Erm, who knows what specializations, investments and technological developments? We're talking about a time when you were lucky to halve most of your children live past 12. Figured an economist would construct a completely useless model forgetting any realistic variables and concluding the results are in any way reliable.
It was optimal for its time. Industrialization largely did away with the economic benefits of slavery, but for the technology available at the time of inception and popularity, slavery was very very effective at a time when manual labor was a premium and transportation was limited.
so as soon as you stop 18 year olds from believing in themselves and the promise of their lives to do incredible things, that's when you convince more people not to go to college. so who here wants the task of destroying millions of young people's faith in themselves?
Oh! Me! Me! Having been 18 not too long ago -- or perhaps too long -- I can tell you that 18 year olds are deluded, self-centered, narcissistic, unreasonably entitled and full of themselves; I certainly was.
Complaining about immigrants "taking jobs from Americans"? It couldn't be because immigrants are willing to flip burgers, clean toilets and basically work hard at non-glamorous jobs for low pay. There will always be more ditch diggers than scientists and telling every single kid that they're special without qualifying it effectively makes everyone not special.
How about instead of "you can grow up to be anything you want to be", we tell kids "as long as you work hard and do right by others, there's no shame in not being Joe McMansion"?
I think the point is you don't have to. One chromospore emits light that is received by another chromospore, which then transmits it to another. Each time, the wavelength of light changes.
Problem being, this isn't a 3-terminal gating effect. I'm having a difficult time imagining how they make logic gates from this.
His point was that Google has built an empire, not necessarily made money. One of the things about Google's business model is that any additional use of the internet is better for them -- whether or not they make money directly.
Yes, youtube is just a giant money sink. But think of how many more people use the internet on how many more devices (phones, tablets, netbooks, etc.) because of it. Now think of when all those people need to find something related to a video they saw.
The basic ad-for-money model only works if the market expands. Google helps that market expand by making services that bring more people online for more hours of the day.
The relative complexity of a C++ program vs what someone can realistically do in HDL is vastly different. Try coding Office in HDL and watch as you go Wayne Brady on your computer.
We're talking millions per year worth of money for powering their server farms. The custom blade design alone saved them a few million annually. We're not talking small money, especially as they keep expanding to meet the needs of the tubes.
Google has been spending a lot of effort -- from custom power structures inside their buildings to buying that magic box that generates power form minerals to custom-making their own server blades -- to reduce power and make energy efficient servers; they have so many of them after all.
These guys, while formerly PA Semi, focused their new business on energy-efficient server CPU's. So I wouldn't so much expect a gPad. It's likely the consumer will never see the chips that are being produced here.
That really depends. The instruction binary itself is more compact compared to say, ARM, but the actually implementation is far from.
For instance, almost every x86 processor these days pre-decode and store that data into cache. These are invisible to the programming model but it also means that you really aren't getting that many more instructions-per-bit of cache compared to say, ARM.
If you look at how x86 is implemented, and how most compilers use it, about 80% of the ISA isn't even used and the implementation is almost all about accelerating the RISC-like instructions (load, store, execute, permute) rather than the CISC-ish instructions (execute with a memory operand).
That's not to say it's that much worse than its ARM counterparts when it comes to area. But it is a lot worse when it comes to power.
Adding more cache isn't free. Ignoring power and area, indexing more cache also increases access times. This is why Intel has always opted to use small sizes for their L1 cache.
You can add more L3 on-die, but there are vast diminishing returns to that due to the high-latency access times of L3 cache.
Society has, is and will always be on a pyramidal level, at the bottom you have basic workers, in the middle you have skilled workers, at the top you have managers. We will always need more basic workers than skilled workers and more skilled workers than managers. Society has, is and will always fall into this order. However, the definitions of skilled workers, basic workers and managers have changed throughout the ages.
No one is arguing that. However, it is the degree of difference that matters. The capitalists in the Western world saw what happened in Russia and China; they saw what happened if you push people too far and make the standard of living for the bottom demographic too much worse than those at the top. They learned a quick lesson: while free trading leads to great wealth and prosperity, it also has the nasty side effect that the rich will get richer and the poor will get relatively more poor. It's natural; those that have power (money) are more able to get more. They also learned that there is a breaking point. If a select few gain too much of the overall production power, the lower demographics will either become discouraged from working (sense of no hope) or will outright revolt.
