Who Will Benefit From Hyper-Threading?

qoncept asks: "I've read a number of reviews of Intel's new Pentium 4 3.06ghz processor with multithreading and I've noticed that perhaps it is being reviewed as an option to the wrong people, and in fact Intel may even be marketing it to the wrong people. It seems that, as a business move, Hyper-Threading may not have been worth Intel's investing in it. Most reviews show that in single threaded benchmarks, there are literally no benefits to using HT. In multithreaded processes, the results are moderate at best. Yet, of course, the reviews seem to say the feel is better. There you go -- it won't increase your productivity by compiling your Java. But, price point permitting, it may be exactly what the casual home user wants -- save money by getting, say a 3.06ghz HT CPU instead of a 3.6ghz CPU without, yet have Internet Explorer, mIRC, AIM and Word run just as 'comfortably.' The benchmarks don't say much for HT, but I'm at least slightly excited about it. What about everyone else?"

Comments

Developers

[IshanCaspian]

I really look forward to being able to run multi-threaded apps on the average user's desktop. There are a lot of advantages to being able to have two lines of logic running concurrently. Although there are few performance benefits right now i'm sure developers will appreciated the ubiquity of SMP and all of the nifty programming techniques that come with it.

Re:Developers

[davincile0]

Neither SMP nor hyperthreading are prerequisites to writing multi-threaded programs. They run just fine on single-cpu machines, so long as the thread library and the OS' scheduler do a decent job. It's nice when you can run multiple threads concurrently on multiple CPUs, or with hyperthreading, but it's a stretch to say "now we can finally start writing multithreaded programs."

Re:Developers

[photon317]

The use of the term "thread" in the statements you quote refers more to a "thread of execution" than specifically to a thread in a multi-threaded app. Their statements were intended to (and do) apply to multiple threads in one multithreaded app, or several seperate multithreaded apps, or several seperate processes. The things they say about true parallelism are basically the same things you could say of SMP vs UP. They built a "half-smp" inside a single processor. Another way to think of it is that they are allowing threads of execution that would otherwise be completely blocked to make opportunistic use of parts of the instruction pipeline that the main active process isn't using.

In any case (up, smp, up-ht, smp-ht), one should code multi-threaded for apps that can benefit from parallelism, and don't bother for those that don't. And in any case, the os will do it's best in the given hardware context to satisfy your needs.

Re:Developers

[PaulBu]

Two points:
They built a "half-smp" inside a single processor.

It's more of smp squared (or at least doubled... ;) ) -- threads in multithreaded app share instructon AND data caches (no cache coherency problems which are going to plague SMP implementations more and more), as well as register files (fast data exchange which does not require access to the main memory nor even to the inter-processor bus).

one should code multi-threaded for apps that can benefit from parallelism

One? Or the compiler? You use s/w available in source code, right? ;-)

Re:Developers

[jsse]

The optimization is transparent to application developers, but it takes extra effort in compiler to compile the source optimized for a particular architecture, like HT.

I like to use this simplest example in lecture. Pentium(I) has two pipelines U-pipe and V-pipe. They're fed in instructions sequentially. Therefore, it'd be better if we could arrange the code during compilation time in such a way that an execution of an instruction does not depend on the result of the previous one.

I'm not sure what it takes to optimize compiled code to take advantage of HT, but I'm sure it can be done.

Re:Developers

[photon317]

I can see the cache thing making a performance boost versus SMP. I don't get the register-file sharing. How can two threads, even if they are in the same process, share register files? Each thread has it's own set of registers, and they can't see each others', as far as I'm aware. But SMP still wins in many cases because you can actually run two cpu-bound processes/threads full-on with 2 cpus - with hyperthreading instead of getting 2x you're getting 1.5x, or 1.1x, or 1.7x maybe depending on the nature of the two processes/threads. One is using the pipeline slack the other one leaves. If the tight loops in these threads/processes are hand-optimized asm that tries to make full use of the pipelines with careful instruction ordering, it would leave virtually zero free for another thread to use. For that matter, if the two threads are doing virtually the same thing, they're likely to get into lock-step with each other since they use the pipelines in a similar fashion.

