[Frontino]

Yeah, I said it! I still don't get why they moved to many-core chips, when most applications (including modern games) requires algorithm dependencies and the best scale you can get is on a dual core, but that doesn't mean you won't get issues with that (lag, asynchronicity, sound effects muting..).
Just put on the market a 5 GHz single core CPU and let me buy that! Don't stick the many-core gimmick on everybody's mouth!

End of rant.

[hoho]

Would you like an integrated power plant with that CPU?

Also, the main bottleneck is memory latency. It's far harder to improve that than to improve CPUs to tolerate the high latency. Going multicore helps quite a bit with that in addition with tons of other benefits it brings.

[Alexko]

Who has moved to many-core CPUs? Most mainstream CPUs still have only 2 to 4 cores, and the high-end ones Turbo up to ~4GHz when only 1 or 2 are loaded.

[Lightman]

Nothing is black or white.
We are living in 'trade-off' world at the moment and multi-core is one of them.
Besides your 5GHz single core CPU will struggle really hard with many modern games, just to name BF3 among others.

[liolio]

Originally Posted by Lightman

Nothing is black or white.
We are living in 'trade-off' world at the moment and multi-core is one of them.
Besides your 5GHz single core CPU will struggle really hard with many modern games, just to name BF3 among others.

Indeed when I read the news, all too often by the way, there is indeed a lot of brown.
Brown seems to be the new grey... and there are a lot of shades of brown...

[french toast]

No thanks...give me a quad @3ghz over 1 at 5ghz....multi tasking would be a nightmare

[hoho]

Originally Posted by french toast

multi tasking would be a nightmare

Yes, that's another point I forgot to mention earlier.

Moving from a single core to even a measily 2.8GHz Northwood P4 with HT was by far the biggest upgrade to system responsiveness I've seen. EVER. And that includes moving from 233MHz P1 with 16M RAM @ 40MB/s to 500MHz P3 with half a gig at some 5x+ higher throughput.

[Blazkowicz]

Originally Posted by hoho

Yes, that's another point I forgot to mention earlier.

Moving from a single core to even a measily 2.8GHz Northwood P4 with HT was by far the biggest upgrade to system responsiveness I've seen. EVER. And that includes moving from 233MHz P1 with 16M RAM @ 40MB/s to 500MHz P3 with half a gig at some 5x+ higher throughput.

I've had similar gains on single core, single thread computers, just by asking the OS scheduler to set the priority of an offending process (firefox) lower.

moving from sempron 64 to athlon II X2, on same mobo, made my PC incredibly better. sure. installing an OS in virtualbox was utter pain (major CPU hog for every simulated I/O)

[hkultala]

Originally Posted by Frontino

Yeah, I said it! I still don't get why they moved to many-core chips, when most applications (including modern games) requires algorithm dependencies and the best scale you can get is on a dual core, but that doesn't mean you won't get issues with that (lag, asynchronicity, sound effects muting..).
Just put on the market a 5 GHz single core CPU and let me buy that! Don't stick the many-core gimmick on everybody's mouth!

End of rant.

The reason is simple:

We have reached the limits of what kind of performance is possible from..
1) single-core with current mfg technologies, AND
2) updating mfg technologies only give very small/slow increase in this.

Given current mfg technology:
1) much higher performance single cores are not possible.
2) slightly higher performance single cores require very high supply voltages leading to huge power consuption and cooling requirements.

If you try to make "bigger, fatter core" it will end up having worse clock speed, and the fatness only gives very small increase in ipc.
If you try to make processor with much longer clock speed, it will end up doing less per clock cycle


So, I'm expecting the performance in non-parallelizable tasks will only about double during next 10 years.

Performance with single thread with data parallel code may increase much more, when wider SIMD-instructions are introduced.

[imaxx]

you're thinking in a too GPU-centric way...

huge parallel work is a natural fit for GPUs, not so much on CPUs. Amdahl's law clearly states that the benefits of multithreading applies only to the parallel parts.
On CPU, it is not unlikely to find a 50% sequential work with the other 50% parallelizable.
so real benefits of multi-threading ends fast after 4-8 cores... whereas even a modest 5% increase in the single core speed would still yield a notable speedup.

It's easy (in a sense!) to cast parallel rays over a grid, not so easy to perform complex analysis with strong interactions, that's why single-core performances are still important.

