They have tremendous demand for mathematicians that can develop models to quantify risk.
This is not a trivial problem. It is quite technically challenging and requires very sophisticated mathematical skills. Oh, and you'll make more money than God.
"30GB should be enough for anybody" -- say the CEO OF A MAJOR FLASH AND DRAM company.
Wow - I'm shocked!
People will always find ways to fill up what you give them. How about redundant storage for reliability? Effectively infinite undo/save for documents (Never worry about "saving" a document again because every keystroke is saved).
etc...
That's just out of my ass. I'm sure other people have many more usages for more space.
>> Can anyone tell me applications of dual core for a on-the-go computer?
The Yonah core does not have SMT (a.k.a. Hyperthreading). So having 2 cores will allow you to run 2 simulataneous threads of execution and will make your computer feel more responsive.
They're creating one of the planet's most powerful computer science think tanks - and they're making tons of money doing it which means they ought to be around for a while.
EU: The Microsoft monopoly has done very well here in Europe. You own, or you control, just about every computer in the continent. The licenses were grandfathered in so there was no problem with the EU Commission....Well, let's cut out the bulls--t. I don't want to spend any more time here than I have to. You can have the license - the price is $2.4 million a day, Mr. Gates.
Gates: Now the price for the license is less than $20,000, am I right?
EU: That's right.
Gates: Now why would I ever consider paying more than that?
EU: Because I intend to squeeze you. I don't like your kind of people. I don't like to see you come out to this clean country with your oily skin and dressed up in those nerdy glasses, and try to pass yourselves off as decent Europeans. I'll do business with you but the fact is that I despise your masquerade, the dishonest way you pose yourself. Yourself and your whole f--king corporation.
Gates: We're both part of the same hypocrisy. But never think it applies to my monopoly.... My offer is this - nothing. Not even the lawyer fees, which I would appreciate if you would put up personally.
It's not capitalism if the GOVERNMENT puts up money to fund something that the market otherwise would not have bothered with.
The market is ALREADY allocating capital to its most efficient use.
By definition, the government is intervening to redirect that capital (tax dollars) towards ends that the market would not otherwise. There is a built-in inefficieny.
Note: I am not saying this is right or wrong - just that it is.
If we completely eliminated NASA, maybe the market would colonize space faster than the government (over the long run) because instead of the tax dollars going to NASA, they would stay in the economy and be put towards more productive ends - perhaps funding unforseen technologies that would assist space travel... but that is hard to say.
SG Cowen is a wall street investment bank. Their analysts get paid to do research on publicly traded companies.
They may have a position in google stock in which case they would try to find information to suggest that the stock should be priced higher. They may do investment banking business with google which introduces a conflict of interest. By law (regulation full-disclosure), the analyst needs to disclose such positions or investment banking relationships if they exist.
Their clients may have positions in google stock or be interested in buying or selling google. They pay the analysts for their reports. Accurate analysts develop good reputations and draw more business to the bank.
You can dramatically reduce leakage by tweaking the process to give you a slightly slower process. It's not the end of the world folks. It's just at this point in time, it makes more sense to have the faster process and pay for it with leakage power. In the future that may or may not be true.
Slower for TSMC or UMC. Intel has no problem with yields at 65nm. Intel is after all going to launch dual core 65nm NOTEBOOK chips early next year.
Intel is also laying down 3.5 billion dollars for a Fab in Israel which will be 45nm for 2007 production. Moore's law continues to have legs. It is ASTONISHING that this is something we take for granted! And in 2009, I predict 32nm. And in 2011, I predict 22nm. Guess where we'll be in 2013?
the area savings on area & increase in performance is no longer that much moving from one process node to another
Explain that to NVIDIA or ATI, where performance scales directly with the number of floating point units and hence is directly proportional to area.
