This discussion has been archived.
No new comments can be posted.
Looking Into The Power Architecture Future
Related Links Top of the: day, week, month.
How many hardware guys does it take to change a light bulb? "Well the diagnostics say it's fine buddy, so it's a software problem."
Speed (Score:5, Insightful)
Re:Speed (Score:5, Insightful)
(Odd...I feel like I just quoted someone. But I can't remember who.)
Re:Speed (Score:4, Insightful)
Except taht most people stop their research into what chip they want when they see a RBFN with the letters "MHz" or "GHz" printed next to it. Nevermind how other factors influnence true optimality of a chip. I personally would much rather see a standard numerical rating be developed (FLOPS may work), except that some (coughIntelCough) wont use taht in marketing materials because it shows inefficiency. (Much like how Hummers dont print their gas mileage on showroom display materials)
Re:Speed (Score:5, Interesting)
The problem is that no matter what metric you use, it won't fit all cases. Different workloads have different requirements. Personally, I'd like to see programmable hardware...Essentially an FPGA section on CPUs. Programs would provide the OS's scheduler with a circuit layout, and the scheduler would have the layout programmed in when needed.
Each program doesn't necessarily have to have access to the whole grid array, either. The scheduler could divide the array into sections. One section would be for speeding up scheduler operations. The rest would be available to have programs loaded in. You wouldn't even need to erase one program's hardware when another program had something it wanted to implement. With the hardware divided, you could load the new program's code into an empty slot, and leave the old code available for the old program's next timeslice. (To prevent having to reprogram the FPGA section every time the program's turn came about.)
Re:Speed (Score:3, Informative)
Re:Speed (Score:4, Informative)
There are only a few FPGAs which use any sort of non-volatile memory (Actel's Pro-Asic being one). Those would have a limited life.
Re-programmable (Score:4, Informative)
There is a trade-off between speed, reliability, cost, and re-programmability.
SRAM types
Are re-programmable but require a rather slow serial load at boot-up. Reliability in embedded systems leaves something to be deisired since any brownout-induced glitch can create errors that are even worse (harder to recover from) than software glitches because wired logic doesn't have anything equivalent to code checksums or interrupt vectors. Well-paid FPGA designers are versed in the arcane art of self-verifying logic.
EEPROM types
Come alive at boot up and are much more resistant to glitches. Their performance, however, is slow. And you have limited (100,000 maybe) rewrite cycles.
Anti-fuse types
are made by Actel [actel.com]. They have the highest performance and best density. They come alive at boot up and are dead-nuts reliable under the worst of conditions; for example, properly qualified, they can survive the cosmic radiation in spacecraft that would leave other types toasted. The big drawback: the anti-fuse process, which works by melting diodes into short-circuits, is not eraseable.
Desktop systems (say, an add-on FPGA card) would be best served by SRAM types, since you already have a processor that requires gluttinous gobs of puritanically clean DC power. Basement hardware hackers would be better served by EEPROM or anitfuse types (depending on performance requirements), since they don't require super-expensive exotic design software.
Re:Re-programmable (Score:4, Insightful)
However, in order to program a Xilinx part in an embedded system (without a PC attached) requires a way to program a serial EEPROM. Programming this might be a pain, but Atmel (for one) makes serial EEPROMS for just this purpose, and will also be happy to sell you a programming cable.
Re:Speed (Score:2, Interesting)
The drawbacks: you either need a FPGA that you can reprogram very quickly (on the order of nanoseconds) or you need a task that takes a long time and can more than benefit from having to take the time it takes to reprogram the FPGA. The latter is not terribly useful for your average desktop machine.
Re:Speed (Score:3, Insightful)
sure... great idea. just like the idea of putting an array of dsps on the pci bus, to act as a generic accelerator unit.
of course, the problem with both of these is: who is going to program for these? some specialized high performance applications, perhaps (openssl, linux kernel, autoca
a stick is a club... (Score:4, Insightful)
so long as everybody has their needs met, it's a good thing(tm).
Re:Mirror , just in case (Score:2, Funny)
Wow, what a paradigm shift (Score:2)
Sorry, I'm a bit bitter today.
Re:Wow, what a paradigm shift (Score:3, Funny)
Re:Wow, what a paradigm shift (Score:2, Insightful)
when comparing his purchase to that of his neighbor.