Yes I use software available in source code form, what does that have to do with my statement that one should code parallelizable apps in multi-threaded style as appropriate regardless of hardware and it will always be of benefit.

Everyone misses

[psavo]

The point in 2+ CPU systems. It's not about getting multithreaded apps gettin faster, it's about getting more programs run together better.
When using 2 cpus (or HT), when one process takes all the juice there's still some left for everything else, and system will appear more responsive.
So, there You go, You can encode some divx, and still browse comfortably net, or listen to mp3's, or watch some divx. (Of course I don't know how effective HT is, but my 2xAthlon lets me do just that).

Re:Everyone misses

[RzUpAnmsCwrds]

Not a problem on my 1x AthlonXP 1667 (2000+). Most modern operating systems can balance processes pretty well (I can burn a CD while playing a game, for example; my burner rarely has to use BURN-Proof). The point is that a process should never take over the system because it should never be allowed to take over the system. I'd say that Windows does a pretty good job of that.

Re:Everyone misses

[ConceptJunkie]

But since NT/2000/XP has all disk I/O in critical regions, your system still grinds to a standstill.

The day Microsoft OS's are not ridiculously I/O-bound, this will make a much bigger difference... for Microsoft users.

Of course, I guess there's a point to helping your filesystem remain intact when Granny or the baby flips the Big Red Switch without shutting down...

The moral: Use lots and lots and lots of RAM with Microsoft

Re:Everyone misses

[Sloppy]

The benefit you describe is just as achievable with good scheduling even on a single processor. (Amiga users have known this since the 1980s.) You can run a bunch of CPU-bound processes and the system's responsiveness will not be influenced at all.

The real benefit of SMP/SMT on a modern OS is that it lets you get stuff done faster. It has nothing to do with responsiveness; that's the scheduler's job. SMP/SMT only help with responsiveness if your scheduler is defective.

Re:Everyone misses

[Parsec]

As with km790816's comment, this interests me. I find pre-emptively scheduled OS's to not schedule I/O according to process priority. Thus a lower priority process can tie up the hard drive and slow down a higher priority process, possibly causing it to fail in its performance (skip audio/video). This seems to be a problem on both Windows and *nix. Does anyone have an answer / solution to this?

Re:Everyone misses

[ConceptJunkie]

My understanding is that Windows does this to help maintain file system integrity... turning off a Windows box without shutting down is much less risky than with Linux, as I understand it. The trade-off is performance, and when I have lots of stuff going on, sometimes the whole system will freeze for several seconds or more while the disk thrashes. If I get really stupid and launch something memory-intensive when the machine is already low on RAM the system can go dead for several minutes. It's pretty easy to avoid that, usually.

Re:Duh

[weird mehgny]

They also run just fine on a PIII 450.

Yah, but they may not start or perform tasks instantaneously.

I have dual p3 1ghz already

[DrSkwid]

and a 500Mhz p3

and 99% of the time you wouldn't know the difference

Re:I have dual p3 1ghz already

[Myrkur]

I watch porn 99% of the time too! But a better processor makes on hell of a lot of difference to that 1% of the time when I'm not.

Re:I have dual p3 1ghz already

[ctr2sprt]

Are you including time that the machines are idle and you're not using them? That's the only way I can make sense of your claim. Even if you're not a hardcore developer (where MP is a big bonus) or gamer (where the faster CPU make all the difference, and it doesn't matter how many of them you have), the difference is still going to be visible for ordinary desktop tasks, like ripping a CD and surfing the web at the same time.

And while we're comparing experience, I have a 2-way PPro-200 system, a 2-way P3-450 system, and a 1-way P4-1.6 system. Both of the MP machines are far more responsive for, well, every task that I throw at them: the only reason I don't have only MP boxes is the cost.

