This will all put you back $20,000+, but it will be fast. I will live vicariously through your budget.
posted by clawsoon at 12:07 PM on January 27, 2017 [4 favorites]

Best answer: " 3D rendered images derived from real-word medical imaging--for example, creating a heart from a CT scan, converting high quality MRI images into 3d Mesh Models of soft tissues."

This is good to know. The CPU you need depends entirely on how well the application that does this work scales with the number of cores. What that means is, "Does this piece of software get faster if I go from one CPU core to two, from two to four, or from four to 16?" How much faster does it get? Does performance double with the number of cores, or does adding cores at some point create diminishing returns?

3D rendering can be infinitely scalable, where more cores = more speed up to the limit of cores on a current computer. But that doesn't mean this is true for your application.

The other question is whether the rendering in your software might benefit from either a gaming or workstation graphics card. Sometimes the gains from a specific type of graphics card, with the right application, are many times greater than the gains from any CPU.

tl;dr
Ask the people who wrote the 3D modeling software. What CPU and graphics card should I buy? Cross-check their answer with the community of users if at all possible.
posted by cnc at 12:41 PM on January 27, 2017

Response by poster: When you say "My current computer cannot run the imaging/modelling software," do you mean "my computer is unacceptably slow running the software" or do you mean "the software doesn't run at all"?

Thank you for your question. The answer is: it's unacceptably slow.

As some have suggested, no single component, in isolation, is the cause. Instead, the slowness stems from the interplay of a few primary components/metrics--CPU, GPU, RAM--and that many more secondary components/metrics.

(which is not to say I have the slightest idea how this interplay actually works, esp. with regard to making my system too wimpy to run Schwarzenegger-class software)

FWIW, my current computer key spec's are

CPU:Intel i7-6700 CPU @ 3.41GHz
GPU: Nvidia GeForce GTX 960
Memory: 16 GB RAM
posted by BadgerDoctor at 12:52 PM on January 27, 2017 [1 favorite]

That shouldn't be... horrifically slow, but the 960 is basically a rebadged 760, so even a mid-level modern GPU will be 2-4 as fast.

My completely-ass-pull theory is that your software doesn't support consumer (read: gaming) GPUs at all and requires a pro level Quadro/FirePro card.
posted by Oktober at 12:56 PM on January 27, 2017 [2 favorites]

My thoughts on that setup go direct to my wonder if you are using traditional hard-drives. Whenever reading or writing any sort of data, including pulling the relevant data into your RAM, that would certainly bottleneck. Now, the load-out specs you're talking about go way beyond what I'm familiar with, so I'll bow out, but you might look here at this article to get a feel for how you could get easy improvement in this regard, specifically at the CrystalDiskMark figures (comparing HDD vs. SSD vs NVMe m.2).
posted by a good beginning at 1:04 PM on January 27, 2017 [2 favorites]

Best answer: (which is not to say I have the slightest idea how this interplay actually works, esp. with regard to making my system too wimpy to run Schwarzenegger-class software)

Time to pull out your stethoscope. :-)

Best thing to do right now: Open up Task Manager if you're in Windows, "top" if you're in Linux, Activity Monitor if you're on a Mac, and then start doing some work on your computer that's unacceptably slow.

As you work, note what's happening:

- Is your hard disk indicator light blinking like crazy? If it is, that's a sign that either a) your disk is too slow and you should get an SSD, or b) you're running out of memory and RAM is being swapped back and forth to disk and you need more RAM.

- Check the memory usage of your application in Task Manager. Is it over ~75% of the total memory you have in the system? In that case, you need more RAM.

- Does the memory usage of your application go up to 2GB or 4GB and then stop growing? If so, you're probably running into 32-bit limitations, which might be the fault of your software and might be the fault of your CPU.

- Is your CPU, as reported in Task Manager, pinned at 100%? This means a better CPU - newer, more cores, more speed - would help.

- Does Task Manager report multiple cores? Let's say you have 4 cores: Is your CPU usage pinned at 25%? If that's the case, then your software is probably single-threaded and you need to look for the highest GHz in the newest CPU you can find; more cores won't help much.

- Is the CPU usage very low, i.e. under 10%? If so, your CPU is probably spending most of its time waiting for data from disk or network.
posted by clawsoon at 1:40 PM on January 27, 2017 [5 favorites]