The problem with designing for smaller geometries is that the ASIC tools aren't mature enough in time for the new process. Design rules increase geometrically as the process shrinks. Intel and AMD have in-house custom tools that deal with these problems since there is a very tight relationship with the Fab. TSMC and the like are more de-coupled from companies like Cadence or Synopsys.
To quote good old Jerry Sanders - "Real men have fabs"
I haven't seen major PC distributors selling clean boxes
If there were a market for clean PCs, don't you think someone would have built a company around that already?
There must not be enough profit to make this venture worthwhile or you'd see a plethora of companies trying to do this. Any costs that can be wringed out of a PC can be used to undercut competition and directly benefit the bottom line. I recently read a story about how Dell wants to get rid of all of the Intel and NVIDIA stickers so that can shave a few pennies off of the cost of the PC (and also reduce assembly times). If they're fighting to remove stickers, surely there is a huge incentive to cut out $35 of software expense.
The non-existence of such options seems to suggest that there is no money to be made there. People simply don't want the free alternative. They derive value from Windows and are willing to pay for that. The value is that the number of applications available on Windows vastly exceeds the number available on Linux. Couple this with the fact that a large number of popular linux applications also run on windows (cygwin, openoffice, firefox etc...) So for $35 (or whatever the OEM windows cost is) you get almost unrestriced choices in software applications.
It's extremely difficult for Linux to have the same number and quality of applications that Windows has. Linux has such a small market penetration, there is little financial incentive to develop software for it. The only reason to develop profitable software for Linux is if Linux can do something that Windows can't. There are many cases for which this happens, but the applications are mainly restricted to the server world. In the desktop space, nobody really has any answer to Microsoft.
For linux to succeed in the desktop, it has to provide a value above and beyond what Microsoft offers. I have to say - in my experience, it doesn't. Sure it can come close to COPYING microsoft, but where is the innovation that allows users to do things that Microsoft can't do? Keep in mind that we're talking about the OS here. If you develop a novel application for Linux, chances are it can be ported to windows (think Firefox). There simply isn't much value in kernel that the average user cares about.
The story is of course different for servers, which is why you see real corporations use Linux in the enterprise.
Look to Apple as an example of something that is successful in the desktop. They know how to innovate.
So this raises the question of why Linux can't innovate beyond copying windows? The answer, I believe, is that there isn't enough financial incentive to innovate with Linux. Linux innovation is dependent on a community which doesn't get paid. Do you go to work and code for free? I thought not. The best and brightest want to get paid.
The only hope is that hardware companies like Sun and IBM want to reduce the cost of OS software to 0 since it reduces the costs of their systems. They have the incentive to pay developers to innovate for Linux. They just, quite frankly, haven't been able to compete with Microsoft or Apple. Hopefully that will change in the future. I'd like to have cheaper computers just like everybody else.
That's what I meant when I said "They might lose a strategic bargaining position with Intel"
There might be a dozen or more reasons why Dell might not want to go with AMD. It's actually pretty easy to speculate on why this might be the case. Dell actually has the data to evaluate what's best for Dell (as opposed to some random slashdot person saying Dell is "dumb" for not choosing AMD) - and given the success of the company, I'd say they're doing a pretty fine job.
Yep. Michael Dell sure is fucking it up. I mean - he's only the 10th richest man in America and he's already 40. What a slacker.
If ONLY he used AMD. Then he could really make it.
I mean, they only made 3 billion dollars in pure profit last year. They somehow gain marketshare every single year.
They should adopt a strategy like Gateway or IBM. Oh, wait, IBM is out of the PC business because they can't make any money. Gateway loses money every single year...and HP makes the majority of their money from printer INK.
Perhaps using AMD would increase costs MORE than it would save them. Since they'd have to now support Intel and AMD motherboards and chipsets, effectively doubling the number of configurations they need to support, configure and manufacture. They might lose a strategic bargaining position with Intel causing their Intel costs to rise. Maybe they have visibility into Intel's roadmap that makes them believe Intel will offer more competitive processors in the future. Maybe they are worried about AMD as a reliable supplier with their relatively small capacity. Whatever... I'm sure the bean counters are on top of it and it makes no financial sense to do it at this time.