That's why we have public schools, roads, welfare, medicare, etc. Yes, there will always be the rich, the moderately comfortable, the poor, etc. But if left unchecked, this arrangement is self-destructive. The socialist programs counter this. But you have to be careful here too. Too strong of a social program and you begin to stifle the natural growth. Too little, and the distribution of wealth begins to become unhealthy to a productive society.
...Or they can and pass it on to customers. It is a lot more fair. For example, if Pizza Hut has 20 cars they use to deliver pizza, they pay for 20 car's worth of wear on the roads and raise their price of delivery so it evens out. Same with the rest. UPS and FedEx can simply add on a bit more to shipping. In the end it evens out and becomes more fair
That would work if taxes were paid on a per-use basis. Such that every road was a toll-road. But again, these things aren't meant to be "fair" in the ownership, free market sense. Public services are meant to provide everyone a basic standard of living such that the poor don't get *too* poor.
Then why doesn't all of civilisation collapse? It is because humans generally follow the law.
And who provides that law? The legal system, rights of ownership, deeds, policing, rules on what counts as trespassing, what counts as ownership, what is "fair sale" and "fair price", who provides that? Oh that's right, the people you pay taxes to. There is a finite amount of land and everyone wants more. If you want the system to protect your claims, you better the hell pay "protection money".
It's not that simple. The corporations go where they can sell; that's what they do. Currently, the U.S. the biggest spender in the world. Corporations aren't here out of some "oh, the tax isn't so bad, so let's stay in America" type of lethargy; they're here because they have to be in order to sell to U.S. customers. What has happened is that we've made it far too easy for overseas companies and "subsidiaries" to import. While this would be a great idea if we were trading with other countries with similar economies (Europe), when dealing with countries with horribly lax corporate regulations and incredibly low spending power (China), it doesn't work in our favor.
Yes, eliminating corporate income tax would attract some companies back but that's a band-aid solution; one in which further tips the balance of power between supplier and consumer in the direction of the supplier. IMO, the U.S. should never have signed on to so many free-trade agreements with countries that don't have anything to provide in terms of product. Free trade only works in your favor if you have something to sell to the other party as well as something to buy.
The same argument could be made for netbooks. The only advantage they have over a laptop is portability and longer battery life. If you think about it, laptops really aren't that "mobile". You can lug them around but you really need to plop yourself down somewhere to use it.
The form factor in this allows you to pull it out anywhere and look up something. Yes, you can do that on a phone but the screen size is very limiting for many things.
The difference is Dell's more powerful computers today does the exact same function as Dell's less powerful computers years ago. They still run word processing, spreadsheets, games, a task manager, etc.
The "superphone" is more an indication that it's taken on far more functionality, not just getting better at what it previously did.
My understanding is that even DX10+ compliant GPUs still suffer badly when conditional branching occurs. They can do it, but it basically causes them to throw away everything.
That's entirely up to the implementation. Today's generations of GPU's don't pay much heed to conditional branching but the upcoming Fermi from nVidia, for instance, does introduce branch prediction and tracking. The API supports conditionals and loops.
As for Larrabee, while it was designed as a GPU in some ways, I got the impression it still hewed to CPU roots. It was integer based, not floating point based
*boggle* no it wasn't. The thing was a bunch of 486 CPU's each with a gigantic 128-bit SIMD (read: vector floating point) unit attached. It obviously was not made to do anything but the most rudimentary CPU tasks. Hell, it doesn't even support branch prediction or OoOE.
They wanted to make all those college raytracer programs practical for use, replacing the current model which is somewhat more fuzzy and less accurate, but *way* faster.
Erm, no. While it's true that SSE supports 64-bit FP and may have been faster than the double-precision data on current graphics cards *per core*, in aggregate, it still wouldn't be any faster than a typical graphics card. And with Fermi, nVidia has vastly improved its double-precision processing anyway.
I don't know the specifics, but I suspect bengie is correct that the heat isn't the biggest issue. Both have thermal ranges that are close to eachother and have not that substantially different heat generation.
