The Future of Intel Processors

madison writes to mention coverage at ZDNet on the future of Intel technology. Multicore chips are their focus for the future, and researchers at the company are working on methods to adapt them for specific uses. The article cites an example were the majority of the cores are x86, with some accelerators and embedded graphics cores added on for added functionality. "Intel is also tinkering with ways to let multicore chips share caches, pools of memory embedded in processors for rapid data access. Cores on many dual- and quad-core chips on the market today share caches, but it's a somewhat manageable problem. "When you get to eight and 16 cores, it can get pretty complicated," Bautista said. The technology would prioritize operations. Early indications show that improved cache management could improve overall chip performance by 10 percent to 20 percent, according to Intel." madison also writes, "In another development news Intel has updated its Itanium roadmap to include a new chip dubbed 'Kittson' to follow the release of Poulson. That chip will be based on a new microarchitecture that provides higher levels of parallelism."

Interesting! Cell is making waves after all...

[seebs (15766)]

I think Cell's taught us two important things about heterogeneous multicore:
1. It's fairly hard to develop for.
2. It's bloody fast.

Looks like Intel's gonna be running with it some; that's good news for anyone making a living selling compilers! :) Buy stock in gcc...

gcc?

[everphilski (877346)]

Buy stock in gcc..

Yeah, cause, you know, Intel doesn't make their own http://www.intel.com/cd/software/products/asmo-na/ eng/compilers/284132.htm [slashdot.org]">compiler...

Re:

[walt-sjc (145127)]

It's a joke son. Gcc is GPLed.

Re:

[seebs (15766)]

You know, come to think of it, IBM has a compiler too.

Maybe, uhm... A joke?

(That said, stuff like this IS good news for anyone working on gcc professionally, potentially, although it does have the short-term impact of creating a class of apps where gcc isn't going to be as good as the industrial and research compilers for a while.)

Multicore vs. implicit parallelism

[BritneySP2 (870776)]

While multicores, obviously, have their use, the future belongs to CPUs with massive internal implicit parallelism, IMHO.

Re:

[MikShapi (681808)]

While CPUs with massive internal implicit parallelism, obviously, have their use, the future belongs to electric cars, IMHO.

Re:

[BritneySP2 (870776)]

move that into the processor

In a manner of speaking, yes. For a compiler of a programming language to be able to implement the language's constructs efficiently, there must be an adequate support of those constructs by the target hardware.

On a more general note, the boundaries between hardware and software are always blurred, in that you cannot completely abstract one from another without hurting the performance of the system.

Let's see where this takes us

[keithjr (1091829)]

With process sizes getting smaller and smaller, it is interesting to watch new ideas for as to what to do with that newfound area. The elementary choice seemed to always be "throw on more cores" but the prospects of accelerators and bridges moving into Systems-on-Chips looks like it might have much nicer prospects.

The average parallism factor for most programs tends to hover around four. I think Intel might have figured out that this is a decent stopping point for hardware parallelism as well.

But gee

[MrNonchalant (767683)]

What I really want is a dialogue with Intel engineers about this piece of Intel-themed news. Why can't you add something like that to the site? You could call it something like Opinions With Intel or Intel And Opinions or Center for Intel. No that's not quite right.

This story should be posted 8 times

[Timesprout (579035)]

So we can can have comments in parallel.

Re:This story should be posted 8 times

[MyLongNickName (822545)]

Be patient. The other five articles should appear soon.

Re:

[wbren (682133)]

So we can can have comments in parallel.

I think you ramped up your clock frequency too much. Your instructions are overlapping, causing data corruption in the pipeline and grammar mistakes. :-)

Cell and parallel processing. Answer this for me.

[zymano (581466)]

Why isn't parallel processing used more since more of us will need graphics/math intensive processors? We don't need faster word processors. The threading direction seems misguided to me. Is the state of parallel processing compilers not workable. I don't want to hear about the stupid '4 diggers to dig 1 ditch' analogy. Cliche.

