HT = The Future of Multi-Core Processors?

A thought recently struck me: The perfromance of Pentium 4 processors greatly increased with the invention of Hyper-Threading. It allowed a single processor to run multi-threaded applications, as well as multiple applications simultaneosly, faster, and more efficiently. Well I was thinking... What if the processor cores in multi-core processors had HT? The performance increase would likely be substantial, maybe even redefine the multi-core processor again. And then again, maybe that's already happened, and I'm simply oblivious. Anyways, it was just a thought... I'd like to see what others think of it.

 

The future is multi-core processing, 2, 4, 8...

HT on a P4 was mostly marketing. It didn't really work, and at times, performance actually suffered because of it. You have a neat idea, but it's back to the drawing board.

 

Intel says they Don't offer it because of manufacturing cost.
& that also means more heat.

 

Well as technology advances and multi-core processors become yet even cheaper, cooling methods become more advanced and effective, and the multi-core CPU evolves some more, maybe then they'll be able to integrate it? Or perhaps they could reinvent HT to be suited for integration in a multi-core CPU?

 

Multicore is definitely the way to go. But you already knew that.

HyperThreading+Multicore is an intriguing idea. I suspect that the first generation of such chips will be superior performers compared to non-HT multicores. Although many of the benefits behind the original idea of HT can already be seen in multicore processors nowadays, a combination of the two approaches would yield an exponential increase in performance (whereas a strictly multicore approach would yield linear increases per core at best given the overhead). The reason behind this has to do with abandoning the concept of temporal multithreading and adopting more of a simultaneous multithreading (SMT) approach.

Anyways, it'd be interesting to see how Intel goes about implementing this.

 

Actually, the Pentium D Extreme Editions already had this. They were Pentium D's with HT enabled, and they performed roughly the same as the regular Pentium D's without HT, but cost much, much more (not to mention were still smashed to pieces by the Athlon 64 X2). HT never had any noticeable effect on performance, quite frankly, and putting it on a dual-core CPU is no exception. Plus the Core microarchitecture doesn't support the technology, I don't believe.

 

A single core processor with HT could never hope to compete with a real dual core machine.

As for HT itself, the concept is good. The execution and support, apparently, is bad.

But the fact that Intel hasn't dropped the idea entirely, and are insisting upon implementing it in their future chips, makes me think that they have some improvements in mind to make the final product market feasible.

 

First, I should point out that hyperthreading is only Intel's implementation of a much more general technology called Symmetrical Multiprocessing (SMT)

Why do I say this? Because HT sucks, but SMT is actually a very good idea.

However, it's obviously not free. True, you only need one core instead of two, but that core still becomes more complex. There is a cost associated with it, and in some cases, it might just be more efficient to add another standalone core instead.

But SMT is definitely an interesting technology. I read that IBM made their own SMT-supporting CPU's which provided 80% extra performance, at a cost of 50% extra transistors.

Compare this to the multicore approach, which gives you 100% extra performance, at a cost of 100% more transistors.
SMT actually offered a better performance relative to the cost in this case.
Of course this was just one example (I think it was a server CPU, if it was launched commercially at all), and in other scenarios, the added performance might not have reached these 80%. But the potential is there. If implemented right, it could allow us to run more threads at a higher level of performance, than the pure multicore approach.

You probably have to approach it from the opposite direction of HT though.
HT was basically "We have one core, meant for one thread, but let's keep a backup thread ready to soak up unused CPU resources"
If they'd instead gone "we'll make one core, but make it wide enough to normally run two threads. And then, if only one thread is running, it can take as much as possible from the other half of the core"

 

So why didn't Intel just adapt SMT in the first place, instead of creating an inferior attempt at it?

 

Well, isn't that what they did?
SMT is just a broad name for the idea in general.
It's nothing more specific. They can't just adapt it because it's not an industry standard or anything. It's just a concept, like "multi-core". It's just a description of a fancy way to make our CPU's run more code faster. It's up to the individual manufacturer to come up with a way to implement it on their CPU.
Intel tried to retrofit it onto a single-threaded CPU core as well as they could.
They didn't do a particularly bad job of it, given circumstances.

for SMT to really be worthwhile, you have to design your CPU for it from scratch. (Just like to make a good dualcore CPU, you have to design for it from the beginning. You can't just glue another core in and expect it to work well. Look at Pentium D)
Intel started out with an existing CPU (the P4), and they just tried to get as many of the benefits of SMT as possible, without having to completely scrap the existing CPU.

 

HyperTransport (HT)
Hyper-Threading Technology (HTT)

May seem minor but it is HTT always has been.

 

So if they created a brand new multi-core CPU, designed specifically to accomodate SMT, might that be successful?

And also, what has AMD accomplished in terms of SMT?

 

Maybe AMD will integratre SMT into multi-cores first?

 

Yep. Potentially, anyway.

Nothing, as far as I know.

 

HT is coming when Intel releases the Nehalem microarchitecture next year.

 

I thought Hyperthreading died away with Intel's Netburst architecture. Like it was only designed to compensate for the fact that the Intel Northwoods and Prescotts had ridiculously long pipelines which churned out low IPCs.

 

For that incarnation of HT, yeah.

Ironically, HT (or SMT) becomes much more useful with wider, shorter pipelines, like you get on more modern CPU's. (Simply because there's better room for two threads to run in parallel)

 

Been done, IBM multi core SMT here is a Link good easy read.

 

Great read, thanks.

 

Symmetric multiprocessing is usually called "SMP". And multiple core chips are actually a version of SMP as well. So are multiple CPU machines, assuming that they share memory in the symmetric way.

HT is also a version of SMP, but it's actually more in the spirit of an old idea called "multiprogramming" which was certainly available on DEC-10. There was a similar facility on IBM System/360 which was called "MFT" (multiprogramming with a fixed number of tasks).

Revival of HT is not likely to be the future because it's not thermally as nice as multiple cores. HT requires the multiple threads to run on the same silicon, where having more cores means you can throw threads onto sillicon which isn't always hot already.

 

You guy's are mixing terms again, and I blame it all on incorrectly calling HTT, HT. Going down from there.
Jalf meant Simultaneous Multithreading (SMT). As Symmetric multiprocessing, or SMP, is a multiprocessor computer architecture where two or more identical processors are connected to a single shared main memory. Most common multiprocessor systems today use an SMP architecture. SMT permits multiple independent threads of execution to better utilize the resources provided by modern CPU's. And finally the comparable IBM technology to Intel's HTT is IBM's Hardware Multithreading (HMT). The reason is both wait for CPU to be idle before executing the other thread. SMT can execute two threads Simultaneous with out waiting for idle to switch between the two.