I didn't get that as his point but it wasn't a difference in heat generation, it was the sum. Your GPU currently requires anywhere from 1-2 fans and a large heatsink to dissipate the heat not to mention 2 PCI-E power rails. Your CPU requires a separate 2.5V power rail that can go up to ~130W.
Consider if you consolidated that into one ~1500 pin package. The number of pins you'd need just to feed ~300-400W would be enormous not to mention getting the heat away with a sufficient heatsink/fan.
one of a multitude of issues is: how can you get the instruction sets to sync up? One is a general purpose unit and the other is a specific instruction unit. We're also talking about say, 1600+ cores on one versus 16 to maybe 32 (or say 64-256 in the near future) cores on the other. Remember, this will likely be done on enterprise before it hits consumer level - rendering farms would love it.
The way AMD is planning to do it is by having the CPU front-end feed the GPU back-end. Meaning that all of the essential instruction/thread tracking would still be done by the CPU whereas the heavy math would be done by the GPU. Sound familiar? Open up your PS3 and you'll see one.
nVidia's current GPU's already track hundreds of threads among their cores just fine. The programming model is fairly elegant, breaking instructions up into blocks that can be executed in parallel.
Actually, Larrabee *was* designed to be a GPU. The approach was unique in that each individual core was x86 but the heart of each core was a powerful SIMD (SSE) unit. The idea was that you could program it like you did a multi-threaded x86 program (like you would with multi-core x86 CPU's) but it would run very fast like a GPU.
GPU's nowadays (ever since DirectX 8 I believe) support branching, conditionals, and all have local pockets of cache to deal with the DRAM they have to interface to. GPGPU compilers like CUDA and OpenCL all support this.
Really, nowadays, there isn't much of a difference. Think of modern GPU's as a lot (hundreds) of parallel, light CPU's. And think of multi-core CPU's as a couple of really really fast GPU's.
The biggest problem will be heat. GPU's currently consume and dissipate upwards of 200W. Likewise for CPU's. To get a single die or even package to consume and dissipate that much power and heat will be a challenge not just for the silicon designers but the system guys as well.
To some level, CPU's have been moving to be more GPU like for a long time. SIMD (SSE, AltiVec, NEON) are GPU features that made their way to CPU's. Ditto for parallel, long pipelines. Remember the Pentium 4? That was a huge step in the GPU direction.
There are two problems with that approach:
1. Code that isn't pure number-crunching doesn't run well on such a compute model. 2. The model is almost entirely memory-starved. GPU's have up to a GB of high-speed, dedicated RAM on the card itself. CPU's have to live with high-density (relatively) slot-loaded memory.
AMD is moving in a direction where the GPU compute parts are fed by the CPU front-end. As we move forward, I suspect we'll see more of a "fusion" if you will (don't sue me) of the two compute models.
And you also mean the porting of thousands and thousands of x86 apps as well? If the ARM version of Windows can't run the apps people want, they aren't going to by an ARM netbook.
I think Apple has pretty much proved that the actual number of applications people want to be able to run isn't anywhere near in the thousands and thousands. The vast majority of people basically want a browser, document editor, spreadsheet editor, presentation editor -- basically Microsoft Office and IE/Firefox/Chrome.
A media player, a media store, and various games. That's about it. In that list, the only category in which individual efforts must be made to port the program are games. But as we've seen with the iPhone, if you build the platform, they will come.
I think netbooks/tablets will be the point where traditional PC apps stop being desirable. With the form factor and the desired battery life, people will not want to run the same programs on a netbook as on a laptop/desktop. A 10" screen will be eaten up by Office ribbon and there's no way to play WoW on such a thing.
The question isn't whether Windows 7 will be ported to ARM. The question is whether Windows Phone 7 (which primarily runs on ARM) will soon be preferred over Windows 7 for everything up to laptops.
SMT is probably one of the best architectural features you can get in terms of complexity of implementation vs speedup. It isn't trivial by any means, but far less hazardous (get it?) than an OoO back-end or speculative execution.
Slashdot Mirror
Erm, who knows what specializations, investments and technological developments? We're talking about a time when you were lucky to halve most of your children live past 12. Figured an economist would construct a completely useless model forgetting any realistic variables and concluding the results are in any way reliable.