Re:Cell and parallel processing. Answer this for m

[keithjr (1091829)]

Well, the analogy I've always heard was "1 woman can have 1 baby in 9 months, but 9 women can't have 1 baby in 1 month." Lesson here: not everything is as "parallelizable" as digging a ditch. Data dependency in single execution threads means there often simply isn't enough independent work that can be done at once. Moreover, it is often left up to the user (or third party vendors) to create the application library to take advantage of parallel processing. Almost all code being run at this moment was wri

Re:Cell and parallel processing. Answer this for m

[LWATCDR (28044)]

Okay how is threading not parallel processing?
One of the great difficulties of the Cell is asymmetrical in nature. With a Cell you have to do a lot more resource management than with symmetrical multiprocessor system. I have not worked with the Cell but some of the issues I could see cropping up is that it maybe a little light in none floating point resources. With only one PPC core there may be issues with keeping all the SPEs busy.
The 360 is no slouch when it comes to floating point but has a lot more g

For the long term

[ClosedSource (238333)]

Intel needs to develop new processor technologies to significantly increase native performance rather than just adding more cores. Whether multi-core processors can significantly increase performance for standard applications hasn't yet been proven and even if possible, will depend on the willingness of developers to do the extra work to make it happen.

If software developers can't or won't take advantage of the potential benefits of multi-core, Intel and AMD may have to significantly cut the price of their processors because upgrading won't add much value.

Re:For the long term

[timeOday (582209)]

Intel needs to develop new processor technologies to significantly increase native performance rather than just adding more cores.

Figure out how to do that and you will be a rich man. The move to multi-core is a white flag of surrender in the battle against the laws of physics to make a faster processor, no doubt about it. The industry did not bite the bullet of parallelism by choice.

Re:

[0xABADC0DA (867955)]

That sounds like what they are doing, improving performance by making more things native.

For example, they could put a Java bytecode interpreter "cpu" into the system. Java CPUs didn't take off because a mainstream processor would always have better process and funding, and you had to totally switch to Java. But if everybody had a Java "cpu" that only cost $0.25 extra to put in the chip and got faster as the main CPU got faster, then it might actually be useful (incidentally .NET bytecode is too complicat

Re:

[Vellmont (569020)]

I think this is this most intelligent reply I've heard about multi-core processors. Everything I've heard up to this point is the standard "But multi-threaded programming is both hard, and has diminishing returns". Which is very true. I've often wondered how the hell I'd break my programs into 80 different independent parts.

Ultimately I think you're right. Processors started out general, and have become increasingly specialized. First we had the "floating point co-processor", next stuff like an MMU, th

Clock Speed?

[tji (74570)]

It seems that Intel very rarely mentions clock speed in any of their roadmap briefings. The clock speed increases over the last five years or so have been pretty minimal. Moore's law talks about the rate transistor density increases. But, clock speed has followed a similar curve until recently. The last 4-5 years has to be the longest plateau in the history of the industry.

Yes, I know they changed to a new architecture that put less emphasis on raw clock speed. But, given that more efficient architecture, clock speed increases are still going to be a major benefit.

So, what's the story? Has the industry hit a wall? How long will it take to get back to above 3GHz for a mainstream processor, or even to the 4GHz levels that the old Pentium IVs were pushing.

Don't get me wrong, I am a huge fan of the power efficiencies of the new chips. For my primary purposes (laptop, HTPC) the new chips are a godsend. And, the thought of specialized "accelerator" cores is fantastic (a video decoder core for MPEG2 & H.264, please). But, doing that same thing at 4GHz is even more compelling (of course, with the speedstep++ stuff to shut down cores when not needed, and throttle back to low GHz to save power).

Re:Clock Speed?

