What is Digg?Sponsored by Gilt Man
It's a Gilt Man's world view!
giltman.com - Get gear and gadgets at up to 70% off. Shop like a guy, dress like a man. You're invited.
58 Comments
- SgnDave, on 10/12/2007, -2/+23I don't see why that is inaccurate... reverse hyperthreading means the same to me as anti-hyperthreading. The article, or the translation that you cite, says that multiple cores will look like one. Hyperthreading makes one core look like many.
- thebeck, on 10/12/2007, -3/+20They price based on supply and demand because they can only produce so many chips. If they were cheeper there would be shortages.
- jaguarsavages, on 10/12/2007, -0/+7As much of a supporter of AMD as I am, the upcoming Intel Conroe looks like it will be some seriously tough competition for AMD. Not that that is really bad news though. AMD has been ahead of Intel in terms of performance for the past 2-3 years and some intense, neck-to-neck competition will good for consumers in the end, I think :)
- SgnDave, on 10/12/2007, -1/+8There is some potential for paralellizing some operations. This is the goal of such technologies as out-of-order execution (Intel's term for it) or superscalar/superpipelined processors. Those concepts make "parallel" use of processor resources (think memory, cache, and ALU); paralellizing across cores would be harder, but it IS theoretically possible.
- ZapWizard, on 10/12/2007, -0/+7It is simple.
Hyperthreading is NOT by any means two virtual cores. Read Intel's whitepapers on their website.
Instead Hyperthreading splits Floating Point operations from Integer Operations and can run them at the same time.
This means you only see a boost in performance when doing those two operations at the same time.
This is why Hyperthreading is application specific, and has to be hardcoded.
Intel's biggest marketing ploy was to get microsoft to show two seperate cores in the task manager, leaving many to think that each virtual core was equal.
That is why Dual-Core is better overall then Hyperthreading, both cores are equal, and can do either task.
Making both cores run on a single thread can increase the speed greatly on single-threaded tasks.
That is why Intel's hyperthreading does nothing for single-threaded applications.
I look forward to this technology from AMD, it seems like a perfect solution to bridge today's single-thread apps with Dual-Core processors. - fivefootstep, on 10/12/2007, -3/+10Most important concept to take away from all this: If you're running one thread, and you have a physically multithreaded system (ie, two CPUs/cores), one of them ie idle while the other does all the work. A simple set of instructions for the new K8s could shift power from the idle threads to speed up those processes that *are* running. What AMD is doing is basically a hardware-level version of what multi-CPU system programmers have been doing for years. They (we) made (make) the programs run faster by instantiating them on all threads and piping the clocks theirselves (ourselves). This would allow for a single instance of the final instruction set to occupy RAM rather than several copies of these "CPU plumbing" programs - thus reducing overhead, and solving through hardware a problem that has been around for quite a while.
Translation, if you didn't catch (or care to read) all that: this will allow AMD CPUs to run multi-threaded programs in a multi-threaded mode and single-threaded programs in a single-threaded mode.
Further translation: no more wasted flops. =) I'm happy. Anyone else?
I don't see how this is inaccurate at all, by the way. I speak French, and the quoted text here was apparently handled with Babelfish or some such because it's kind of mangled. The article does, indeed, call it what should be propertly translated to English as Counter-Hyperthreading, "Anti-Hyperthreading," or "Reverse Hyperthreading." -- that's exactly what it is: emulating a single CPU with two or more... albeit with some loss from overhead (crossoverhead? /coin)
+digg to article
-digg to "inaccurate" crier - j0keR, on 10/12/2007, -1/+7Yay, and then they can invent anti-reverse-hyperthreading that emulates multiple cores on the emulated single core in order to increase performance... wait... none of this makes any sence.
- jwolf, on 10/12/2007, -0/+6This sounds wickedly cool. If they can pull it off, it will change the way people think about multi-processor or multi-core systems.
On the other hand, if technology such as this becomes mainstream, I hope that does not detract from the movement toward multi-threaded software design. - inactive, on 10/12/2007, -1/+6Don't bother, every comment above from people that actually know computer architecture and are calling ***** on this is getting modded down. Who could have expected such a thing on a site overrun by ignorant "I don't need to know how the stinkin' compu-tar works" Apple trolls. Coincidence? On the other hand, their comments do provide great comedy for anyone with said computer architecture knowledge. Digg comment moderation is a great example of how the ignorant stay ignorant, how ignorance is a choice.