Re:I have dual p3 1ghz already

[MBCook]

It seems to me that games should see some benefit, although that isn't panning out right now. Don't you think performance would jump in games if you could devote an ENTIRE processor to one, so it didn't have to worry about sharing with other processors, not having it's stuff in the cache, etc? The other processor would handle dealing with the hard drives, NIC, and other 'mundane' things that would just 'tie down' the other CPU.

Because games do many things at once (sound, graphics, input, AI) if they are well written to use SMP, they would see a large benifit. But even in the situation I gave above, there should be a benefit.

Re:I have dual p3 1ghz already

[DrZaius]

I've run a multitude of dual and single processor machines. I've never noticed a difference as far as application usability goes. Last time I ever had a problem ripping a CD and surfing the web at the same time was when I was still running my 200mmx.

Also, dual CPU will not help the "feel" of your system. If it slows down, it is because your IDE bus is busy and Windows is trying to access files off of your HD. You can increase your clock cycles and the number of processors all you want, but the PCI bus is still only 33MHz on a 32bit bus.

Dual processor makes very little difference to PC users who don't have penis size issues.

In a server, totally different story. When most OLTP databases or webservers are running, they are generally servicing more than 10 users at once.

who will benefit? intel of course

[Splork]

they've just invented a feature that their marketing department can say without lying that their chip has that others don't.

the fact that it doesn't do anything useful for most uses at the moment makes no difference.

Marketing. . .

[Cokelee]

it's the only way with Intel. They can't really make a faster processor, so they're always coming up with new ways to make it "feel" faster, or make the clock speed higher.

I'm not excited at all. What about resonance? Multithreading with simultaneous and common processes may cause it to run SLOWER!

Re:Marketing. . .

[qoncept]

I believe its called optimizing. I suppose there are alternate routes -- they could just slap more chips on and require an external powersource, a la Voodoo5. But I'm sure as we all know, without optimization, computers would have gone nowhere.

Sure Intel is using Hyper-Threading as a buzz word, but that doesn't mean its worthless. Your beloved AMD copied SSE, and made their own 3DNow! and you'd have a hard time convincing me either of those will have the impact hyperthreading does. I saw someone compared the price of a p4 with hyperthreading to that of a dual athlon, but thats not the point. Its the technology. Don't you think AMD and everyone else who makes cpus for anything would be interested in taking advantage of it? If it was AMD (and I really dont think anyone else, except maybe Motorola) who had introduced HT, slashdotters would love it.

Of course I don't like Intel (and of course I hate Bill Gates more, and Steve Jobs has everyone beat), but that doesn't mean having them around isn't healthy for the entire industry.

By the way, has anyone else noticed nvidia trying the brute force tactic like 3dfx did right before they went out?

Re:Duh

[psavo]

Funny thing is. I run Word 97 on my 2x1800XP, and you know what?
The fucking PoS(hit) still has those 2-3 sec pauses all the time. It didn't sped up from a PPro200 even a tiny bit.

Another review

[KarateBob]

Heres a new review of the 3.06 HT at Sharky Extreme

Re:who will benefit? intel of course

[dpilot]

Don't forget that this is on a processor with NetBurst Architecture, so it makes the Internet faster. Now with NetBurst AND HyperThread those web pages will be on your screen before you've clicked the button.

This is a great feature

[Joe Tennies]

As much as I am really rooting for AMD, I must say that I wish the Athlon's had this feature. The average user is not going to notice a big difference right now because most applications have been so optimized for single processor computers that they perform poorly on SMP computers. The big thing that hyperthreading is going to do is allow for more registers on the X86 architecture w/o changing the instruction set at all. This is the big enhancement and why I am so excited about it.

Re:This is a great feature

[MBCook]

True. Maybe the x86-64 processors (Hammers) will get it soon after launch (or at the next major die change). That said, I think that Intel could wipe the floor with AMD performance wise easily. They already have a great processor. If they could just get it's FPU performance near that of the Athlons (or faster), then AMD would only have the price arguement on their side. IMHO, of course.