[ZachPruckowski (918562)]

Penryn (a die shrink of the Core2 Duo/Quad plus some SSE4) should have 3 GHz+ models. The real performance issue isn't clockspeed, it's instructions per second. When you make 128-bit SSE take fewer cycles, and you add execution units, improve scheduling logic, and reduce access latencies (through pre-fetching or larger caches, or faster buses), you make processors faster. A processor that runs at 2 GHz with 3 Instructions per clock is just as fast as one that runs at 4 GHz with 1.5 IPC. The reason clockspeed hasn't been increasing is because performance gains have been coming from other areas. Intel could probably sell a juiced-up 3.6 GHz Core 2 Extreme, but it'd run at 180 Watts or something, and cost like $1500.

Re:

[timeOday (582209)]

The real performance issue isn't clockspeed, it's instructions per second.

Bull. The fact is, the MHz "myth" is mostly true. The vast majority of improvement in processor speed over the past 30 years is due to clock rate, not IPC. The performance gains from other areas over the last 5 years have not kept pace with the rate of progress for the preceeding 25 years, not even close.

Re:Clock Speed?

[644bd346996 (1012333)]

Sure, for most of the past 25 years, it has been the clock speed that's been improving. But that's changed in recent years. When Intel switched from Prescott to Core, they pretty much cut the clock speed in half without really sacrificing performance. That's because they increased the IPC a lot in Core, so that it had comparable IPS.

When comparing different processors with the same ISA (ie x86), IPS is the best measure of CPU performance, not clock speed.

Re:

[Vancorps (746090)]

Tell that to the Amiga guys and to AMD when they chose IPC over clock while the P4 was around. Both are very important. The industry spent years ramping up the clock and now they're spending a few years working on IPC. It makes perfect sense to me. Moore's law also doesn't refer to the frequency of a chip but to the number of transistors which has kept pace especially now with the 45nm processes.

Personally I think for the moment IPC is far more important than frequency given computers are doing more and m

Re:

[Doctor Faustus (127273)]

So, what's the story? Has the industry hit a wall?
Yes. There was a big story about three years ago that when Intel got its first chips from some new process shrink (90 nm?), they were startled to find that they couldn't get them to run substantially faster than the previous version. Up until then, they'd always gotten a significant speedup from that with no design changes, but they did hit some sort of physical limit no one was expecting. I haven't heard anything since about whether they figured out what

New term war.

[jshriverWVU (810740)]

I was just checking out this page here [azulsystems.com] which discussed a machine with 768 cores. While I do a good amount of parallel programming this is good news to me. But it seems for the average person, this is turning into another mhz/ghz war, this time cores.

What we really need is for software to catch up. Luckily some programs like Premiere, Photoshop have supported multiple CPU's for a while now. But games, etc can really benefit from this. Just stick AI on 1 core, terrain on another, etc etc.

Re:

[MikShapi (681808)]

"Software" (as in all software ever written) is not a monolithic thing. The vast majority of software in use today is not CPU-restricted by modern (and even 5-year-old) commodity hardware.

Of the little bit that does need oompf, Where SMP can be taken advantage of, people have largely been working on doing so for a while now.

Only the little fraction that remains - projects that CAN USE the extra oompf and haven't been developed in that direction yet - need to catch up.

Your statement hardly applies to most so

Improved cash management

[gEvil (beta) (945888)]

I've found that improved cash management does wonders for me, like allowing me to buy things like new processors.

Remaining Interchangable

[Nom du Keyboard (633989)]

My thought is: How long can Intel and AMD remain interchangeable? For that matter, how interchangeable will Intel be in the same socket, if processors are going to vary this widely? In is this a good thing?

Re:

[drinkypoo (153816)]

For that matter, how interchangeable will Intel be in the same socket, if processors are going to vary this widely? In is this a good thing?

If intel used just one socket, then you would have portions of a socket unused on some systems, but it would cost less to do the design, because there would be only one design. They don't do this because a socket with less pins costs less.

I don't know if that's what you wanted to know...

Intel and AMD could ostensibly remain eternally interchangeable; they are not and

Where all the CPU time will go

[Animats (122034)]

Where will all the CPU time go on desktops with these highly parallel processors?