- CaptainMal, on 10/12/2007, -0/+4This article is apparently inaccurate, but even if it did work the way it says, that would hardly be considered less parallel than with Hyperthreading. The idea described in this translation says it would allow a SINGLE thread to be split among more than one physical CPU. Where as one CPU is used now, they would use more than one to do a single thread. That's more parallel, not less. In HT, you split up parts of a thread and run them at the same time on a single CPU. Here you'd split up the thread and run parts of it on different CPUs.
Would be more faster better if it could work as described without a monstrous amount of overhead, which I suspect would be required to split threads for an app on the fly without instructions from the original designers. - whisperedlie, on 10/12/2007, -0/+4i love getting my technology news from a link to a vague and horribly translated news article from some random french site.
- HappyScrappy, on 10/12/2007, -0/+3Agreed. I can't imagine how this would work, but if it does work, it'll be a huge advantage.
Definitely worth AMD's time to investigate. - inactive, on 10/12/2007, -2/+5The article talks about widthening the pipeline rather than creating a reverse hyperthreading. It was a bad translation of a bad analogy. But hey beleive what you want.
- SuperFarStucker, on 10/12/2007, -1/+4Worthless, as in completely worthless. Even if you were to farm out instructions to the other cores to exploit instruction level parallelism (which is really the only thing some logic in the cpu has a chance at exploiting), it would be prohibitively slow in all but the special cases. Even shared data caches could not rectify this situation. out of order execution units already make the best of the vagaries of system workloads. The best and brightest have long since discarded any ridiculous notions.
- eadnams, on 10/12/2007, -0/+3Isn't it a little late for this? Aren't most apps are moving towards supporting dual-core hyper-threading? By the time this is avaliable to the consumer, wont most of our apps will be set up for hyperthreading?
I'm not an expert tho, so feel free to educate me :) - aggies11, on 10/12/2007, -1/+3Re: SgnDave
Agreed. However improvements to internal core parallelization (isn't that essentially how hyper-threading achieved it's performance gains?) do not nearly add up to the benefits from "treating multiple cores as a single giant cpu".
I guess, best case scenario, is that there will be some ultra smart logic that will be able to determine determine how and what to parallelize, that can achieve performance gains equivalent to if it was done beforehand during software development. But I imagine that is rather unlikely?
But again, the impact of such a development would be huge, as you get all the gains from multi-core/cpu (more processing power at a much cheaper price), without having to go through the difficulty of re-designing your programs to take advantage (which is no trival matter).
I am wary, but hopefull :)
Aggies - inactive, on 10/12/2007, -2/+4There is also the perception that cheaper means worse especially in the corporate world. Or atleast thats one way hector spun the high prices.
- blackax, on 10/12/2007, -0/+2um..... sorry this is bull but if this is real (and by no means is this real) then being able to run a thread on 2 cores is very much a good thing. Now lets sit and think about it if you have 2 cores both at 2ghz and you can use them both to do one thread then you have a 4ghz cpu. so if this could happen then yes oh god yes it is a very good thing!! but it will not happen as they said.
- daofma, on 10/12/2007, -1/+3Hyperthreading is a performance gain for multithreaded applications. Lots of applications run in a single thread, and since their one thread can only be executed by one pseudoprocessor, it goes at half the speed. If you have one core and multithreaded programs, they can only execute one instruction at a time, only simulating more than one thread. With hyperthreading, they essentially occur simultaneously. It has both benefits and drawbacks. Gamers don't like it, people who use their computer professionally generally do.
Gamers will probably continue to like single cores, unless games start to develop for multithreading. - smartpatrol, on 10/12/2007, -0/+1This sounds ridiculous to me. We've pretty much hit the limit with regard to instruction-level parallelism. The A64 can issue 3 or 4 (I don't remember off the top of my head) instructions per clock as it is, but it's rare to max that out. Doubling the width would do almost nothing to increase performance, because the width isn't the limiting factor in most cases.
If it were feasible to execute more than 4 instructions/clock from a single thread on a regular basis, Intel and AMD would already be doing so. Instead, they've chosen to spend their transistors on more cache and multiple cores.
Sounds bogus to me. - aggies11, on 10/12/2007, -7/+8(Comment duplicated in other thread, as not sure which one will eventually win out :)
Re: JackHandy
Are you saying it's innacurate in calling it "Reverse hyperthreading", or that the whole content of the idea (regardless of name), allowing multiple cpu's/cores act as a single "virtual" CPU, is innacurate?
Assuming it's just the name, then whatever you call it, allowing multi cpu/core to act as a single virtual cpu (ie. somehow breaking down a single execution path down for multiple parallel execution) would be truly revolutionary.