I find it ironic that MIPS were dropped from most advertising, in part, because
they were misleading, so manufacturers went back to quoting clock frequency, which
is even moreso.
--
It is still relevent (Score:3, Insightful)
Given that it takes a lot more cycles to process an image - than say - a spreadsheet - I think the consumer very much wants processing speed.
The next hot item will be likely be digital videos actually going somewhere - other than the shoebox
This will be even more cycle intensive.
The market responds to technology in fits and starts - but the analysis says many consumer products are still throttled by the speed on consumer PC
How CONVENIENT. (Score:5, Funny)
SHAME on you, IBM, for causing Steve Jobs' promises not to come true.
Re:How CONVENIENT. (Score:4, Informative)
real speed (Score:5, Funny)
my celeron can probably only do 2 or 3, i'm sure the P4 can top that though.
Re:real speed (Score:4, Funny)
Re:real speed (Score:4, Funny)
G4 doesn't even have a fan on it, and yet it can barely warm a piece of bread!
Correct (Score:5, Funny)
I mean, everybody knows its the cold cathode lights, plexiglass windows, and stickers that make it go faster.
Re: Correct (Score:5, Funny)
You must own a Honda.. perhaps a Civic to be exact?
Re: Correct (Score:5, Funny)
Actually its just a civic LS, but the fake badges fool everyone, and they make my car faster.
I'm also a proud member of the Icy Hot Stuntaz.
Re:Correct (Score:5, Funny)
And I'll let you in on a little secret -- those pins that go in the holes are actually there to slow the CPU down. No need to buy a new processor -- just clip off a couple of the pins on your current PC and it'll go much faster.
Re:Correct (Score:5, Funny)
I tried, but the magic smoke came out of the CPUand it doesn't work anymore (for those who knows, the magic smoke is what makes the chip work; if it escapes, the chip dies).
I agree with Bernie (Score:5, Funny)
Re:I agree with Bernie (Score:5, Funny)
I have a suggestion, get a personlized tattoo on your ARM!
*Rim shot
I apologize.
Re:I agree with Bernie (Score:3, Funny)
Hmmm. What Bernie is desdribing here sounds like an Adaptive Semiconductor System.
Poeple still want more ghz... (Score:2, Interesting)
Re:Poeple still want more ghz... (Score:3, Insightful)
While that's fine for some workloads, with more instructions being executed at the same time, it harms workloads that depend heavily on the results of current calculations to figure out what to do next.
Intel eased the problem by implementing hyperthreading; I'm surprised we haven't seen the same thing come out of AMD's corner.
Re:Poeple still want more ghz... (Score:5, Insightful)
Actually, it makes people think the processor runs faster when it really *does*. Which is why I like their numbering scheme: it compensates for consumer ignorance.
Re:Poeple still want more ghz... (Score:4, Funny)
Also, those drinking straws glued to the top are longer pipelines.
Re:Poeple still want more ghz... (Score:2, Insightful)
Partially. When I bought my 2200+, I understood that it ran at 1800 MHZ. I looked at all the specs first. And then I saw the comparison to the Intel machines that I was looking at. For less money, I bought my AMD parts that put the Intel equivelent to shame. At 1800 MHZ, the Intel couldn't come close. At 2200 MHZ, as AMD wants you to think of when comparing chips, they were pretty close. The Intel chip beat the AMD in a few cata
Re:Poeple still want more ghz... (Score:2)
Re:Poeple still want more ghz... (Score:3, Informative)
My clock speed is 64k hz (Score:3, Funny)
I think it is very important for clock speed - the crystal in my watch runs at 64k hz to keep time which is quite important. Lets see you Solaris or AMD overclockers beat that!
Re:My clock speed is 64k hz (Score:2, Funny)
What else besides games? (Score:5, Interesting)
I do a fair amount of video editing and image manipulation, even still my two year old computer works fast enough for me...
Does the average Joe need the computing power they are given?
Re:What else besides games? (Score:3, Funny)
Depends on your interpretation... (Score:2)
I think you might be seeing a plateau on "average" users purchasing bigger, better processors or computers. If the computer they have can get the job done for them and they aren;t moving on to other computer tasks (ala video editing, rendering, etc..) they don't mind having 2+ year
Re:Depends on your interpretation... (Score:2)
In other words, for Internet and email, Linux is fine.