[pcchen]

Single thread performance is still improving. For example, Sandy bridge is, clock for clock, faster than Nehalem. It's just that you don't really get much more from all the transistors if you focus on single thread performance. You may be able to make a single core CPU as large as a 8 core Sandy bridge, but its single thread performance is probably at most 50% faster and a 8 core Sandy bridge will kill it in most benchmarks (and real world applications too).

[Albuquerque]

Originally Posted by Frontino

and the best scale you can get is on a dual core, but that doesn't mean you won't get issues with that (lag, asynchronicity, sound effects muting..).

Not true. A multitude of games see scaling beyond dual core. And the 'threading' of an application is not linked to all the negatives that you outlined; a single-threaded app is more likely to have those issues than a multithreaded app simply because you could never truly be doing more than one thing at once.

Originally Posted by imaxx

... whereas even a modest 5% increase in the single core speed would still yield a notable speedup. <snip> that's why single-core performances are still important.

Nobody is ignoring IPC nor single-threaded performance. Newer architectures continue to deliver improvements to instructions-per-clock; and newer architectures also continue to deliver higher clocks to lesser-threaded tasks (see also: "turbo" technologies.) But there's only so much that can physically be done on that front.

[imaxx]

Originally Posted by Albuquerque

Nobody is ignoring IPC nor single-threaded performance. [...] But there's only so much that can physically be done on that front.

Check out the 50% line at 8 core. Assuming more than 50% parallelism (unless you're doing raytracing or similar stuff) within a single application is not trivial.
Clearly, it's ok for specific segments (multiple background applications, web-services etc.) with higher parallelism.
It is also interesting to see the 95%/2048 cores curve.

[Grall]

Originally Posted by Albuquerque

And the 'threading' of an application is not linked to all the negatives that you outlined; a single-threaded app is more likely to have those issues than a multithreaded app simply because you could never truly be doing more than one thing at once.

Yeah. Those of us who are a bit older and have followed PC gaming for many years now remember like when you played Quake 2-engine games for example and hit situations or areas your PC couldn't quite handle and the framerate slowed down so much the sound garbled.

This is because many games handled sound essentially at the end of a really long main program loop rather than as its own separate thread, and if the loop took too long to execute then the sound buffers ran dry and playback started stuttering.

With multicore processors, sound will most likely be its own separate thread, and will have a much higher chance of receiving all the processor cycles it ever needs to - especially on quad core or hyperthreaded dual core machines and up.

[Davros]

Many years ago there were rumours that the next amd cpu's would have "reverse hyperthreading"
where multiple cores would be used to run single threaded apps.

[hoho]

Originally Posted by Davros

Many years ago there were rumours that the next amd cpu's would have "reverse hyperthreading"
where multiple cores would be used to run single threaded apps.

Yeah, I remember those and I still don't quite understand what kind of theory those rumors had behind themselves to even make it remotely useful and not require re-programming everything from scratch to be automatically parallelized and thus pretty much the same as using openmp on multicores today.

[CarstenS]

I second the notion - partially.
I'd really love to have high-IPC, high-clocked Few-Core-CPU. Personally, I still haven't seen the dire need to upgrade to a quadcore from my C2D E8500 with a little OC. I might have jumped on the i3-train, if they'd been better overclockable or maybe if Intel had offered a "k"-model.

Most of the programs I use on a regular basis, don't need high performance at all. Some that do, do not make use of more cores, but rather enjoy high IPCs or high clocks (image manipulation with GIMP for example) or single-player RPGs like Dragon Age an it's successor.

So, if they're clocked high enough or if they are overclockable, I really would consider going for an Ivy Bridge dualcore (with Hyperthreading possibly, just in case), instead of a quadcore where I'm paying for many more transistors I rarely use.

Originally Posted by Davros

Many years ago there were rumours that the next amd cpu's would have "reverse hyperthreading"
where multiple cores would be used to run single threaded apps.

And that's exactly what you get when someone drops you a hint about an upcoming technology and you're trying to spin a story from it. Bulldozer's modules are in a way the opposite of Hyperthreading, joining ressources of different cores to work on a joint taks (AVX) - if you want to view it that way. Same thing with Fermi having "not one tessellator" - you could spin it in a way, that Fermi does not have a tessellator (with the emphasis on "a" being lost) and come to the conclusion it does everything in software, therefore must be painfully slow.

[I.S.T.]

Originally Posted by CarstenS

Most of the programs I use on a regular basis, don't need high performance at all. Some that do, do not make use of more cores, but rather enjoy high IPCs or high clocks (image manipulation with GIMP for example) or single-player RPGs like Dragon Age an it's successor.

