hyperthreading ? is it reallu useful ??????

[Liampope]

As interesting as this thread is - the average user doesn't need a deep understanding of how HT works and/or rich analogies/metaphores. We have benches! Lots of them! And they're far more useful than hypothetical rambling. The benches show, for Core i chips, an absolute max gain from HT of about 30% in literally a couple of highly parallel multi-thread optimised tasks. Usually the gain is much less, even in multi-threaded encoding tasks, and in games the gain dwindles to zero, or maybe a few %. This is not going to get any different in the future - clever tweaks are not suddenly going to make HT a performance doubling miracle.

And all that from a feature that is adding 50-60% onto the price of the CPU. That's why HT is barely worth the transistors it's made of, imho. Actually I tell a lie - it's worth the transistors it's made of, which is a very small amount so HT is quite efficient in terms of performance/transistors - it's just not worth the massive premium Intel charge for it. I'd hesitate to recommend an i7 - for a gamer it goes without saying - but even for someone who does a bit of encoding, unless they're doing it 24/7 and time is money.

 

[CAT-THE-FIFTH]

The quad core Core i5 and Core i7 have the same dies. The problem is that Intel wants to charge you much more for a die which has all of its functionality enabled.

 

[Sighn]

@ CmdrTobs
"serving numerous people": I just mean that I do lots of video work for lots of other people.

I agree with your points regarding Hyperthreading misconceptions and Intel's Marketing. I've never been affected by these because I already knew what to expect from a HT CPU, concerning its positive attributes.

My main concern is the pricing for those who may actually benefit from the technology, as I said earlier.

If I could have obtained an actual 8-core CPU back in early 2009, I'd like to have done so, though I'd not like to think how much Intel would have charged for it.

BTW, I no longer switch BIOS profiles. I stopped doing this when I realised I didn't gain much by turning HT off.

 

[ThomasTheTankEngine]

As interesting as this thread is - the average user doesn't need a deep understanding of how HT works and/or rich analogies/metaphores. We have benches! Lots of them! And they're far more useful than hypothetical rambling. The benches show, for Core i chips, an absolute max gain from HT of about 30% in literally a couple of highly parallel multi-thread optimised tasks. Usually the gain is much less, even in multi-threaded encoding tasks, and in games the gain dwindles to zero, or maybe a few %. This is not going to get any different in the future - clever tweaks are not suddenly going to make HT a performance doubling miracle.

And all that from a feature that is adding 50-60% onto the price of the CPU. That's why HT is barely worth the transistors it's made of, imho. Actually I tell a lie - it's worth the transistors it's made of, which is a very small amount so HT is quite efficient in terms of performance/transistors - it's just not worth the massive premium Intel charge for it. I'd hesitate to recommend an i7 - for a gamer it goes without saying - but even for someone who does a bit of encoding, unless they're doing it 24/7 and time is money.

You've got to remember, it's not a feature exclusive to high end chips. The extra performance you'll get on i7's doesn't really warrant the premium they charge, but it's a nice feature for their lower-end i3 chips.

Intel wouldn't have stuck with hyperthreading for as long as they have if it was completely useless. At the moment it serves two functions: A bit of extra performance for those who are willing to pay the the silly premium... and it helps squeeze a bit of extra performance out of their low-end dual core chips.

 

[lee ashmore]

well it seems that starting this thread has lead to some interesting viewpoints , i built (and am currently waiting for my x58 sabertooth rma'd mobo) a new rig with a core i7 950 , 24gb of ram and dual ssds in raid and gtx 460 oc, with 3d modelling in mind , i couldnt see much of a speedup in rendering times with HT on so i turned it off , but it seems that (when i get my board back) i should leave it on ...........

 

[Liampope]

Lol at anyone buying an i7 and turning off HT! Coulda bought an i5.

Turn it back on - it's not giving much but it almost certainly will be reducing your rendering times a little. Maybe do some proper benching to verify. No advantage to having it off. You've paid for it - use it!

 

[ATIorNvidia]

The way to realise the 30% gains some people are touting here is if the application really was an 8 thread application that was putting your 4 cores into 100%. Then those 4 extra thread could execute faster somewhere else (provided the original 4 threads could still put the 4 real cores into full usage)

Despite all of your rambling, you've just told us right there that HT is infact useful.