Comment removed (Score:5, Interesting)
Re:What else besides games? (Score:2)
Re:What else besides games? (Score:2, Informative)
Check your cooling (Score:2, Informative)
Re:What else besides games? (Score:4, Informative)
as a developer i want compilation to be a quick as possible.
also i make heavy use of vmware, which needs as much grunt as it can get.
Re:What else besides games? (Score:2)
Then don't use C++. Seriously.
I've been using Delphi since it was Turbo Pascal, and compile times have been a non-issue for me since I can remember. Even on my 333MHz P2, I've never had a perceptible compile/link time, ever. Even for full rebuilds.
And these days, anyone doing heavy software development should be using Perl, Python, Lua, Lisp, anything dynamic and lightweight.
Re:What else besides games? (Score:2)
Re:What else besides games? (Score:2, Informative)
Granted, it's not an average application, but I do consider myself to be an average joe, so there you go. :-)
Digital Signal Processing (Score:2, Informative)
Contributing (Score:2)
I contribute my spare computing cycles to distributed computing efforts [yahoo.com].
The distributed.net client [distributed.net], at least, does not require the computer to be switched on 24/7 either.
Even playing DVDs only take up about 25% of the CPU time, and I've had no problems with overheating. I sometimes think about all that A/C power and computer cycles being wasted at the university computer rooms.
Maybe O.S. vendors could include a voluntary option doing install to contribute your computing power. They could the
Power is not for PC (Score:5, Interesting)
The PPC is a cousin of the embedded version of the chip, where the performance per watt power usage is relevant. It is hugely successful.
Sales of Apples with desktop POWER chips aren't really significant. Although IBM aren't ready to admit it yet, the battle for the desktop is long over. No amount of performance advantage is going to outweigh the main advantage of the x86 architecture there: backward compatibility, preserving the value of past investments in software for the customer. IBM should know this, as they still make their zSeries mainframes compatible with the 40+ year old 360 architecture for the same reason.
In the PC, unlike most servers, most everything goes through the CPU, which is why for the average Joe raw CPU performance _does_ matter.
Re:Power is not for PC (Score:5, Insightful)
Your post is, for the most part, dead-on and well-put, but I can tell you're not an enterprise Java developer.
Our transaction processing systems were recently moved to Java from C (Solaris on a Sunfire 6800, 8-way SPARC).
Yes, they require more memory. This doesn't really concern us because we spend far less time tracking down dangling pointers and memory leaks now. The increase in memory seems to be about 4x-6x for our system, which still brings it in under a gig.
No, they do not require more CPU. Several parts of our application actually run faster than the C version. I credit the Hotspot on-the-fly optimization crap for this to some degree, but I'm honestly not sure what the deal is. (And I'm our profiling guy. Ain't that sad?
But more importantly, as you mentioned, on big iron the I/O throughput tends to be the bottleneck anyway. Our transaction-processing systems tend to sit happily with significant idle percentages while positively slamming the disks and databases.
We're running inside Sun's Solaris JVM in a hacked-up proprietary version of EJB, using Tomcat for the frontend. I can't imagine that Websphere has much higher overhead, though I could certainly be wrong.
Re:Power is not for PC (Score:4, Informative)
But you're right, I'm not an enterprise Java developer.
Re:Power is not for PC (Score:3, Informative)
That isnt really an issue for COBOL programmers hehe
Re:What else besides games? (Score:2)
Paint. Some effects can be fairly intensive.
CAD and architecture.
Animation and modeling, the interactive 3D part and rendering.
Video editing.
Flash or animated SVG.
Virtual reality, either games or some kind of virtual presence. There is never going to be enough CPU or graphics power to accurately model reality or unreality. As the CPU, GPU and memory continue to grow so will the level of realism and complexity of the virtual world. Snow Crash and Diamond age laid down the gauntlet, its been slow in c
This is why I hate slashdot. (Score:2)
Are you planning on a career in the computing field? If so, please send me your name and address info so I'll know never to hire you.