Most jobs may state minimal requirements "or equivalent experience" but for the most part, HR will not let resumes pass without qualifications.
It was optimal for its time. Industrialization largely did away with the economic benefits of slavery, but for the technology available at the time of inception and popularity, slavery was very very effective at a time when manual labor was a premium and transportation was limited.
so as soon as you stop 18 year olds from believing in themselves and the promise of their lives to do incredible things, that's when you convince more people not to go to college. so who here wants the task of destroying millions of young people's faith in themselves?
Oh! Me! Me! Having been 18 not too long ago -- or perhaps too long -- I can tell you that 18 year olds are deluded, self-centered, narcissistic, unreasonably entitled and full of themselves; I certainly was.
Complaining about immigrants "taking jobs from Americans"? It couldn't be because immigrants are willing to flip burgers, clean toilets and basically work hard at non-glamorous jobs for low pay. There will always be more ditch diggers than scientists and telling every single kid that they're special without qualifying it effectively makes everyone not special.
How about instead of "you can grow up to be anything you want to be", we tell kids "as long as you work hard and do right by others, there's no shame in not being Joe McMansion"?
I think the point is you don't have to. One chromospore emits light that is received by another chromospore, which then transmits it to another. Each time, the wavelength of light changes.
Problem being, this isn't a 3-terminal gating effect. I'm having a difficult time imagining how they make logic gates from this.
His point was that Google has built an empire, not necessarily made money. One of the things about Google's business model is that any additional use of the internet is better for them -- whether or not they make money directly.
Yes, youtube is just a giant money sink. But think of how many more people use the internet on how many more devices (phones, tablets, netbooks, etc.) because of it. Now think of when all those people need to find something related to a video they saw.
The basic ad-for-money model only works if the market expands. Google helps that market expand by making services that bring more people online for more hours of the day.
The relative complexity of a C++ program vs what someone can realistically do in HDL is vastly different. Try coding Office in HDL and watch as you go Wayne Brady on your computer.
We're talking millions per year worth of money for powering their server farms. The custom blade design alone saved them a few million annually. We're not talking small money, especially as they keep expanding to meet the needs of the tubes.
Google has been spending a lot of effort -- from custom power structures inside their buildings to buying that magic box that generates power form minerals to custom-making their own server blades -- to reduce power and make energy efficient servers; they have so many of them after all.
These guys, while formerly PA Semi, focused their new business on energy-efficient server CPU's. So I wouldn't so much expect a gPad. It's likely the consumer will never see the chips that are being produced here.
That really depends. The instruction binary itself is more compact compared to say, ARM, but the actually implementation is far from.
For instance, almost every x86 processor these days pre-decode and store that data into cache. These are invisible to the programming model but it also means that you really aren't getting that many more instructions-per-bit of cache compared to say, ARM.
If you look at how x86 is implemented, and how most compilers use it, about 80% of the ISA isn't even used and the implementation is almost all about accelerating the RISC-like instructions (load, store, execute, permute) rather than the CISC-ish instructions (execute with a memory operand).
That's not to say it's that much worse than its ARM counterparts when it comes to area. But it is a lot worse when it comes to power.
Adding more cache isn't free. Ignoring power and area, indexing more cache also increases access times. This is why Intel has always opted to use small sizes for their L1 cache.
You can add more L3 on-die, but there are vast diminishing returns to that due to the high-latency access times of L3 cache.
Not even close. Look at the die photo of Nehalem, for instance. About ~50% is visible cache and special purpose caches may make up ~10% more.
Why would you ruin this for me. WHY?!
Society has, is and will always be on a pyramidal level, at the bottom you have basic workers, in the middle you have skilled workers, at the top you have managers. We will always need more basic workers than skilled workers and more skilled workers than managers. Society has, is and will always fall into this order. However, the definitions of skilled workers, basic workers and managers have changed throughout the ages.