The gains you get from HT are just fine and dandy so I don't know where you are getting all of this nonsense from. Even with adding 2 real cores onto a quad core (making it a 6-core) you'll never get a full 25% benefit from each core or 50% benefit from both additional cores... it's going to be around the 30-40% ballpark. Considering HT is capable of yielding a 20-30% increase.... I'd again say that's just fine and dandy!

If you compare an i3 to an E8400 you'll see the difference that HT makes. You're right in saying HT is inefficient, but in laymans terms, it's essentially a form of Out-of-order execution which is a good thing however you see it. The only problem is the premium Intel charge for it. Our day to day apps for the most part are not programmed to perfection, and that isn't going to change when apps/progams start using 4-8 cores/threads more regularly.... therefore HT will remain useful.

Amd Bulldozer should be something more in line with what some people expect from SMT (Hyperthreading). Though personally I would rather them give me more Cores and less heat and GTFO.

Bulldozer will have no HT. It's 8 cores and 8 threads period.

 

[joeyjojo]

The 'K' series does not have Virtualisation support so it will run virtual boxes slowly and virtual boxes are the only user time use that can scale well with cores.

I looked for some performance differences between cpus with virtualisation and without but didn't have any luck. I'm running a linux box with vmware on an e7200 atm and would like to know how it compares to say a e8xx.

 

[DragonQ]

The gains you get from HT are just fine and dandy so I don't know where you are getting all of this nonsense from. Even with adding 2 real cores onto a quad core (making it a 6-core) you'll never get a full 25% benefit from each core or 50% benefit from both additional cores... it's going to be around the 30-40% ballpark. Considering HT is capable of yielding a 20-30% increase.... I'd again say that's just fine and dandy!

Depends on the application. Video encoding scales brilliantly with extra cores and HT because you can simply split the image into 4/6/8/whatever equal sections so you'd get very nearly a 100% benefit going from 2 to 4 cores, for example. Other processes where you just need to do something many times, but new results aren't dependent on previous results, will benefit greatly. I have a computing project from a couple of years ago that calculates certain physical quantities according to a specific model with various parameters - it is set up to run as many threads as you want in parallel. In that case, you'd again see a near-100% improvement going from 2 cores to 4 cores.

 

[Cold_And_Miserable]

If you compare an i3 to an E8400 you'll see the difference that HT makes. You're right in saying HT is inefficient, but in laymans terms, it's essentially a form of Out-of-order execution which is a good thing however you see it. The only problem is the premium Intel charge for it. Our day to day apps for the most part are not programmed to perfection.

Bulldozer will have no HT. It's 8 cores and 8 threads period.

The differences between an i3 and a E8400 have little to do with HT. The i3 is 60% faster than an overclocked E8400 at X3: Terran Conflict which is single threaded.

Bulldozer was rumoured to have some sort of reverse HT that allowed it to process a single threaded application over 2 cores. It could be very handy.

 

[lee ashmore]

Lol at anyone buying an i7 and turning off HT! Coulda bought an i5.

And how would i have got the 24gb in an i5 ??????

 

[Corasik]

Still dont get it, if you have a program that uses your quad core at 4 * 100% how is splitting your 4 core into 8 * 50% cores any better.

Each core is split into several different execution units, capable of logical, integer, floating point, and SSE instructions. There are several of each type of execution unit in each cpu core.

Often a thread may be waiting for say an SSE instruction to be completed, and in the meanwhile the logic, and integer units may be sitting doing nothing.

Hyperthreading allows another thread to make use of those unused parts, so in theory at least 1 core can execute 2 threads so that they overlap, and it can complete both tasks faster than if hyperthreading wasnt sharing the cpu resources out.

The reason its not 100% successfull, is if both assigned threads end up waiting for the same execution units, then one thread has to be given priority, and the second thread will stall.

 

[Corasik]

The differences between an i3 and a E8400 have little to do with HT. The i3 is 60% faster than an overclocked E8400 at X3: Terran Conflict which is single threaded.

Bulldozer was rumoured to have some sort of reverse HT that allowed it to process a single threaded application over 2 cores. It could be very handy.

That rumor has been flying around for the last 2-3 AMD cores!. Even Intel have some patents out on the concept of multicore single thread systems. Its just really hard to achieve in the real world.

