What CPU? Single vs Multi core in real life (no games)? Future?

[Zucca]

Example, just a starting point:

AMD Ryzen TR 2950X vs Intel Core i9-9900K
Userbenchmark Effective CPU Speed: 117% vs 121% 
AMD Ryzen TR 2950X  - 8th / 1143 
Intel Core i9-9900K  - 3rd / 1143

A lot of Software (Altium, Solidworks, Adobe, you name it) can not use all the AMD 2950x "horses", really? So the I9 9900K still wins in real life because it has less horses but "faster"?

I am planing to build a new PC soon and the more I read and research the more I am shocked how apparently it's difficoult to write software code which runs on multi-core  at the same time.

Do you expect a change in the software industry soon? If yes I maybe bank out a 2950x.

Hope you guys can teach me something because I am kind of confused about the situation right now...
HARDWARE INDUSTRY: More core as possible.
SOFTWARE INDUSTRY: Let's continue to use just few of them (i.e. 4), we don't want to rewrite the multi-core code from 0, too much effort, (and impossible?)



I am sorry in advance if my questions are stupid ones, maybe the right answer is "it depends" (as always).

PS: I hate PC games, if I want to play I take my mountainbike or my snowboard out. Workstation (no games) research leads to AMD TR, but if I look at the results in real life, it seems that the single core speed is still what makes the PC faster with the current software.   

[apis]

Since it's hard to improve core speed further, the only viable way forward is more cores. Not all problems can be parallelised though, so the set of problems that can be more efficiently solved in the future is smaller that previously. Old software is written for single core since that used to be the norm. In the future, more and more will take better advantage of multicore but, yes, it is (much) more work to do so at the moment, so it is avoided unless necessary. However, since your computer is often running a few processes simultaneously, these can run on different cores and thus in parallell, making a few more cores useful. Algorithms that benefit a lot from parallelising will be (re)written to take advantage of multicores sooner than other software of course, so if you need that kind of software it will be beneficial, video-editing I would assume is one example.

That said, unless you have need for computing power for some specific task you are better off spending your money on other parts of the computer like memory, good monitor that doesn't strain the eyes, nice comfortable keyboard and mouse and I never seems to have enough hard drive space (but hard drive space is mostly needed for media applications). You don't need a fast computer to write documents, use spreadsheet applications, surf the web, read emails or watch youtube.

You might also want to consider energy efficiency which also affects how much cooling and thus fan-noise is needed.

[T3sl4co1l]

Race conditions, deadlocks, spooky glitches that are impossible to reproduce and debug -- such is the domain of multi-threaded computing.

Think of how buggy single-threaded code is.  Now imagine those same developers writing multi-threaded code.

That is why multithreaded code is uncommon except where absolutely required (and, hopefully, written by an expert that knows their way around these pitfalls).

It's also relatively easy to write a solution, then debug and optimize it a bit before release, or as an update or minor version.  The structure is mostly unchanged, while the implementation changes slightly.  It's a lot harder to make a dramatic, structural and implementation change to a multi-threaded design.  In effect, the initial dev time is wasted.  Time would be better spent doing it right the first time, but then you have the problem of debugging a more complex system, that may not even be possible (see above).

As problems grow ever larger and more complex, devs will have no choice but to do this, but they will hold out for as long as possible because it makes business sense.

Tim

[Ice-Tea]

Unless you're running something that can use the additional cores *now*, don't bother with anything more than 8 cores. Just not worth it.

[Kleinstein]

For many real world desktop applications HD access times can be a big factor - so a good SSD can often be more efficient than a fast CPU.
With applications that really need CPU speed, much of the time critical parts are relative small fractions of the code and often even classical problems (e.g. FFT, linear equation systems, image processing). So the development of code that used more cores is often not that bad.
I tricky part is that the number of cores and also there interaction is different between CPUs. So ideally it would be different versions for different CPUs. 

Another point is possibly sending some tasks out to the GPU.

For really time critical tasks (e.g. FEM, video encoding or maybe even auto-routing) we may find more programs to send out the really time critical part the the GPU. Quite often a reasonable GPU is faster than a high end main (x86) CPU  for quite a few tasks.

With modern time windows things are getting crazy on the desktop anyway. From my feeling most of the processing power it used up for the virus protection and updates -  too slow a PC might not do anything else. At least these cause most of the annoying delays.  So maybe they should rethink the whole PC concept so one could get away with less of these OS related tasks.  If it goes on this way, the time to start a program is back to what we used to see with CP-M from floppy.

[Zucca]

For many real world desktop applications HD access times can be a big factor - so a good SSD can often be more efficient than a fast CPU.

That's why I need a new PC, I need a nvme m.2 disk which is stupid fast compared to SATA.
And of course I start thinking my system is 7 years old and I just need a new horse to ride.

Thanks for all the feedback I am learning a lot.

[Zucca]

Since it's hard to improve core speed further

I heard the same thing 10 years ago, but back on those days the issue was a microelectronic process engineering topic.
Today, I heard, we are going so small the atomic physics is starting to become the new limit.