Look. The PC business is notoriously cut-throat competitive and survives on extremely LOW margins. If Dell could gain a price advantage from using AMD, they would. They're even trying to negotiate getting rid of all the Windows and Intel and Nvidia stickers so they can shave a couple of pennies off of the cost.
As dominant as Dell is, they still only have an 18% market share. There are plenty of other vendors.
Nobody is forcing people to shop at Dell. If Dell is doing a disservice to their customers, those people would shop elsewhere. Clearly this is not happening as their market share increases every single year.
The reason is that many mobile devices don't have the horsepower to do H.264 yet.
Yes, it's higher quality and smaller, but the computational cost (and therefore battery/power) is quite a bit higher.
Re:Writing has been on the wall
on
Dell Dumping Itanium
·
· Score: 4, Insightful
I really do believe it has no technical merit.
I am paid to design processors and have worked on SPARC, MIPS and x86 designs for a span of over 12 years. I spend my days thinking about how to improve processors. That's all I do... all day long.
So please... enlighten me on how the Itanium architecture improves computing on any metric.
Any performance advantage that you see today is solely due to their having much larger die size and pin count budgets vs. other processors just to compensate for their having a crappy ISA. If you give the same budget to a comparable x86 or traditional RISC processor, their absolute performance and performance/watt would far exceed any Itanium.
Put a 9MB cache on an Opteron and see how well it does on SPECFP for example. An Opteron beats the Itanium 2 handily on integer code with just 1MB of cache.
Re:Writing has been on the wall
on
Dell Dumping Itanium
·
· Score: 5, Interesting
x86 processors have a fixed amount of decoder logic overhead vs. RISC. The decoders essentially dynamically translate the x86 instructions into more machine-friendly micro-ops which are very RISC-like.
As transistor budgets increase exponentially (thanks to Moore's law), that fixed overhead gets smaller and smaller each generation - to the point that it's insigificant (less than 5% today and getting smaller tomorrow). So in the early 90s you could make a case for more efficient computing with RISC vs. x86, but today it's just so negligable that you don't care. There are also numerous micro-architectural tricks to get around the limited registers and wacky addressing modes.
Couple this with the fact that 99% of all of the world's software is written for x86 and you have a very large inertia to overcome in order to change the ISA.
Why would any software vendor port their application to a new architecture if that architecture is brand new and nobody is using it initially? This is a very expensive and risky task. Let's say that the incentive is increased performance with a new ISA (highly unlikely given that the ISA doesn't matter anymore given the very large transitors budgets). But let's be generous and give it a 50% performance advantage (again - this is fantasy land). Do you spend the 8 months porting, debugging, testing Photoshop? Or do you just wait 8 months for a 50% faster x86 to come out and instead spend that time improving your product as opposed to keeping it the same on a different architecture?
You'd have to be crazy to take that tradeoff. And so, you see what we have today - x86 everywhere.
Many of the Alpha team went to work on the Opteron. Or did you think that it's resemblance to the Alpha was a coincidence?
The fingerprints are there:) Sure, some of the Alpha engineers went to AMD. But there is still a core design team in Mass. that currently works for Intel. You didn't think that every single engineer just uprooted their family to move cross country to the Bay Area did you?
Intel doesn't just have the Alpha team. They also have the ex-PA-RISC team from HP. I hear they're not too shabby either:)
Bottom line is this - Intel has high quality, experienced engineers in quantities that exceed the competition by a wide margin. This doesn't always guarantee success because management can always screw it up, but it's a pretty significant advantage.
The server market is shrinking, and the laptop market is expanding
Your data is incorrect. The server market is growing very quickly as is the laptop market. Hell, even the desktop market is growing (but at a much much slower pace).