The whole reason that Hyperthreading is a good thing is that the individual cores are already have more execution units than most single threaded apps can handle!.

I suppose the ultimate CPU would be a single core CPU with a lot more execution units than the current processors (so very high IPC), and then a super extreme version of hyperthreading, so that you could push dozens of threads through a single super core!. Trouble is it would probably use a lot more power, and generate a lot more heat than lots of smaller cores.... So we're stuck waiting for programmers to properly make use of multicores

 

[ATIorNvidia]

The differences between an i3 and a E8400 have little to do with HT. The i3 is 60% faster than an overclocked E8400 at X3: Terran Conflict which is single threaded.

Bulldozer was rumoured to have some sort of reverse HT that allowed it to process a single threaded application over 2 cores. It could be very handy.

You're telling me HT does not give an i3 an edge over an E8400 (assuming they're at the same clocks) in heavily threaded apps?

I did not say HT is the only factor in the i3's superiority over the E8400, I merely stated it was a contributing factor.

And no, Bulldozer is not going to have any form of HT or "reverse" HT as you like to call it. Not even close.

 

[CmdrTobs]

The reason its not 100% successfull, is if both assigned threads end up waiting for the same execution units, then one thread has to be given priority, and the second thread will stall.

Yes. Then that compounds with the reality that most apps use synced threads, so if you can hyperthread then that's because your priority processes of any given app has stalled....

That rumor has been flying around for the last 2-3 AMD cores!. Even Intel have some patents out on the concept of multicore single thread systems. Its just really hard to achieve in the real world.

The whole reason that Hyperthreading is a good thing is that the individual cores are already have more execution units than most single threaded apps can handle!.

I suppose the ultimate CPU would be a single core CPU with a lot more execution units than the current processors (so very high IPC), and then a super extreme version of hyperthreading, so that you could push dozens of threads through a single super core!. Trouble is it would probably use a lot more power, and generate a lot more heat than lots of smaller cores.... So we're stuck waiting for programmers to properly make use of multicores

There are ways to reduce wait states by using ridiculously fast memory etc.... I don't think you should wait on programmers there hands are largely tied. There isn't really a way to get the cool physics thread to process that cool explosion before it receives the mouse button click = 1 command from a user interface thread.

You're telling me HT does not give an i3 an edge over an E8400 (assuming they're at the same clocks) in heavily threaded apps?

I did not say HT is the only factor in the i3's superiority over the E8400, I merely stated it was a contributing factor.

And no, Bulldozer is not going to have any form of HT or "reverse" HT as you like to call it. Not even close.

Given that HT demonstrably slows down many applications and at best can only risk the slowing down of all single threaded applications on a reasonably loaded system.... Means HT on the I3 is a contributing factor to the 8400 keeping pace (the very opposite).

Benchmark suits and reviews seem favour the chip industry in that they contain 20% parallel synthetic gains porn. If you are like the majority of USERs who Browse, Ms Office and game, program hyperthreading is a null feature. And if you are willing to Overclock (You could say everyone is in the context of this board) hyperthreading is NEGATIVE as you save much heat and gain many hz for turning it off.

Though, people on hear talk like this is 'renderfarmers.co.uk' to dishonestly portray hypertheading in it's one semi-sanguine role. Even in that it's dubious as these days for A/V at least 5mins of render time is Hrs of work. So the productivity gain is nowhere near the 20% gains in final render speed.

Bulldozer is not quite that, it's just got a extra ALU so hyperthreading will be faster for common cpu duties.

 

[ATIorNvidia]

Given that HT demonstrably slows down many applications and at best can only risk the slowing down of all single threaded applications on a reasonably loaded system.... Means HT on the I3 is a contributing factor to the 8400 keeping pace (the very opposite).

Benchmark suits and reviews seem favour the chip industry in that they contain 20% parallel synthetic gains porn. If you are like the majority of USERs who Browse, Ms Office and game, program hyperthreading is a null feature. And if you are willing to Overclock (You could say everyone is in the context of this board) hyperthreading is NEGATIVE as you save much heat and gain many hz for turning it off.

Though, people on hear talk like this is 'renderfarmers.co.uk' to dishonestly portray hypertheading in it's one semi-sanguine role. Even in that it's dubious as these days for A/V at least 5mins of render time is Hrs of work. So the productivity gain is nowhere near the 20% gains in final render speed.