If you think the only thing a fast CPU is used for these days is playing "Unreal Tournament", you have no business managing (or even logging in to) the kinds of UNIX boxes some of us deal with on a daily basis. Yes, a faster CPU (or 24 of them) is more than necessary for some of the applications we run in
Re:What else besides games? (Score:4, Insightful)
Sensationlist statement (Score:5, Insightful)
Sure, you may be able to optimize a few very common pathways. But you simply can't optimize all of them. Thus a "perfect" algorithm for pathway adaption would again net you one of those 10-20% increases on a general processor. A dedicated machine (e.g. One attempting to calculate PI to infinity) could of course see several times the performance, but then you have to weigh an expensive programmable chip against a cheap custom chip.
Re:Sensationlist statement (Score:5, Insightful)
He is saying that people have run out of the easy optimizations. That it is more important now to concentrate on the performance of the whole package, not just the core.
To that end people providing their own macro designs will allow Power to extend in ways IBM isn't planning on. Need better I/O handling? Somebody might sell it to you. Need a cache controller that handles a high number of outstanding cache requests because your software isn't cache friendly? Somebody might have that too. Need to find these people with these designs? They'll all be talking to each other as part of PowerPC consortium
This opens up avenues for more creative uses that compliement the basic core, and helps bring down design time. Before you might have not even contemplated a custom chip based on a PowerPC design. In a few years, you might be able to glue a few building blocks together to get it.
Re:Sensationlist statement (Score:2)
What history has shown is that general computing is general computing (
Re:Sensationlist statement (Score:5, Interesting)
I think you grok this well.
Clock speed has never been the main factor in the performance of your computer - it's just been a number that works well for marketing. Your typical modern cpu is idle most of the time anyway. When you increase the clock speed, it does increase performance, but not linearly - doubling the clock speed on your chip might only give you a 10% boost or so in terms of real world performance.
I remember back when the Pentium first came out, having two systems with P60s to compare, the only difference between them being that one had 4 times the cache memory onboard and, I believe, a better cache-logic implementation on board. The system with the superior motherboard was in a whole higher class, performance wise, in regards to every task we threw at it, although the effect was much more pronounced on some tasks than others, it was striking in every case.
As CPU power has been growing far faster than IO capabilities, I would expect the same sort of testing with new systems today would show even more dramatic effects.
Better IO handling is very important for many different applications. Just look at the difference between running an application that will fit in cache against one that requires constant work with your main RAM bank. It's huge. So is the difference between a program that will fit in main RAM and one that requires page swapping with VM. Massive difference. Increasing clock speed shaves a microsecond off here or there, but it does nothing about all the wasted cycles while the CPU waits on IO.
CPU speed over the past 20 years has increased incredibly, but IO capabilities in the PC haven't improved at anything like the same rate. Making CPUs smarter (not necessarily faster, but more efficient at using the speed they already have,) using bigger better designed caches and improving IO systems are likely to be much more efficient ways of increasing real world performance than cranking up the clock speed.
Re:Sensationlist statement (Score:4, Interesting)
IBM is surely going to continue to shrink the die and increase clock rate, but it seems as though for the same amount of R&D they feel that there are more gains in performance to be had by looking elsewhere.
My Asynchronous Processor (Score:5, Funny)
RISC (Score:2, Interesting)
i would think the clock speed has more meaning for RISC processors.
Re:RISC (Score:2)
Re:RISC (Score:3, Informative)
Re:RISC (Score:2)
If you really want to compare computers, alway
Re:RISC (Score:3, Interesting)
I'm not an AMD fangirl or anything but they've been at the "interpret x86 instructions with RISC architecture" game for coming up on a decade. Intel only caught on relatively recently with Banias and Dothan (aka Pentium-M, "Centrino") and are finding out the uncomfortable truth that these chips that operate at a slower clock speed are more efficient and do more per cl
Same old (Score:5, Informative)
Seems IBM is embracing open standards (Score:4, Insightful)
If they succeed it doesn't bode well for the x86 architecture, which seems to be a victim of it's own success. They seem to be trapped into just adding faster clocks instead of changing the architecture. They still have neat things like Centrino, but the marketing droids seem to have control over the engineers there. Every update just seems to be a faster clock speed without regards to how much it actually increases performance(I think this is evident in a lot of consumer pc's were they put in the latest and greatest pentium processor but then add in a paltry amount of RAM) I'm not saying I know more than the Intel engineers, I think they are doing a fabulous job with what they have to work with, but...I don't know where I am going with this, I'll just sit back and burn some karma now....damn ADD
Re:Seems IBM is embracing open standards (Score:2)
It seems you haven't been introduced yet. Foidulus, this is x86-64. Him and the opteron family want to have a little talk with you.