Itanium is all but dead... relegated to the supercomputer niche - and we all know what happens to supercomputer companies:)
Intel has spent billions on Itanium and seen an effective return of 0%. Investors won't tolerate this for much longer. AMD's x86-64, and Intel's subsequent introduction of EMT64 (same thing), have finally pushed this ill conceived idea into its well deserved death spiral.
It has no technical merit. But technical merit sometimes is a secondary matter in the business world. However, the economics don't make any sense - you can't introduce a new ISA into a mature software market and expect it to fly just because you're Intel.
It was a mistake - write it off and move on.
This should free Intel to deploy those valuable Itanium engineers (like the ex-Alpha team) to work on something that actually generates cash (like x86 servers). So while AMD might have a short term lead - the giant resources of Intel are more than enough to catch up and re-assert their leadership position.
The good news is that if the particular algorithms you're interested in have a favorable ratio of computation to memory access (aka good data reuse), then you are loving the new parallel world and your code is getting faster at the rate that computational capability is increasing (which is greater than Moore's law.)
How can you scale faster than Moore's Law?
If your application is compute bound and parallel, then you increase capability by increasing functional units which are directly proportional to transistor density (aka Moore's Law).
When NVIDIA draws their pretty graphs showing faster than Moore's Law scaling, it's because their die size hasn't been maxed-out in the past and is increasing in size. When NVIDIA chips become reticle limited (upper limit die size), they will also scale with Moore's law.
Also, there aren't many restrictions on STLF, mostly size (a load of 128 bits can't be forwarded from a store of 32 bits, although the converse is possible) and alignment.
There are partial address match problems because the P4 doesn't use all of the bits.
In common situations and if the compiler generates proper code, STLF restrictions shouldn't be very restricting at all.
Yes, store forwarding does help alleviate latency, but the forwarded data doesn't come from the renamed registers, it comes from a store-buffer (in some cases it can be just as fast as a regfile). Also, because detecting store-load dependencies is hard (think about how to do it in hardware), there are many restrictions on the kinds and types of forwarding that can be done at full speed.
As you mentioned, it's very inefficient to have spill/fills instead of register accesses - you can acheive similar performance, but most likely lower performance-per-watt.
In particular, there has been a longstanding disconnect between the growth in the amount of memory bandwidth available to chips versus the amount of computation that can be done on them. Computational capacity is growing much more quickly than memory access
I think you're confusing memory latency and bandwidth. Memory latency isn't improving as fast as computation or bandwidth.
In order to hide latency, you do more things in parallel while you wait for your long latency memory operation to complete. Finding many things to do in parallel in a single thread is hard to do automatically by the hardware (ie - out of order execution). And by "hard to do", I mean power hungry. A more power efficient way to do this is to put the burden on the programmer to partition the problem into parallel tasks and explicitly map it onto multiple execution units.
Register renaming does nothing to alleviate the scarcity of registers in x86. The programmer does not see the renamed registers and so must use only 8 registers (or 16 with x86-64) and still spill/fill to memory.
Register renaming is required for out-of-order execution and does not eliminate the spill/fills. You're mixing two different concepts.
Slashdot Mirror
Go to Wall Street.
They have tremendous demand for mathematicians that can develop models to quantify risk.
This is not a trivial problem. It is quite technically challenging and requires very sophisticated mathematical skills. Oh, and you'll make more money than God.
"30GB should be enough for anybody" -- say the CEO OF A MAJOR FLASH AND DRAM company.
Wow - I'm shocked!
People will always find ways to fill up what you give them.
How about redundant storage for reliability?
Effectively infinite undo/save for documents (Never worry about "saving" a document again because every keystroke is saved).
etc...
That's just out of my ass. I'm sure other people have many more usages for more space.
The answer is pretty complicated but I can condense it into a slashdot sentence. You do it because it's in your own self interests to do it.