Bulldozer is not quite that, it's just got a extra ALU so hyperthreading will be faster for common cpu duties.

At "reasonable" loads any performance losses you could ever get would be extremely minimal (inside 1-3% most likely), at higher loads HT will show obvious gains.

Not everybody builds a PC to play Call of Duty. There's other things besides gaming, y'know? Show me any programs that use more than 4 cores and don't benefit from HT, until then you don't really have a case.

Given that HT demonstrably slows down many applications and at best can only risk the slowing down of all single threaded applications on a reasonably loaded system.... Means HT on the I3 is a contributing factor to the 8400 keeping pace (the very opposite).

Rubbish. Prove it.

 

[Liampope]

Given that HT demonstrably slows down many applications and at best can only risk the slowing down of all single threaded applications on a reasonably loaded system....

Whilst I agree with your general principle (HT = meh), you seem to be theorising an awful lot without reference to benches/data. Where, for example, are benches showing HT 'demonstrably slowing down many applications'? I don't think I've seen any.

 

[CmdrTobs]

Despite all of your rambling, you've just told us right there that HT is infact useful.

The gains you get from HT are just fine and dandy so I don't know where you are getting all of this nonsense from. Even with adding 2 real cores onto a quad core (making it a 6-core) you'll never get a full 25% benefit from each core or 50% benefit from both additional cores... it's going to be around the 30-40% ballpark. Considering HT is capable of yielding a 20-30% increase.... I'd again say that's just fine and dandy!

Why's that? Why are those new cores different from old? In a parallel scenario I can get ~99.9% benefit from each new real core. There is NO logic to say I should get some made up percentage like "25%".

HT is only capable of 20% to 30% in those very applications where another real core would give 100%. It's seems you are slightly mixed up.

If you compare an i3 to an E8400 you'll see the difference that HT makes. You're right in saying HT is inefficient, but in laymans terms, it's essentially a form of Out-of-order execution which is a good thing however you see it. The only problem is the premium Intel charge for it. Our day to day apps for the most part are not programmed to perfection, and that isn't going to change when apps/progams start using 4-8 cores/threads more regularly.... therefore HT will remain useful.

Not really. 'out of order execution' is thread and CPU level. Hyperthreading is OS level. It's up to the OS to prioritise and allocate threads and up to applications to request threads. That's a whole different ball game.

The fact HT needs additional threads to process to get any gains means at the software design level the program must be broken up. That's actually not too difficult. The problem is balancing the load as most applications will have one main thread that is the busy one (A GFX engine for example) and other ones that will be chicken feed that could all run on one core happily using stand time splitting(a network connection to the COD4 server for example). Most of this for most apps can all fit happily within 4 real cores, and any maxed cores are maxed by ~1 (synced)thread(s) by 1 application.

Hypertheading thus is not only superfluous for the most part but risks race conditions and all that nasty stuff that means windows is best not using it for the most part, demonstrated by consistently worse performance on games with it on. The games that demonstrate no performance issues are likely the ones with some logic to turn it off.

Bulldozer will have no HT. It's 8 cores and 8 threads period.

Really?

 

[CmdrTobs]

Whilst I agree with your general principle (HT = meh), you seem to be theorising an awful lot without reference to benches/data. Where, for example, are benches showing HT 'demonstrably slowing down many applications'? I don't think I've seen any.

Use google there's plenty out there.

Remember I am talking at the CPU and code levels if it is indeed invoked. If an OS or application is smart enough to 'turn off' hypertheading and thus not impair its execution that's not score +1 for hyperthreading. That's score +1 for turning it off it of and overclocking. Leaving in the hyperthreading victories and calling draws when software manages to ignore it misses the subtle point.

 

[ATIorNvidia]

Really?

Yes, really. 8 cores and 8 threads. I can even try and dig out a post made by John Fruehe (AMD employee) over at the OCN forums where he stated that. Each module has two threads and two cores.

At "reasonable" loads any performance losses you could ever get would be extremely minimal (inside 1-3% most likely), at higher loads HT will show obvious gains.

Not everybody builds a PC to play Call of Duty. There's other things besides gaming, y'know? Show me any programs that use more than 4 cores and don't benefit from HT, until then you don't really have a case.

Yawn.