Probably because it sounds largely impossible (there are some tasks that are simply not parallelizable[is that even a word? heh] ). But if this is actually true and can be implemented, then it would effectively solve the main challenge to multi-core architectures: Having to develop algorithims that utilize parallelism, which can be notoriously difficult/complicated.
I will digg purely on the potential, and let the rest of the details sort themselves out :)
Aggies - iSlayer, on 10/12/2007, -0/+1AMD's performance has to do with the number of pipes. The shorter the pipelines the better and more efficient of getting things done faster. Conroe are doing this right now with their chips where they have shortened the pipeline to get faster speeds in performance with apps. and it seems like they have pulled it off with of course adding a few other features. Why do you think they reduced their clock frequency? With 100FPS in a game compared to the fastest AMD chip available. I certainly hope AMD have something up their sleeves to kill intel as I am a AMD fan and hope it stays that way. Although its not looking good for them.
- longofest, on 10/12/2007, -0/+1...OR shared cache
- socket, on 10/12/2007, -0/+1This is complete and unadulterated *****.
- t35t0r, on 10/12/2007, -0/+1anti-hyperthreading means message passing in hardware ..there's going to be lots of additional logic on die for this to occur.
- thebeck, on 10/12/2007, -1/+2I don't think it is feasible to increase single thread instruction per clock significantly more than it is now. What I think this technology will do is allow AMD have several small cores using the same amount of transistors as a few large cores. Instead of having several cores with several pipelines each as it is now, each core can have only one of each type of pipeline and barrow idle core pipelines as needed. This way you get the best of both worlds. Good single thread performance, about as good as it is now, and wicked fast multi tread performance thanks to the CPU having several times as many cores as otherwise possible.
- danpsmith, on 10/12/2007, -1/+2Raped on pricing? I bought an X2 that was a much better deal than the PCs packaged with intel chips.
- inactive, on 10/12/2007, -4/+5"How is this different from doubling the width of the pipeline, except being (almost certainly) worse performance-wise?"
It's not. That's exactly what AMD is actually doing. The digg sheep don't understand processor design but anti-hyperthreading / reverse-hyperthreading SOUNDS REALLY COOL so digg !!!!
So please, report this inaccurate. - Luuvitonen, on 10/12/2007, -1/+2I, for one, welcome our new anti-hyperthreading overlords!
Out, out ye demons of stupidity! - mooninite, on 10/12/2007, -3/+4They have already. Mitosis
Intel actually came up with this before AMD. - inactive, on 10/12/2007, -4/+5Ah, I see. I've reported it. What a bunch of dumbasses. I'm in a graduate advanced computer architecture course this semester, so my ***** meter pinged off the scale before I even finished reading the title (funny what a little knowledge does for you).
- HappyScrappy, on 10/12/2007, -0/+1@ZipWizard:
I'm sure it's the fault of Intel's marketing info, but your understanding of HT is completely wrong.
HT is simply having two virtual processors in one core. This used to be called a dual-stream processor. You can schedule instructions from either stream at any time. In Intel's case, they schedule the 2nd stream when the first stream experiences a pipeline bubble.
This was great for NetBurst (P4) because it had that huge pipeline that stalled all the time. But they haven't done it on more recent processors.
Intel's HT doesn't help with single threaded apps because, just like true dual-core system, if there is only one thread doing all the work, you can't take advantage of additional cores, whether real or virtual.
It doesn't have to do with floating point and integer specifically. Intel has been able to schedule simultaneous integer and floating point instructions since the first Pentium with its U and V pipes. - whalesalad, on 10/12/2007, -0/+1Sounds pretty damn cool
- CaptainMal, on 10/12/2007, -0/+1Most of the limits of multi-threaded software design are logistical, not a lack of foresight. If development time doubles in order to multi-thread, but performance increases only 4%, can that be considered a good investment?