For more information about the evolutionary reasons for this, I'd suggest reading "The Origins of Virtue" by Matt Ridley.
>> Can anyone tell me applications of dual core for a on-the-go computer?
The Yonah core does not have SMT (a.k.a. Hyperthreading). So having 2 cores will allow you to run 2 simulataneous threads of execution and will make your computer feel more responsive.
I'd say it's more like Bell Labs.
They're creating one of the planet's most powerful computer science think tanks - and they're making tons of money doing it which means they ought to be around for a while.
EU: The Microsoft monopoly has done very well here in Europe. You own, or you control, just about every computer in the continent. The licenses were grandfathered in so there was no problem with the EU Commission....Well, let's cut out the bulls--t. I don't want to spend any more time here than I have to. You can have the license - the price is $2.4 million a day, Mr. Gates.
...
Gates: Now the price for the license is less than $20,000, am I right?
EU: That's right.
Gates: Now why would I ever consider paying more than that?
EU: Because I intend to squeeze you. I don't like your kind of people. I don't like to see you come out to this clean country with your oily skin and dressed up in those nerdy glasses, and try to pass yourselves off as decent Europeans. I'll do business with you but the fact is that I despise your masquerade, the dishonest way you pose yourself. Yourself and your whole f--king corporation.
Gates: We're both part of the same hypocrisy. But never think it applies to my monopoly.
My offer is this - nothing. Not even the lawyer fees, which I would appreciate if you would put up personally.
Cue dead hooker...
It's not capitalism if the GOVERNMENT puts up money to fund something that the market otherwise would not have bothered with.
The market is ALREADY allocating capital to its most efficient use.
By definition, the government is intervening to redirect that capital (tax dollars) towards ends that the market would not otherwise. There is a built-in inefficieny.
Note: I am not saying this is right or wrong - just that it is.
If we completely eliminated NASA, maybe the market would colonize space faster than the government (over the long run) because instead of the tax dollars going to NASA, they would stay in the economy and be put towards more productive ends - perhaps funding unforseen technologies that would assist space travel... but that is hard to say.
This is the first post I've read from someone who gets it.
Mod this +1000
To all the "visual" editor developers...please don't make me use the mouse. It is the most inefficient input device when it comes to coding.
There is a commercial elisp package that seems to support decent semantic support for C++. http://www.xref-tech.com/xrefactory/main.html
But I haven't tried it. Has anyone tried this?
SG Cowen is a wall street investment bank. Their analysts get paid to do research on publicly traded companies.
They may have a position in google stock in which case they would try to find information to suggest that the stock should be priced higher. They may do investment banking business with google which introduces a conflict of interest. By law (regulation full-disclosure), the analyst needs to disclose such positions or investment banking relationships if they exist.
Their clients may have positions in google stock or be interested in buying or selling google. They pay the analysts for their reports. Accurate analysts develop good reputations and draw more business to the bank.
Why is everyone so worried about leakage?
_ lithography_promises_reduced_leakage_for_small_dev ices
You can dramatically reduce leakage by tweaking the process to give you a slightly slower process. It's not the end of the world folks. It's just at this point in time, it makes more sense to have the faster process and pay for it with leakage power. In the future that may or may not be true.
http://www.tgdaily.com/2005/09/20/new_intel_65_nm
With billions of dollars at stake - it is unwise to underestimate the ingenuity of device physicists.
Slower for TSMC or UMC. Intel has no problem with yields at 65nm. Intel is after all going to launch dual core 65nm NOTEBOOK chips early next year.
Intel is also laying down 3.5 billion dollars for a Fab in Israel which will be 45nm for 2007 production. Moore's law continues to have legs. It is ASTONISHING that this is something we take for granted! And in 2009, I predict 32nm. And in 2011, I predict 22nm. Guess where we'll be in 2013?
the area savings on area & increase in performance is no longer that much moving from one process node to another
Explain that to NVIDIA or ATI, where performance scales directly with the number of floating point units and hence is directly proportional to area.