No one is arguing that. However, it is the degree of difference that matters. The capitalists in the Western world saw what happened in Russia and China; they saw what happened if you push people too far and make the standard of living for the bottom demographic too much worse than those at the top. They learned a quick lesson: while free trading leads to great wealth and prosperity, it also has the nasty side effect that the rich will get richer and the poor will get relatively more poor. It's natural; those that have power (money) are more able to get more. They also learned that there is a breaking point. If a select few gain too much of the overall production power, the lower demographics will either become discouraged from working (sense of no hope) or will outright revolt.
That's why we have public schools, roads, welfare, medicare, etc. Yes, there will always be the rich, the moderately comfortable, the poor, etc. But if left unchecked, this arrangement is self-destructive. The socialist programs counter this. But you have to be careful here too. Too strong of a social program and you begin to stifle the natural growth. Too little, and the distribution of wealth begins to become unhealthy to a productive society.
...Or they can and pass it on to customers. It is a lot more fair. For example, if Pizza Hut has 20 cars they use to deliver pizza, they pay for 20 car's worth of wear on the roads and raise their price of delivery so it evens out. Same with the rest. UPS and FedEx can simply add on a bit more to shipping. In the end it evens out and becomes more fair
That would work if taxes were paid on a per-use basis. Such that every road was a toll-road. But again, these things aren't meant to be "fair" in the ownership, free market sense. Public services are meant to provide everyone a basic standard of living such that the poor don't get *too* poor.
Then why doesn't all of civilisation collapse? It is because humans generally follow the law.
And who provides that law? The legal system, rights of ownership, deeds, policing, rules on what counts as trespassing, what counts as ownership, what is "fair sale" and "fair price", who provides that? Oh that's right, the people you pay taxes to. There is a finite amount of land and everyone wants more. If you want the system to protect your claims, you better the hell pay "protection money".
It's not that simple. The corporations go where they can sell; that's what they do. Currently, the U.S. the biggest spender in the world. Corporations aren't here out of some "oh, the tax isn't so bad, so let's stay in America" type of lethargy; they're here because they have to be in order to sell to U.S. customers. What has happened is that we've made it far too easy for overseas companies and "subsidiaries" to import. While this would be a great idea if we were trading with other countries with similar economies (Europe), when dealing with countries with horribly lax corporate regulations and incredibly low spending power (China), it doesn't work in our favor.
Yes, eliminating corporate income tax would attract some companies back but that's a band-aid solution; one in which further tips the balance of power between supplier and consumer in the direction of the supplier. IMO, the U.S. should never have signed on to so many free-trade agreements with countries that don't have anything to provide in terms of product. Free trade only works in your favor if you have something to sell to the other party as well as something to buy.
The same argument could be made for netbooks. The only advantage they have over a laptop is portability and longer battery life. If you think about it, laptops really aren't that "mobile". You can lug them around but you really need to plop yourself down somewhere to use it.
The form factor in this allows you to pull it out anywhere and look up something. Yes, you can do that on a phone but the screen size is very limiting for many things.
The difference is Dell's more powerful computers today does the exact same function as Dell's less powerful computers years ago. They still run word processing, spreadsheets, games, a task manager, etc.
The "superphone" is more an indication that it's taken on far more functionality, not just getting better at what it previously did.
My understanding is that even DX10+ compliant GPUs still suffer badly when conditional branching occurs. They can do it, but it basically causes them to throw away everything.
That's entirely up to the implementation. Today's generations of GPU's don't pay much heed to conditional branching but the upcoming Fermi from nVidia, for instance, does introduce branch prediction and tracking. The API supports conditionals and loops.
As for Larrabee, while it was designed as a GPU in some ways, I got the impression it still hewed to CPU roots. It was integer based, not floating point based
*boggle* no it wasn't. The thing was a bunch of 486 CPU's each with a gigantic 128-bit SIMD (read: vector floating point) unit attached. It obviously was not made to do anything but the most rudimentary CPU tasks. Hell, it doesn't even support branch prediction or OoOE.
They wanted to make all those college raytracer programs practical for use, replacing the current model which is somewhat more fuzzy and less accurate, but *way* faster.
Erm, no. While it's true that SSE supports 64-bit FP and may have been faster than the double-precision data on current graphics cards *per core*, in aggregate, it still wouldn't be any faster than a typical graphics card. And with Fermi, nVidia has vastly improved its double-precision processing anyway.
