 |
Intel builds 80-core prototype
Author: Tim SmalleyPublished: 18th January 2007
The Tera-Scale Teraflop Prototype chip looks promising for the future - energy efficiency is a major focal point.
The prototype was built so that the chip giant's researchers could investigate the best way to make such a large number of processing cores communicate with each other. This was in addition to researching new architectural techniques and core designs.
The chip, dubbed the Tera-Scale Teraflop Prototype, is just for research purposes and lacks a lot of necessary functionality at the moment. However, R&D Technology Strategist Manny Vara said that the company will be able to produce 80-core chips en masse in five to eight years.
Currently, the prototype chip consumes less than 100W of power, which is less than the 130W consumed by the quad-core QX6700. Of course, the prototype currently lacks some key functionality, which could potentially throw the power consumption characteristics out of proportion, but it's an impressive feat nonetheless.
Vara added that although there are many more cores on the Tera-Scale prototype, they're a different type of core than the ones used in today's microprocessors. "The new ones will be much simpler. You break the core's tasks into pieces and each task can be assigned to a core. Even if the cores are simpler and slower, you have a lot more of them so you have more performance."
Today's microprocessor cores are very flexible, while Intel believes that tomorrow's microprocessor cores will be much more specialised, but of course, there will be many more of these simpler cores. AMD's Fusion project appears to be going down the route of scaling what we've already got, while Intel is moving towards what would be a more flexible approach to energy efficiency.
Before you get too excited though, this is all on paper at the moment; the real war of the cores won't be decided until both companies have released their respective massively multi-core processing architectures in a few years time.
Recent articles on bit-tech.net
Comments (30)
Stats: 0.185 seconds
I think that this may be a step in the right direction, but it could just turn out to be the next GHz race.
Love Microsoft personally, Since I use them everyday in my Job and at Home, I'm just a bit biffed at my Vista installation last night blue screening, then hogging all my resources when it was finally installed :)
Though on another note, yes you are correct about the speed race. Just hope they keep the number of products down, as I can Imagine them doing 40 core versions @ 8ghz, 40 Core versions at 8.5ghz, then 60 Core versions of same values etc, could become a bit messy!
If you specifically dedicate a core to hd video and suddenly nobody wants to do hd video does that core become redundant? If this is the case an out of date cpu will have lots of useless cores doing nothing.
Am I wrong?
But that aside, I'd love this. I really hate interacting with my fileserver, and that's solely because it's the only single-core machine I have left. Since getting a dual-core system, I've vowed to never use less again. I can only imagine how much better a quad would be not having used one, so this is just a mindgasm.
Jamie - I don't think that the cores are permanently dedicated to a specific task, I think that pic was just demonstrating how you could assign tasks and seemed to imply it would have a few reserved for such (not entirely unlike how Via CPUs have that hardware encryption that kicks the hell out of any madly clocked AMD or Intel chip by about an order of magnitude, with a quarter the clock speed). Of course, if you're talking 80 cores, I don't think having a few specialized would be a bad thing, especially since they tend to be massively faster at task X if it's designed solely for that purpose. You're talking no more than 5% wasted cores IF (!) that task was to go away forever, which seems pretty unlikely in the case of HD and graphics at least.
Sure they can perform all of these tasks (possibly simulaniously, assuming sufficient bandwidth) But how much of a performance gain would you really get? They are dedicated but simpler. As far as I can tell they would boast no performance gain over traditional proccessors.
So yeah, what Firehed said...
ch424
It would work if they are basic CPU cores that can redistribute tasks on the fly, and configure the CPU array flexibly to suit the application. Ideally, to the software, the array would appear like a single CPU --just a really powerful and ideally suited one.
In terms of orders of ten, let's say current quad-core CPUs have 1 billion-ish transistors, and this has comparable thermal/electrical properties for 100ish cores, that means about 10 million-ish transistors per core... so Pentium II/III sort of level!
As for redistibuting tasks on the fly, that's presumably what these engineers are investigating.
There are other advantages of having 80 symmetric cores of course: you can put the manufacturing yeilds waay up by disabling the one or two cores that have defects, like they do with graphics chips (the X800GTO2 cards for example).
ch424
I think this is a step in the right direction, but I'm not sure how many cores you will need, 80 just seems like overkill, but I'll have to see 5 years from now how things are.
I think that's the idea they should run with... I can see PC's looking very different in years to come. I envision being able to emulate operations and functions that are currently hardwired to the mobo with these processors, rendering the mobo to a simple socket and a few connections for hardware. Probably wouldn't be quite that simple, but I can see the potential to do so much with very little.
As a (Totally inaccurate, but representitve) example, it'd be the difference between doing standard multiplication (9*9) and repeating the values and adding them up (9+9+9+9+9+9+9+9+9). In essence, they're the same action, but because the RISC processor is missing the functionality to do the * bit, it has to go the long way round, slowing everything, especially as instructions in software are an order of magnitude slower than those hard-coded into a chip (1 or 2 cycles per instruction for hard-coded, compared to 10-20 for software driven, what with all the memory reading and writing required).
And consider, Apple moved away from RISC/PowerPC, as the performance was nowhere near as good as it's CISC/x86 equivalent.
Well, TFA states that we'll be seeing these in 'five to eight' years. If we take that as being about 6 years (it's an easier number too), and look at how Moore's Law predicts CPU manufactuing process (halves in size every 18-24 months):
72 months, divided by 18 = 4 cycles of Moores Law.
So if we're at 65nm gate lengths now, we could be at 8.125nm in 6 years. :p
It's horribly inaccurate but the point is they'll be tiny compared with what we have now. Supposedly a decade ago, everything was built on 500nm processes!
There would be other huge factors involved in actually getting to that scale of manufacturing in reality, too.
Warning: You could get lost in this link for days :p
I think the planned order is 65nm --> 45nm --> 32nm --> 22nm --> ??
And I'm not sure, but, I think the "halves in size" relates more to area than length...so 65/2 isn't equal to 45...but 65^2 = 2*45^2 (roughly). For the sizes in my memory (130nm, 90nm, 65nm, 45nm, 32nm, 22nm), the halving of "area" seems to hold more true than the halving of "length." If that holds true, then after 22nm we'll see something around 15 or 16nm.
Ummm, sortof. In my version of Windows XP-Pro (and on my old Hyperthreaded Intel Pentium 4, that looks like 2 processors to the OS), if I go into task manager (ctrl-alt-delete), and look at the programs running, I can right click and set the processor "affinity" for each one -- so I can make individual programs run exclusively on either of my virtual CPU's. I'd imagine you can do the same on true dual and quad-core processors.
Found it pretty interesting myself :)
If you plotted date/process size, you'd see a pretty strong correlation along x=y, but it's obvious that the Law isn't much of a Law, as instead a general rule of thumb. There's something like this on the Wiki article, which better explains it.
It's been sagging under expectation, but climbing back again. :)
While Windows Vista may be multi-threaded it damn sure is not massively multi-threaded enough to make of more than about 4 cores or so.
Then all programs will have to almost completely re-written to actually make use of this kind of processor - or it will just end up running slower than before.
So, I think these might make it into server markets, but not consumer markets.
Thanks for the further info...I got more interested, since the image you linked didn't go far enough to include Core2Duo...so I actually looked up some more numbers. Interestingly, I found that the Pentium D 900 series had 376 Million transistors, while the Core 2 Duo dropped back down to 291 Million. Of course, the Core 2 Quad jumps back forward to 582 Million transistors.
I found the info in the tables on this and this page.