The problem with designing for smaller geometries is that the ASIC tools aren't mature enough in time for the new process. Design rules increase geometrically as the process shrinks. Intel and AMD have in-house custom tools that deal with these problems since there is a very tight relationship with the Fab. TSMC and the like are more de-coupled from companies like Cadence or Synopsys.
To quote good old Jerry Sanders - "Real men have fabs"
I haven't seen major PC distributors selling clean boxes
If there were a market for clean PCs, don't you think someone would have built a company around that already?
There must not be enough profit to make this venture worthwhile or you'd see a plethora of companies trying to do this. Any costs that can be wringed out of a PC can be used to undercut competition and directly benefit the bottom line. I recently read a story about how Dell wants to get rid of all of the Intel and NVIDIA stickers so that can shave a few pennies off of the cost of the PC (and also reduce assembly times). If they're fighting to remove stickers, surely there is a huge incentive to cut out $35 of software expense.
The non-existence of such options seems to suggest that there is no money to be made there. People simply don't want the free alternative. They derive value from Windows and are willing to pay for that. The value is that the number of applications available on Windows vastly exceeds the number available on Linux. Couple this with the fact that a large number of popular linux applications also run on windows (cygwin, openoffice, firefox etc...) So for $35 (or whatever the OEM windows cost is) you get almost unrestriced choices in software applications.
It's extremely difficult for Linux to have the same number and quality of applications that Windows has. Linux has such a small market penetration, there is little financial incentive to develop software for it. The only reason to develop profitable software for Linux is if Linux can do something that Windows can't. There are many cases for which this happens, but the applications are mainly restricted to the server world. In the desktop space, nobody really has any answer to Microsoft.
For linux to succeed in the desktop, it has to provide a value above and beyond what Microsoft offers. I have to say - in my experience, it doesn't. Sure it can come close to COPYING microsoft, but where is the innovation that allows users to do things that Microsoft can't do? Keep in mind that we're talking about the OS here. If you develop a novel application for Linux, chances are it can be ported to windows (think Firefox). There simply isn't much value in kernel that the average user cares about.
The story is of course different for servers, which is why you see real corporations use Linux in the enterprise.
Look to Apple as an example of something that is successful in the desktop. They know how to innovate.
So this raises the question of why Linux can't innovate beyond copying windows? The answer, I believe, is that there isn't enough financial incentive to innovate with Linux. Linux innovation is dependent on a community which doesn't get paid. Do you go to work and code for free? I thought not. The best and brightest want to get paid.
The only hope is that hardware companies like Sun and IBM want to reduce the cost of OS software to 0 since it reduces the costs of their systems. They have the incentive to pay developers to innovate for Linux. They just, quite frankly, haven't been able to compete with Microsoft or Apple. Hopefully that will change in the future. I'd like to have cheaper computers just like everybody else.
That's what I meant when I said "They might lose a strategic bargaining position with Intel"
There might be a dozen or more reasons why Dell might not want to go with AMD. It's actually pretty easy to speculate on why this might be the case. Dell actually has the data to evaluate what's best for Dell (as opposed to some random slashdot person saying Dell is "dumb" for not choosing AMD) - and given the success of the company, I'd say they're doing a pretty fine job.
Yep. Michael Dell sure is fucking it up. I mean - he's only the 10th richest man in America and he's already 40. What a slacker.
If ONLY he used AMD. Then he could really make it.
I mean, they only made 3 billion dollars in pure profit last year.
They somehow gain marketshare every single year.
They should adopt a strategy like Gateway or IBM. Oh, wait, IBM is out of the PC business because they can't make any money. Gateway loses money every single year...and HP makes the majority of their money from printer INK.