• Virus scanning. Multiple objects can be virus scanned in parallel.
• Adware/spyware. The user impact from adware and spyware will be reduced since attacks will be able to use their own processor. Adware will be scanning all your files and running classifiers to figure out what to sell you.
• Ad display. Run all those Flash ads simultaneously. Ads can get more CPU-intensive. Next frontier: automatic image editing that puts you in the ad.
• Indexing You'll have local search systems indexing your stuff, probably at least one from Microsoft and one from Google.
• Spam One CPU for filtering the spam coming in, one CPU for the bot sending it out.
• DRM One CPU for the RIAA's piracy searcher, one for the MPAA, one for Homeland Security...
• Interpreters Visualize a Microsoft Office emulator written in Javascript. Oh, wait [google.com].

Re:

[walt-sjc (145127)]

Keep in mind that many of those tasks are also very I/O intensive, and our disk speed has not kept up with processor speed. With more cores doing more things, we are going to need a HELL of a lot more bandwidth on the bus for network, memory, disk, graphics, etc. PCI SuperDuper Express anyone?

Size doesnt matter to me.

[jshriverWVU (810740)]

I for one do a lot of cpu intensive coding, so I *would* use a 1thz processor. One thing I dont understand, they kept wanting to get more ghz for the same size an eventually hit a barrier. So why are we stuck on having a processor so small? I recently bought a 3ghz CPU and it was about the size of a 50 cent piece, and the actual core was smaller than a dime! 3ghz in less space than a dime! Cool, but why can't they just extend outwards?

I wouldn't mind going back to the days when computers were bigger if i

Re:

[bcmm (768152)]

I don't think size is an issue really. Faster cycling doesn't come from adding transistors, it comes from making things happen faster. If anything, putting things closer together helps.

Re:

[smoker2 (750216)]

3ghz in less space than a dime! Cool, but why can't they just extend outwards?

Three words :
Speed Of Light
The clock speed (of a cpu) is limited by the speed of light, and the bigger the chip, the further stuff has to travel. Even at light speed, you can only go so far and get back again in a certain time.
I'm not brilliant at explaining this, but I'm sure someone else will pick this up.
In the meantime, have a look at this interesting paper [www.gotw.ca] from 2005.

Programmable Cache/Storage

[Doc Ruby (173196)]

Cache's are cool, because they're automated to solve a common chip problem of faster access to more frequently used data, without any extra programming. But they're a pain, because they're a blob that extra programming can't do anything else with. If Intel could just add some programmatic access to core caches (including flushing and swap in/out to main or other-core memory), which otherwise could serve higher performance at some cycles, they'd solve a lot of these problems with little investment.

Conversely

Re:

[serviscope_minor (664417)]

Intel have added some programmer control over the cache. Look at the prefetch, movnt and sfence instructions. They're only really hints, but they do help.

Time to dig out your instruction set manual... :-)

More energy efficient chips...

[Nim82 (838705)]

I'd much rather they focussed on making chips more energy efficient than faster. At the moment barring a few high end applications most of the cpu power on the majority of current processors is largely unused.

I dream of the day when my gaming computer doesn't need any active cooling, or heat sinks the size of houses. Focussing on efficiency would also force developers to write better code, honestly its unbelievable how badly some programs run and how resource intensive they are for what they do.

Re:

[drinkypoo (153816)]

I'd much rather they focussed on making chips more energy efficient than faster.

Primary enemy of electronics is heat caused by inefficiency. By moving to a smaller process we reduce voltage, thus we reduce power (P=VI) and thus we reduce heat. So we can go faster. But we can also not go faster, and go lower power. VIA is the current leader, AFAIK, in low-power x86-compatible processors/systems. But beyond their equipment, much of which is very sad and slow, you can simply underclock any CPU and depending

Energy Efficiency

[zentec (204030)]

The thing that is the future for Intel is not only the bizillion cores and cheaper/faster, but to do so with outstanding energy efficiency. This is obviously important for portable computing, but it's also important to reduce heat load and power consumption in large data centers. Cost of ownership comparisons have yet to include power consumption, but as green house gas taxes start making their way onto electric bills, it's likely to be a selling point.