[rsjsouza]

That always depends, but I would say there are a lot more applications where multi thread is used - at least cokpared to my old Windows NT4 days where absokutely nothing was like that.

I have recently updated my system from a quad core i7-920 (2.66Ghz) to a hexa core Xeon W-3690 (3.46GHz), which scales very well on video compressing and 3D model rendering.

By the way, the system above runs on a 2008 Gigabyte EX58 Extreme and the only thing I changed was the processor. Everything else is the same.

[Berni]

Yep as people said if you don't have a task right now that is capable of using many cores then don't bother.

Most things that electronics engineers do like PCB tools, compiling code etc are not multithreaded. What you want is a CPU with about 4 fast cores and a fast SSD. If you can get more cores at the same speed go for it, if more cores forces you into getting slower cores then avoid it.

By the way this is also the sort of system that runs games well. They don't like using many cores in general. If you are doing video editing, graphics, 3d rendering etc then you can make good use of the high core count Ryzen CPU, otherwise not so much.

Im personally running a i7 4790K (4 core, 4GHz base, 4.4GHz boost) for a few years now and the main reason for it is that when it launched it was the worlds fastest x86 CPU when measured in single threaded performance. This incidentally also made it the fastest CPU in the world for gaming so it became popular and its price skyrocketed soon after because of it(Now that it has settled down its reasonably priced again). The new 9th gen i7 CPUs take the crown now so gamers flock to those instead.

Here is a pretty convenient chart showing single threaded performance of CPUs:
https://www.cpubenchmark.net/singleThread.html

[Kilrah]

Actually quite a lot of tasks can make good use of many cores, maybe not 16+ but 8-12 for sure.

I have a mini-itx machine with an overclocked (5GHz) quad-core 7700K for when I travel and my home machine is a 4 year old 8-core, 4.2GHz overclocked 5960X. Obviously the home machine can do better but only on tasks that use multiple cores well, but that is actually the case and the difference happens to be significant with most of the things I run (code compiling, probably the best use of multi-core nowadays, video editing, photo retouching...). General responsiveness while it's busy on some heavy load is much better as well, with the 7700K it's typical to get little freezes when I'm running a large batch export in Lightroom, but on the 8-core that becomes enough for it not to hog everything to the last drop and let me switch to a browser window without blocking 2 seconds...

[Zucca]

So basically I need some Virtual Machine Voodoo to use the entire power of the AMD TR?

I can imagine the AMD eng bashing their heads on the benches, so much design effort and only few applications can use all its immense power.

Right now thinking about an I7 or I9 Intel.

[Kilrah]

I can imagine the AMD eng bashing their heads on the benches

Certainly not, they designed to be future-proof.
And they also have lower core count offerings, at prices way below intel....

If I was building a PC now I'd be for once seriously starting to consider them. Or I'd wait until the "leaked" 3rd gen that probably won't be long.

[Zucca]

This incidentally also made it the fastest CPU in the world for gaming so it became popular and its price skyrocketed soon after because of it(Now that it has settled down its reasonably priced again).

Ok got it, what bothers me is that Intel I9 Motherboard do not support ECC or many PCIe lines as AMD does. My guess? I9 is 95% sold to gamers who don't care about those Workstation needs, but they buy RAM with RGB leds on it  , wait another one  .

BTW what is now the definition of "Workstation" anyway? Right now  if we forget the graphics card circus it seems the most productive PC is the gamers one.

I feel lost 

[Zucca]

Certainly not, they designed to be future-proof.

When that future will become the present day?

[Kilrah]

Workstation typically refers to machines meant to hopefully have higher reliability e.g. ECC-equipped stuff / Xeon on intel side but probably the distinction is now most importantly the businessy style service contracts and guarantees coming with them.

When that future will become the present day?

Well as said a lot of stuff makes use of it right now already and it's only going to improve, so a year or 2?

it seems the most productive PC is the one of the gamers.

In terms of performance per dollar for sure. Businesses will however be happy to pay twice as much for a workstation-class machine that's got less risk of breaking down and will be repaired/replaced quicker if it does and thus might avoid leaving a paid employee unable to work.

[Red Squirrel]

Where high core count shines the most is server stuff.  VM servers, or even dedicated apps that spawn a lot of child processes to distribute load. Ex: copys of the same task but different instances of it. Web server serving multiple clients for example.

For desktop some apps will also take advantage of it but lot of stuff also won't.  Games are kinda hard to design to use multiple thread as they don't really involve tasks that can be distributed, it's more like 1 big task that's running in real time.  I'm sure some do but it's not as simple to design that way.

[Ice-Tea]

Im personally running a i7 4790K...

So am I. Unless the magic smoke gets in the air it won't be replaced anytime soon.

Now, for that stupid ass dual core i7 in that damn laptop...

[Zucca]

Thanks Kilrah. I am enjoing the discussion. You cleared the air about the "workstation" animal pretty good.

Anybody can point me to a I9 Motherboard which is more "Workstation" as possible? 