Perhaps using AMD would increase costs MORE than it would save them. Since they'd have to now support Intel and AMD motherboards and chipsets, effectively doubling the number of configurations they need to support, configure and manufacture. They might lose a strategic bargaining position with Intel causing their Intel costs to rise. Maybe they have visibility into Intel's roadmap that makes them believe Intel will offer more competitive processors in the future. Maybe they are worried about AMD as a reliable supplier with their relatively small capacity. Whatever... I'm sure the bean counters are on top of it and it makes no financial sense to do it at this time.
Look. The PC business is notoriously cut-throat competitive and survives on extremely LOW margins. If Dell could gain a price advantage from using AMD, they would. They're even trying to negotiate getting rid of all the Windows and Intel and Nvidia stickers so they can shave a couple of pennies off of the cost.
As dominant as Dell is, they still only have an 18% market share. There are plenty of other vendors.
Nobody is forcing people to shop at Dell. If Dell is doing a disservice to their customers, those people would shop elsewhere. Clearly this is not happening as their market share increases every single year.
The reason is that many mobile devices don't have the horsepower to do H.264 yet.
Yes, it's higher quality and smaller, but the computational cost (and therefore battery/power) is quite a bit higher.
I really do believe it has no technical merit.
I am paid to design processors and have worked on SPARC, MIPS and x86 designs for a span of over 12 years.
I spend my days thinking about how to improve processors. That's all I do... all day long.
So please... enlighten me on how the Itanium architecture improves computing on any metric.
Any performance advantage that you see today is solely due to their having much larger die size and pin count budgets vs. other processors just to compensate for their having a crappy ISA. If you give the same budget to a comparable x86 or traditional RISC processor, their absolute performance and performance/watt would far exceed any Itanium.
Put a 9MB cache on an Opteron and see how well it does on SPECFP for example.
An Opteron beats the Itanium 2 handily on integer code with just 1MB of cache.
x86 processors have a fixed amount of decoder logic overhead vs. RISC. The decoders essentially dynamically translate the x86 instructions into more machine-friendly micro-ops which are very RISC-like.
As transistor budgets increase exponentially (thanks to Moore's law), that fixed overhead gets smaller and smaller each generation - to the point that it's insigificant (less than 5% today and getting smaller tomorrow). So in the early 90s you could make a case for more efficient computing with RISC vs. x86, but today it's just so negligable that you don't care. There are also numerous micro-architectural tricks to get around the limited registers and wacky addressing modes.
Couple this with the fact that 99% of all of the world's software is written for x86 and you have a very large inertia to overcome in order to change the ISA.
Why would any software vendor port their application to a new architecture if that architecture is brand new and nobody is using it initially? This is a very expensive and risky task. Let's say that the incentive is increased performance with a new ISA (highly unlikely given that the ISA doesn't matter anymore given the very large transitors budgets). But let's be generous and give it a 50% performance advantage (again - this is fantasy land). Do you spend the 8 months porting, debugging, testing Photoshop? Or do you just wait 8 months for a 50% faster x86 to come out and instead spend that time improving your product as opposed to keeping it the same on a different architecture?
You'd have to be crazy to take that tradeoff. And so, you see what we have today - x86 everywhere.
Many of the Alpha team went to work on the Opteron. Or did you think that it's resemblance to the Alpha was a coincidence?
:) Sure, some of the Alpha engineers went to AMD. But there is still a core design team in Mass. that currently works for Intel. You didn't think that every single engineer just uprooted their family to move cross country to the Bay Area did you?
:)
The fingerprints are there
Intel doesn't just have the Alpha team. They also have the ex-PA-RISC team from HP. I hear they're not too shabby either
Bottom line is this - Intel has high quality, experienced engineers in quantities that exceed the competition by a wide margin.
This doesn't always guarantee success because management can always screw it up, but it's a pretty significant advantage.
The server market is shrinking, and the laptop market is expanding
Your data is incorrect. The server market is growing very quickly as is the laptop market. Hell, even the desktop market is growing (but at a much much slower pace).