I think you're wrong about the average user seeing no performance increase. I think that that's exactly who WILL see the increase. Developers might not. It's the guy who sits there surfing the web, playing MP3s, and ripping a CD who will benifit from this. A dual 600 feels zippy doing things that a much faster computer has problems with, because it's got a second processor to help process UI clicks, etc.

Re:This is a great feature

[Sloppy]

As much as I am really rooting for AMD, I must say that I wish the Athlon's had this feature.

Me too; I wish all CPUs had it. But don't forget: for about the same price of this P4, you can buy two Athlon MPs. That's two whole cores, and they'll cream a single SMT P4.

This is a neat enhancement, but it's hardly "big." It just gets Intel somewhat closer to catching up to AMD on bang/buck.

Not Me

[Konster]

I read several reviews, most notable among them are here and here. Although the technology seems compelling when looking forward a few years, its infancy just doesn't sell me the product, especially when I consider that a dual Athlon MP 2000 (1.6Ghz) is respectably close to the $700 PIV 3.06GHz with HT, and costs a LOT less.

3.06GHz PIV + motherboard + 512MB DDR RAM = $1025
2 Athlon MP 2000 + motherboard + 512MB DDR RAM = $695....for 80-90% of the performance of the HT PIV?

Sorry, but I can get the basics for an SMP system for $5 less than Intel wants for its new flagship CPU.

Now, if I could get 2 PIV 2.4 GHz CPUS with HT, that might be a different story...

Re:Not Me

[MBCook]

Hyperthreading is a good concept. I don't know about you, but I'd pay a slight increase in price to get it over an equivelent CPU. This is, when you're using multiple apps, free performance. That said, I agree with you. If the price is similar, I'd rather have a true SMP rig that was slower than a faster HT CPU.

What this boils down to is this: if given the choice between two CPUs of near equal speed and they cost nearly the same, what would you rather have? The one with or without HT?

Besides, let's face it, this is for marketing. It does give benefit, but this gives Intel the ablility to say "things are more responsive with Intel Pentium 4 than the other guys". Not only that, but now people can market P4 computers and say "Sure it costs more, but it's like getting a free second processor. That's worth like $700 right there." It wouldn't be the first less-than-half-truth we've heard from people in the PC industry before...

(*cough* netburst *cough*)

Re:Not Me

[SuiteSisterMary]

People today, it seems to me, have this odd notion that the first iteration of something should be the best iteration.

Sure, hyperthreading won't do much for you now. But what about the next generation? Or the third? Once this trickles down to the point that it winds up on chips automatically, great.

Re:who will benefit? intel of course

[foniksonik]

Hmmmm.. seems to me that a 28k modem will still only give you 28k. Does anyone with cable/dsl/T1 have speed issues with their web browsing? http pipelining seems more useful. If only 'the rest' of the web would cater to it.

How about...... everyone?

[Bleedy20]

There's a few people saying "Right now, it doesn't do anything."

"Hey man, have you heard about this new invention? It's called a Compact Disc!"
"Forget about it man, they're worthless. My record player gets crap sound out of them."

How about MMX? What about 3D cards? We had to wait until software came out that really took advantage of them before we could see what it could really do. Some apps are developed for multithreading, but the hardware's got to be in wide release before it's worth it to developers to write for it.

I think in the end we'll all benifit, but just like every other technology it'll take some time.

Rack Density

[stmfreak]

In our preliminary tests of a unit Intel donated, we were able to run four instances of a single threaded process on a dual-proc HT machine. The performance was somewhat greater than two instances on the same box.

Admittedly, not conclusive results and we've yet to run more controlled tests, but our initial take is that you might achieve higher rack density of processes and throughput using this architecture.

Sorry I don't have specific data, we're still studying HT.

Digital editing, of course!

[eMilkshake]

Check out this article that answers your question. It shows how the new Intel chip in a Dell workstation blows the pants off a dual-cpu Apple computer for less cost.

Hyperthreading as HZ

[Animats]

Think of hyperthreading as turning the HZ value up to 1,000,000,000.

It's a big win if your OS's scheduler sucks, but supports multiprocessors.