More and more there's a need for extremely energy efficient, low footprint devices for special purpose applications. It just doesn't make a lot of sense to have PC sucking 60 watts when all you need is something to run Minicom to a simple 15" LCD screen.

Multiple cores appear as one

[gilesjuk (604902)]

That's what they need to do. Rather than make one chip look like two, it's easier to get max performance by making more than one core appear as one.

Re:Instead of more power

[CajunArson (465943)]

That would also improve overall security too.

I hate to break it to ya, but in a low-level language like C, doing proper bounds checks and data sanitization required for security does not help performance (although it doesn't harm it much either, and should of course always be done)
    There is a lot of bloated code out there, but the bad news for people who always post "just write better code!" is that the truly processor-intensive stuff (like image processing, 3D games) is already pretty well optimized to take advantage of modern hardware.
    There's also the definition of what "good code" actually is. I could write a parallelized sort algorithm that would be nowhere near as fast as a decent quicksort on modern hardware. However, on hardware from 10 years from now with a big number of cores, the parallelized algorithm would end up being faster. So which one is the 'good' code?
    As usual, real programming problems in the real world are too complex to be solved by 1-line Slashdot memes.

Re:

[Nikker (749551)]

I think with all of these cores and such an increase in on die cache we should be asking what can we accomplish by staying on-die? As the number of cores increase so will on-die cache, when we start to get into 10MB+ area we could likely do some pretty fancy stuff, also treating registers as memory on idle cores will add to this. With all this micro-logic maybe even simple operations add + move ops will be added to the off-die ram as a type of pre-processing.

The more cores they add the more the system w

Re:

[fitten (521191)]

Define "bloat". For example, do you classify 'features', as in adding more of them, as bloat? I think the word "bloat" is thrown around so much that few people have a good definition of it anymore. For example, features (what lots of people call 'bloat') that aren't used *shouldn't* cause performance issues as the code for them isn't executed.

Besides, if we stopped adding features, we'd still be using things like ed for editing (and 'word processing'), our games would still be like Pong, and our remote a

Re:

[morgan_greywolf (835522)]

Better code = less bloat = better performance and security.

The parent's point is that in code where it makes a difference, the code is already thoroughly optimized, in general. Slimming down the code for Microsoft Word or XEmacs or Firefox or Nautilus or iTunes (there, now we've slaugthered everyone's sacred cow!) isn't likely to make much of a difference because apps like these already run plenty fast on modern hardware. Sure, bloat is bad, but it's a lot harder to remove bloat from existing code without

Re:Instead of more power

[ichigo 2.0 (900288)]

For some reason javascript is slow on all browsers. I believe there is a W3C spec that mandates it.

Re:

[Vellmont (569020)]

Better code = less bloat = better performance and security.

The thing you've failed to realize is that "bloat" is relative. One mans bloat is another mans "gotta-have-it" feature. Also the point of the poster was that "better performance" is a moving target.

Programmers don't design software for one guy, with one computer, that's run only next week. They design software for a hundred/thousand/million guys that runs on 200 different computers of different speeds, and for the next several years.

The basic tak

Re:

[Vexorian (959249)]

hmmm how about:?

Optimization = more specialized code = less maintainability = bugs are worse = adding features adds bloat = security issues

More powerful processors = less need for optimization

More powerful processors = Compilers take less time to do their job and developers get more time to work on their applications efficiently

Re:

[WrongSizeGlass (838941)]

Who's going to need 80 Cores? *ducks*

Any one wanting to run Areo on Vista Ultra Optimum Utmost Paramount Ultimate Quintessential Home Edition?

I, for one, am betting Intel loses its shirt on this 80 Core hodgepodge. That's why I'm investing my entire retirement saving in Transmeta's Crusoe line.

Re:

[walt-sjc (145127)]

What would a duck do with 80 cores? Quack in harmony?