Re:Seems IBM is embracing open standards (Score:3, Informative)
It also seems that IBM is a few years late [sparc.org] in that respect. (See: IEEE Standard 1754-1994)
If they succeed it doesn't bode well for the x86 architecture, which seems to be a victim of it's own success. They seem to be trapped into just adding faster clocks instead of changing the architecture.
As much as I can't believe I'm defending the Intel architecture, Intel *has* been modifying their chip design. Out of order instructions, Superscalar execution, instruction p
Re:Seems IBM is embracing open standards (Score:2)
It also seems that IBM is a few years late in that respect. (See: IEEE Standard 1754-1994)
But isn't SPARC (the standardized part) only an instruction set? IBM appears to be opening up much more, revealing details about the actual implementation of the architecture.
What I still can't tell is, will it be possible for a guy with a fab to retool his factory for building IBM CPU's, selling them indefinitely w/o paying the license.
Now *that* would be Open Hardware. Being able to improve
Re:Seems IBM is embracing open standards (Score:5, Insightful)
x86 processors have managed to bump the clock and improve the architecture. You have to do both to be successful. Having higher clock rates IS a benefit even if you do nothing else, so long as the rest of the system can keep up. IMO you can give most of the credit for the improvement of x86 to AMD, which really pushed its limits admirably. They also didn't manage to push clock rates as far (at least, not as soon) so they had to add more functional units and tie them all in, which is exactly what you're talking about, an alternative to increasing cycles per second.
Re:Seems IBM is embracing open standards (Score:5, Interesting)
But AMD very-much uses the x86 architecture, and has long emphasized things other than clock speed. They've already put into action several of the things IBM's Bernie Meyerson seems to think he brilliantly came up with:
- Efficiency: Athlons just plain get more performance per clock than an Intel. There are a lot of factors that contribute to this including the length of the pipeline, but the design just gets more done with each tick. That's less complexity and less...
- Power usage: Athlons have 10-12 pipeline stages compared to the Pentium 4's 30. Between that increased efficiency, and less need for a large cache (big pipeline means frequent cache hits), it can use far less power than a P4 for the same performance - and consequently generate much less heat.
- Interacting with software: Also not new - more recent desktop AMD chips internally clock themselves up and down depending on whether you're idling or running an app. Laptop chips have done this for years. That means the invalid assumption PC novices make that leaving a PC on while they eat lunch will not use much power becomes valid. For the power user, the PC churns out less heat overall since it only pumps heat under peak usage.
There are things the guy lists that are just freakin' out there:Uh, dude, this isn't an episode of Transformers, it's a CPU. AMD and Intel already resolved this issue by building very strong chips that don't fail. Even if physically modifying the chip to lop off the bad parts is possible, I can only see it leading to a reduction in quality of chips produced, with manufacturers knowing that worst case, if it fails, it'll just lop itself to pieces.
thanks for the link (Score:5, Funny)
I hate it when articles are posted about small, obscure companies and then I can't find their website.
It's all marketing (Score:5, Insightful)
Re:It's all marketing (Score:5, Insightful)
Both a Kenworth over-the-road tractor and a Formula 1 car have about 1000 horsepower. But one will accelerate a LOT faster than the other. And one can tow 20 tons of stuff behind it.
Even IF MHz were directly comparable, you still couldn't judge the speed of a computer without considering what that computer was built to do.
Re:It's all marketing (Score:2)
even on my workstation on the job, my P4 2.8GHz HT processor regularly waits and waits for the puny IDE HD to load or seek through a data file or complete a search of the filesystem. it's like wasting access to a genius:
It like guys with short penises... (Score:2, Funny)
Make the chips run cooler (Score:5, Insightful)
So hey, if you're listening harware vendors, see if you can't simply make the dang things run cooler on less power before you speed them up!
Old news (Score:3, Interesting)
Clock speed really does not have a direct correlation to computer speed anymore. It seems like we will see more of the trend of newer, better technology that runs at a lower rate but executes a lot more in one tick, so it is much faster. It seems that it will start at 1GHz and move up to 3 before somebody gets a new idea, makes a new "slow" processor and starts it over...