[Zucca]

Interesting the https://www.userbenchmark.com definition:

https://www.userbenchmark.com/Faq/What-are-the-UBM-performance-classifications/93

Desktop

Surfing, email, office apps, music/video playback.
This formula is weighted towards single-core CPU performance:

50%*DCPU Bench + 10%*GPU Bench + 30%*SSD Bench + 10%*HDD Bench
(DCPU Bench = 80% Single core, 10% Quad core, 10% Multi core

Gaming

3D Gaming and graphics.
This formula is weighted towards GPU performance:

25%*GCPU Bench + 50%*GPU Bench + 15%*SSD Bench + 10%*HDD Bench
(GCPU Bench† = 30% Single core, 60% Quad core, 10% Multi core)
†For games that can't use multiple cores the desktop CPU Bench is a better indicator of performance.


Workstation

Number crunching, virtual machines, databases, audio/video encoding.
This formula is weighted towards multi-core CPU performance:

40%*WCPU Bench + 20%*GPU Bench + 25%*SSD Bench + 15%*HDD Bench
(WCPU Bench = 10% Single core, 10% Quad core, 80% Multi core)

[Lukas]

Most things that electronics engineers do like PCB tools, compiling code etc are not multithreaded.

I'm glad that you mention that since I'm here to state pretty much the opposite. For large projects and enough RAM/fast storage you can launch as many compile jobs as you've got CPUs at your disposal and things will get faster.

In my PCB tool https://github.com/carrotIndustries/horizon/ at least, DRC is multithreaded. I would be surprised if more upmarket EDA tools aren't multithreaded for compute-intensive tasks as well.

[rhb]

If you're expecting good multicore performance from commercial software, it will be a long time coming.  I don't see those developers being good enough to master making it work without crashing.

Personally, I won't touch threads with a barge pole.  But shared memory programming is well developed in the Unix world.  SGI made some really good shared memory machines.

 MPI works quite nicely on multicore machines, though not as efficiently as shared memory unless they have modified MPI to use shared memory on mulicore machines.

They should have done that by now, but I've not messed around with high performance computing in a long time.

There are a lot of algorithms which cannot be parallelized.

The Hermes language had a cool feature that you could send a message to another process by transferring ownership of a portion of memory.  But it never gained any traction that I know of, so it's probably a dead language now.

I'm considering buying a 12-16 core  dual socket HP Z series machine as they are quite cheap now on eBay.  Apparently these were used for running SolidWorks and are coming off lease.  I've seen 20 core machines with 256 GB of DRAM for around $2k or less.  Sort of blows my mind having spent 4 years working on a VAX 11/780 with 4 MB of DRAM.

I don't think I'd spend that much, but I want to do a bunch of computational electromagnetics.  Finite difference is not hard to write and there are already several open source codes around such as OpenEMS.

[Red Squirrel]

I've written some multi threaded programs, nothing super complicated, but even a simple enough program has a lot of things to take into consideration so it gets tricky.

Not all programs in nature can be written to fully take advantage of multithreading, it really depends what kind of task the program does, and if it involves work loads that can be split up and worked on independently without relying on each other.   So for example if you are brute forcing something, you could split it up in multiple threads.  Thread 1 tries "password", thread 2 tries "password2" etc so you can try multiple passwords at once.   So now you can try say, 8 passwords at once instead of doing them serially.

One project I always think would be cool to take on is a MMO/minecraft type game but I'd want the world physics to be real time, like water flow, plants growing, NPCs, any AI stuff etc would constantly be working.  Say someone builds a dam and punches a hole then the water will flow even when they arn't logged in.  So imagine all these mechanics working, now I'd want to be able to make this multithreaded so that each physics is basically a "task" that is loaded off and then the result returns. (modified blocks etc).  This would definitly pose a challenge but if done right would make for a very scallable application.  Ideally, each thread does not even need to reside on the same physical machine, you can have it act as a cluster.  That's when things get really fun.    All this would be the easy part compared to the actual game client though, I know nothing about graphical application design. 

[Zucca]

thanks Lukas,Red Squirrel and rhb. Very interesting.

Now I9 9900K does not support ECC RAM, this is irritating. Intel is selling his ass to teenager gamers  , and they are right 'cause there are tons of money there.

I'm considering buying a 12-16 core  dual socket HP Z series machine as they are quite cheap now on eBay.  Apparently these were used for running SolidWorks and are coming off lease.  I've seen 20 core machines with 256 GB of DRAM for around $2k or less.  Sort of blows my mind having spent 4 years working on a VAX 11/780 with 4 MB of DRAM.

Tempting but either I do some Virtual Machine out of it or you will have a lot of unused horses laying around. They are wonderful machine, porno stuff really but I don"t see for example Abode or Altium running there as fast as on a I9 9900K.

[NiHaoMike]

Having a lot of cores also comes in handy for heavy multitasking. If I had a bunch of highly efficient cores and RAM with no better use of them, my first thought would be to spin up a bunch of earnhoney mining VMs and get back some of the hardware cost.

[BravoV]

..... I need a nvme m.2 disk which is stupid fast compared to SATA.

Once you're "nvmed" , you will never look back.