I don't know the specifics, but I suspect bengie is correct that the heat isn't the biggest issue. Both have thermal ranges that are close to eachother and have not that substantially different heat generation.
I didn't get that as his point but it wasn't a difference in heat generation, it was the sum. Your GPU currently requires anywhere from 1-2 fans and a large heatsink to dissipate the heat not to mention 2 PCI-E power rails. Your CPU requires a separate 2.5V power rail that can go up to ~130W.
Consider if you consolidated that into one ~1500 pin package. The number of pins you'd need just to feed ~300-400W would be enormous not to mention getting the heat away with a sufficient heatsink/fan.
one of a multitude of issues is: how can you get the instruction sets to sync up? One is a general purpose unit and the other is a specific instruction unit. We're also talking about say, 1600+ cores on one versus 16 to maybe 32 (or say 64-256 in the near future) cores on the other. Remember, this will likely be done on enterprise before it hits consumer level - rendering farms would love it.
The way AMD is planning to do it is by having the CPU front-end feed the GPU back-end. Meaning that all of the essential instruction/thread tracking would still be done by the CPU whereas the heavy math would be done by the GPU. Sound familiar? Open up your PS3 and you'll see one.
nVidia's current GPU's already track hundreds of threads among their cores just fine. The programming model is fairly elegant, breaking instructions up into blocks that can be executed in parallel.
Actually, Larrabee *was* designed to be a GPU. The approach was unique in that each individual core was x86 but the heart of each core was a powerful SIMD (SSE) unit. The idea was that you could program it like you did a multi-threaded x86 program (like you would with multi-core x86 CPU's) but it would run very fast like a GPU.
GPU's nowadays (ever since DirectX 8 I believe) support branching, conditionals, and all have local pockets of cache to deal with the DRAM they have to interface to. GPGPU compilers like CUDA and OpenCL all support this.
Really, nowadays, there isn't much of a difference. Think of modern GPU's as a lot (hundreds) of parallel, light CPU's. And think of multi-core CPU's as a couple of really really fast GPU's.
The biggest problem will be heat. GPU's currently consume and dissipate upwards of 200W. Likewise for CPU's. To get a single die or even package to consume and dissipate that much power and heat will be a challenge not just for the silicon designers but the system guys as well.
To some level, CPU's have been moving to be more GPU like for a long time. SIMD (SSE, AltiVec, NEON) are GPU features that made their way to CPU's. Ditto for parallel, long pipelines. Remember the Pentium 4? That was a huge step in the GPU direction.
There are two problems with that approach:
1. Code that isn't pure number-crunching doesn't run well on such a compute model.
2. The model is almost entirely memory-starved. GPU's have up to a GB of high-speed, dedicated RAM on the card itself. CPU's have to live with high-density (relatively) slot-loaded memory.
AMD is moving in a direction where the GPU compute parts are fed by the CPU front-end. As we move forward, I suspect we'll see more of a "fusion" if you will (don't sue me) of the two compute models.
And you also mean the porting of thousands and thousands of x86 apps as well? If the ARM version of Windows can't run the apps people want, they aren't going to by an ARM netbook.
I think Apple has pretty much proved that the actual number of applications people want to be able to run isn't anywhere near in the thousands and thousands. The vast majority of people basically want a browser, document editor, spreadsheet editor, presentation editor -- basically Microsoft Office and IE/Firefox/Chrome.
A media player, a media store, and various games. That's about it. In that list, the only category in which individual efforts must be made to port the program are games. But as we've seen with the iPhone, if you build the platform, they will come.
I think netbooks/tablets will be the point where traditional PC apps stop being desirable. With the form factor and the desired battery life, people will not want to run the same programs on a netbook as on a laptop/desktop. A 10" screen will be eaten up by Office ribbon and there's no way to play WoW on such a thing.
The question isn't whether Windows 7 will be ported to ARM. The question is whether Windows Phone 7 (which primarily runs on ARM) will soon be preferred over Windows 7 for everything up to laptops.
SMT is probably one of the best architectural features you can get in terms of complexity of implementation vs speedup. It isn't trivial by any means, but far less hazardous (get it?) than an OoO back-end or speculative execution.