More generic cores can only take you so far. I think a move to a PPU (like the move to the GPU) will mean more for leaps in performance gain by splitting out tasks that generic CPUs aren't that "good" at, rather than just getting more CPUs that are equally bad at executing certain types of code. - inactive, on 10/12/2007, -1/+2"I don't see how this is inaccurate at all, by the way. I speak French, and the quoted text here was apparently handled with Babelfish or some such because it's kind of mangled. The article does, indeed, call it what should be propertly translated to English as Counter-Hyperthreading, "Anti-Hyperthreading," or "Reverse Hyperthreading." -- that's exactly what it is: emulating a single CPU with two or more... albeit with some loss from overhead (crossoverhead? /coin)"
Apparently you didn't read the article in french then. It was an analogy that got mangled and these blogs took it as gospel. The article is about increasing the pipeline width not combining operations cross-die. There was some talk about an AMD version of macro-ops fusion as well that also got lost in translation. - CRAZYSWEETGUY, on 10/12/2007, -0/+0The multi-core processors is replacing hyper threading in a way improving upon that in a sense. It is new way of doing something that was done before by having multiple processors. It made sense to hyper thread when there was only one processor. Having two or more processors changes things. When you get multiple processors the possibilities are increased. Hopefully in the future an x number of processors will be able to run certain program(s) while an x number of processors runs another. Right now we may have one processor run one program and another processor run another program with the duo's. In the future we are talking about 4, 8+ processors. This means instead of having one processor per program we need to implement 2-3 processors per program. This would be like hyper threading and revers hyper threading merging together in a sense. Hyper threading could be used on each processor to operate parts of each program if the programs can be written much better in the future. I do not know that much about these things though, still learning about this, so maybe some of what I am saying does not make as much sense as other things being said.
- longofest, on 10/12/2007, -2/+2read the above comments... i thought just like you did until i got thinking... they are basically hyperthreading by taking operations a single thread's operations (floating point vs. integer) and distributing one to one core and the other to another, and will thus be able to make a single thread's operations and distribute them and execute them simultaneously. You won't see quite the speedup of multi-threaded apps, but you could see a marginal speedup for single-threaded apps, which is what the technology is meant for.
- inactive, on 10/12/2007, -4/+4How is this different from doubling the width of the pipeline, except being (almost certainly) worse performance-wise?
- inactive, on 10/12/2007, -1/+1Well, assuming this to be true then it sounds excellent
Although the gaming PC arms race will just get one more step out of hand, SLI graphics cards and two processors on the same core? Double that even? Starting to get into the realm of clusters here, just in one box - jstroot, on 10/12/2007, -0/+0Is mitosis really the same thing as this "reverse hyper-threading"? On the surface, I think no. However, looking into architecture, and how this could feasibly run, it seems they may be nearly the same.
Here is a link to a mitosis explanation.
http://www.intel.com/technology/magazine/research/speculative-threading-1205.htm#section3 - r00723r0, on 10/12/2007, -0/+0This could be something. Perhaps programmers wouldn't have to write for the other cores, and rely on AMD's "anti-HT" technology. It might not be as much of a boost in terms of speed to rely on it, but it would allow people to be lazy without too many drawbacks :D And if one took the time to write for other cores, perhaps the speed increase would be even higher (hopefully) due to AMD's optimization (hopefully) for instruction sets across cores (HOPEFULLY).
While all of this dual-core smack-talk between AMD and Intel isn't sorted out, I'm staying single-core. - mojotooth, on 10/12/2007, -0/+0Hmm... Perhaps this was just a poorly-timed April Fools Joke? It has all the makings.
"For better automotive performance, we have done something that no one has done before. We have removed the six-liter Hemi and attached a HORSE to the front of your car. This blast from the past is the greatest thing to occur since papyrus." - inactive, on 10/12/2007, -1/+1new architecture AM2 chips have yet to be released, i'd wait until then to draw conclusions.
- iMatt711, on 10/12/2007, -3/+1thank god something we can use
- inactive, on 10/12/2007, -6/+4Portuguese: http://www.htk.com.br/noticia.php?noticia=525
- puffarthur, on 10/12/2007, -22/+20Booyah AMD will trump Intel once again! But hopefully they won't rape us on pricing like they sort of did with the Athlon x2
- mattspammail, on 10/12/2007, -5/+3imagine a beowulf cluster of reverse-hyper threaded boxen!
- -Jaguar-, on 10/12/2007, -5/+2Intel would just come up with their own version.
- millixaw, on 10/12/2007, -4/+1If one physical CPU processing two threads (Hyper-Threading) has proven to enhance performance, then why would two physical CPUs processing one thread between them in any way enchance performance?? Sounds to me like all it results in is each core processing half of a thread. Then again, I have a feeling the title is inaccurate due to digg hype (anything AMD gets you all way too excited).
Plus, the whole idea of this deters software vendors from pursuing multi-thread design. Somehow I have a feeling only something like Gentoo would possibly benfit from this. - theratdotus, on 10/12/2007, -7/+45) BEER
-
Show 51 - 59 of 59 discussions
© Digg Inc. 2009
— Content created and posted by Digg users is