GHz is the wrong metric (Score:5, Insightful)
Even on the base CPU, the most important metric, I find, is "MIPS per watt". That's what determines how much horsepower you can get off a given amount of cooling, which is the real limiting factor for CPU speed.
*Everyone* knows? (Score:5, Interesting)
So I "dumb" my speach down a bit and give it again. The masses don't want to know how processors work, they don't want to know about architecture, they want an even base line to measure performance. Most people think the Comp Usa rep is ripping them off and they are trying to feel good about their purchase, being an un-educated consumer.
By buying the high clock speed they can compare it to their neighboors and in their heads they have a Super-Fast PC.
I'd like to note that most people I talk to look at AMD like most people look at a Yugo. (remember those cars?) In spite of my advice that an AMD is like a new Honda for $2,000.
That's my 2 cents
Re:*Everyone* knows? (Score:2)
Re:*Everyone* knows? (Score:3, Funny)
How about Mazda?
(yea, ok, fine, I'm one of those people that wont touch a honda - except a prelude SH or an S2000 - with a 10ft pole)
Bogo Mips (Score:5, Interesting)
Re:Bogo Mips (Score:4, Insightful)
"+5 Funny" would be appropriate!
The "Bogo" stands for "Bogus" for those of you who don't know.
I got a chuckle out of this.
Re:Bogo Mips (Score:4, Insightful)
Good example of what morons most moderators are, eh? (my pet peave for a while now)
thats old news? (Score:2)
Uh, can you spell AMD (Score:5, Interesting)
who in the world (Score:2)
what I want to see (Score:4, Insightful)
The most direct processor measurement (Score:5, Funny)
So, the new processor measurement that gets right to the heart of what's important for a CPU performance geek is going to be henceforth the PU. The PU stands for 'penis unit' and it indicates to their fellow cpu performance geeks how big their dick is relative to everyone else's.
The Article: Distilled (Score:3, Funny)
The world is full of incredibly bright people.
If you can beat 'em, change the rules of the game. (Score:5, Insightful)
IBM would prefer customers to come to them and ask IBM, "Which processor is better?" rather than rely on an external, easily verifiable, though not accurate, single number indicator.
The truth, as we all know, is that there is no single metric since each processor has strengths and weaknesses and various applications rely on these strengths differently.
They are also opening their processors to the end users a little more, almost as a jab at Intel. Intel has microcode, but you'll never see it or get to modify it. But the very presence of microcode in almost every modern general purpose CPU means that performance can be enhanced and tailored for each application with very little processor change.
So IBM is letting people get closer with the processor to enhance performance with very little risk or effort.
The kicker is that it's not simple, so only a few large manufacturers and some dedicated homebrewers will really have anything to show for it.
Thus it's a marketting ploy intended to raise questions about current performance metrics in the minds of indecisive consumers.
But then, when has the CPU war ever been about anything but marketting?
-Adam
thoughts on clockspeed and CPU power/ratings (Score:4, Interesting)
When you buy a car, you don't just consider how fast it goes, you consider fuel economy/comfort/quality etc. this could be applied to CPUs
for instance, you buy a new DeLL/HP/whatever, the machine has 3 numbes on it. 12/16/65 - which means, 12 is the general office app benchmark, 16 is the gaming benchmark, and 65 is the mean power usage of the machine.
so a good office machine is a 14/2/30, but if you are playing games, you need a 6/26/130, you don't care as much about the power bill or how fast office computes a6:c6*c14 whatever. these numbers would be linear, as in the nex-gen would just have higher numbers.
Each computers label would ave a description of the rating above the label saying "look at the killer gamer system" or whatever.
I can see the arguement of the system being confusing, but i'd take the least confusing method that was effective, and i think this would be effective.
------------------
Something like this could translate over to server side with web/fileserving/powerreq or something, but it would allow companies like AMD and IBM who have not pushed the MHz myth to the extreme to allow their product to compete on merrit not Mhz.
thoughts??
Re:Ask the damn question! (Score:3, Interesting)
Remember back when all the Pentium chips had "MMX" slapped onto the end of their names? MMX stood for "MultiMedia eXtensions," which were a set of operations programmers could use to speed up certain