Dragon Age 1 actually likes 3 core minimum CPUs a lot more than dual cores...

[CarstenS]

Originally Posted by I.S.T.

Dragon Age 1 actually likes 3 core minimum CPUs a lot more than dual cores...

You're right. I might have remembered had I not played it at my monitor's native resolution with Supersample-AA, which kind of limits the Fps in my case to mid-40ish, which is pretty much exactly what my CPU can handle too.
http://www.pcgameshardware.de/aid,69...dventure/Test/

(Which brings up another point of multicore CPUs in games: Mostly their benefit is evidenced by low-res, medium-detail, no-AA style of benchmarks. In reality, people who care about a decent level of image quality would be needing multicore CPUs less than the average benchmark might imply.

[Arwin]

I've been looking into CPU/GPU bottlenecks because my wife has an old dual-core desktop, and I was wondering if I could drop in a big GPU there and then give her my quad-core low power machine that only holds low-profile GPU cards. But looking at the graphs, it seems that a dual core processor holds back a lot of higher end GPU cards in many metrics, limiting them to about half their maximum performance.

My work laptop has an 8-core i7, but of course the GPU in there barely matches my low-profile 5570 in my desktop, so that doesn't help much.

[Albuquerque]

Originally Posted by CarstenS

(Which brings up another point of multicore CPUs in games: Mostly their benefit is evidenced by low-res, medium-detail, no-AA style of benchmarks. In reality, people who care about a decent level of image quality would be needing multicore CPUs less than the average benchmark might imply.

I generally agree, but there are some odd outliers. Skyrim really loves six cores even at the highest details:

I suppose it might be caching, but really seems to love threads. Bizarre.

[swaaye]

I remember during the P4 and Athlon 64 days, when dual cores first showed up people were saying that we were at a dead end on x86 performance per clock per core. Maybe that was mainly the marketing push of the day (reviews like to regurgitate that stuff). Intel's vast army of brilliant minds continues to bring considerable improvements with even their tick/refresh CPUs. AMD seems to be heading backward though.

[suryad]

I think it would be silly to go back to a single core no matter how high the IPC might be. As it is we can alreayd see Intel is improving not only the electrical and thermal efficiency of the processors but it's IPC and number of cores as well which I think is the right way to go. I went from a quad core to a hexacore and the only apps I noticed the biggest improvement was in Java development and games as well. I have not overclocked my cpu and I have played the same games before on a quad core and on the hexa core so my conclusion is that IPC helped in that department. But when I am running a git pull, and a maven build, and firing up multiple instances of Intellij, and opening up sql developer to run some sql queries, etc etc, ther performance improvement has been drastic...which I attribute to the number of cores as well as the IPC improvements. Now all I really need to maxmixe performance is an SSD and that time is coming near as it seems one of my velociraptors is starting to fail.

[Albuquerque]

I don't see how your reply has anything to do with mine.

Total instructions computed per clock continues to increase, however there is a hard physical limit as to how 'fast' a single core will be able to perfom. Your chart has nothing to do with the physical limit I'm talking about.

You can point to "ZOMG SINGLE CORE PERFORMANCE MATTERS" and I've said nothing to counter that, because it's true. But that has no bearing on the physical limits of our world and the manufacturing capabilities we have at our disposal. Sorry, you will have to find another way to 'express' that work, or else simply accept that work is not going to get markedly faster in the forseeable future.

[imaxx]

Originally Posted by Albuquerque

Total instructions computed per clock continues to increase, however there is a hard physical limit as to how 'fast' a single core will be able to perfom. Your chart has nothing to do with the physical limit I'm talking about.

You can raise performance by increasing clock, IPC, multithreading. Increasing clocks hit the wall with netburst, where they reached 8Ghz with a longer pipeline (32 stages if I remember correctly) with poor performance results, so they did revert to PM project. Increasing IPC - especially in x86 architecture- has a natural limit due to registry dependencies. Increased multithreading has a natural limit on the amount of parallelizable work you can have within a single application. This is the physical limit i'm referring to -the fact you can add more cores, but their usefulness decrease more than linearly until they are almost an useless way to improve performance for a single application.

I was simply pointing that both orthogonal limits (say IPC/threads) are nearer, more or less. So that you might get higher overall speed by boosting single-thread performance than adding another parallel thread/core to the pool, unless you are executing an highly parallelized process.

Physical limits are also in function of the CPU architecture (see the 8Ghz Netburst), but so are the price they pay for mispredictions and so on.