Itanium is all but dead... relegated to the supercomputer niche - and we all know what happens to supercomputer companies :)
Intel has spent billions on Itanium and seen an effective return of 0%. Investors won't tolerate this for much longer. AMD's x86-64, and Intel's subsequent introduction of EMT64 (same thing), have finally pushed this ill conceived idea into its well deserved death spiral.
It has no technical merit. But technical merit sometimes is a secondary matter in the business world. However, the economics don't make any sense - you can't introduce a new ISA into a mature software market and expect it to fly just because you're Intel.
It was a mistake - write it off and move on.
This should free Intel to deploy those valuable Itanium engineers (like the ex-Alpha team) to work on something that actually generates cash (like x86 servers). So while AMD might have a short term lead - the giant resources of Intel are more than enough to catch up and re-assert their leadership position.
EE Times puts out a yearly salary report which is a good gauge for engineers.
9 00112
. gif. gif
http://www.eet.com/showArticle.jhtml?articleID=30
The following charts are especially enlightening:
http://img.cmpnet.com/eet/news/04/august/SALARY_3
http://img.cmpnet.com/eet/news/04/august/SALARY_2
The good news is that if the particular algorithms you're interested in have a favorable ratio of computation to memory access (aka good data reuse), then you are loving the new parallel world and your code is getting faster at the rate that computational capability is increasing (which is greater than Moore's law.)
How can you scale faster than Moore's Law?
If your application is compute bound and parallel, then you increase capability by increasing functional units which are directly proportional to transistor density (aka Moore's Law).
When NVIDIA draws their pretty graphs showing faster than Moore's Law scaling, it's because their die size hasn't been maxed-out in the past and is increasing in size. When NVIDIA chips become reticle limited (upper limit die size), they will also scale with Moore's law.
I'v searched Intel's manuals a bit and there's no mention whether store buffers or renamed registers are used to perform STLF
i ssue01/art01_microarchitecture/p03_netburst.htm
:)
It's actually a store buffer in the P4 (and most other x86 processors for that matter). Trust me, I'm paid to know this.
http://www.intel.com/technology/itj/2004/volume08
Also, there aren't many restrictions on STLF, mostly size (a load of 128 bits can't be forwarded from a store of 32 bits, although the converse is possible) and alignment.
There are partial address match problems because the P4 doesn't use all of the bits.
In common situations and if the compiler generates proper code, STLF restrictions shouldn't be very restricting at all.
Somebody please explain this to Adobe
Yes, store forwarding does help alleviate latency, but the forwarded data doesn't come from the renamed registers, it comes from a store-buffer (in some cases it can be just as fast as a regfile). Also, because detecting store-load dependencies is hard (think about how to do it in hardware), there are many restrictions on the kinds and types of forwarding that can be done at full speed.
As you mentioned, it's very inefficient to have spill/fills instead of register accesses - you can acheive similar performance, but most likely lower performance-per-watt.
I'm nitpicking, your original post is excellent.
In particular, there has been a longstanding disconnect between the growth in the amount of memory bandwidth available to chips versus the amount of computation that can be done on them. Computational capacity is growing much more quickly than memory access
I think you're confusing memory latency and bandwidth. Memory latency isn't improving as fast as computation or bandwidth.
In order to hide latency, you do more things in parallel while you wait for your long latency memory operation to complete. Finding many things to do in parallel in a single thread is hard to do automatically by the hardware (ie - out of order execution). And by "hard to do", I mean power hungry. A more power efficient way to do this is to put the burden on the programmer to partition the problem into parallel tasks and explicitly map it onto multiple execution units.
I'd like to nitpick a misconception.
Register renaming does nothing to alleviate the scarcity of registers in x86. The programmer does not see the renamed registers and so must use only 8 registers (or 16 with x86-64) and still spill/fill to memory.
Register renaming is required for out-of-order execution and does not eliminate the spill/fills. You're mixing two different concepts.