Poincare found that after intensely studying a problem and then walking away and "forgetting" about it for a while led to new insights about the problem during the "forgetting" period.
He recommended that his students take breaks to go think about something else entirely.
For a more in-depth analysis of this idea I would recommend reading Jacques Hadamard's "The Psychology of Invention in the Mathematical Field"
The next time you suffer from a block and you just can't get anywhere... take a break from the problem:) The latent memories of the problem interact with your unrelated activity in the subconsious in novel and unexpected ways that may lead to better insights into the problem.
Users of EDA software care about performance. Time to market is EVERYTHING the highly competitive ASIC markets. Just about everybody is moving to x86 due to it's superior performance - The 64-bit x86 chips from AMD are only going to accelerate this move.
You're describing SOEMT (switch on event multithreading). When the process blocks on an "event" like a cache miss, switch to another context.
The machine you were thinking about was the TERA machine which switched on cache misses.
This is NOT hyperthreading. Hyperthreading is SMT. The switching and interleaving is much more fine-grained. Thread switch heuristics and resource sharing algorithms exist in many places in the pipeline.
Why is revenue used as a metric? Why is revenue per share important? Companies can issue arbitrarily many shares so this metric is useless. Employees are only as good as their value to the bottom line.
Profit should be the metric. The whole reason to go into business in the first place is to make a profit.
I am reminded of that Saturday Night Live skit where you had a bank that would make change for you. You could give them a dollar and they'd give you back 4 quarters. How do they make profit? Volume...
I could setup a business that had billions in revenue and 0 profit.
Let's compare profit:
Microsoft: 13.10B IBM: 12.93B
It's closer than you think - and Microsoft is still ahead. Microsoft is a cash machine. They accumulate money at a rate of about 1 billion a month. They can't re-invest it because what is a more profitable use than Microsoft's existing (and paid for) business?
They make so much money that they literally don't know what to do with it.
Perhaps this is why they are considered to be the largest technology company around.
It would be slower for most things and faster for some things.
With 64 bits you need to make the critical path of the processor wider (the ALUs, the register file etc...). The makes your circuits slower. You have the lost opportunity cost of frequency -- You could have made the processor run at a higher frequency with a 32 datapath - For those of you with a background in computer arithmetics, think about the carry-chain in an adder.
64 bit addressing means that your pointers now take up twice as much space as your old 32 bit pointers. This exerts more pressure on your memory system and makes your L1 and L2 caches appear smaller. It also makes your memory bandwidth appear narrower since you need to ship more data across the bus between the processor and DRAM.
The benefit of x86-64 is that it provides more registers. This means that there are fewer loads and stores which are used to spill and fill stack variables. This reduces the number of instructions required. It doesn't necessarily speed up the processor because stack variables would live in the store-forwarding buffer and would bypass the cache lookup, but you do have fewer instructions which means less pressure on your instruction cache and lower instruction-fetch bandwidth requirements.
There is a benefit for applications that naturally use 64-bit data types. This is a minor effect because many apps do not need the full 64 bit arithmetic.
The other big benefit is that you can now address a 64-bit virtual address space which is convenient for accessing large data structures.
In a commodity based business, the lowest cost producer wins.
A commodity is something that is NOT unique. It is easily replaceable and has almost no differentiation from one producer compared to another.
Given that all of these PC makers are basically selling the same machine (commodity), the one with the lowest cost structure will have the highest profit.
Dell wins hands down. They have literally only about 2 hours of inventory (meaning that a computer will live in a warehouse for about 2 hours on average). This is an amazing feat for a company of this size.
The size of Dell also works to it's advantage. It's called economies of scale - the larger you are - the cheaper it is for you to produce your goods.
With Dell being so large, they can act like Walmart and squeeze their suppliers (Intel, Taiwanese mobo makers, Korean/Chinese DRAM makers etc...) for sweetheart deals on components.
Let's say you have the same machine from Dell, HP and Gateway and Dell offers the lowest price, what do you think the majority of consumers will buy?
PCs are commodities - there is very little that can differentiate one producer from another. The minute one producer offers an advantage on their PC, 3 nanoseconds later all of their competitors have copied them and offer the same thing. The only way to differntiate youself is through price. Cheaper wins.
A few things you're forgetting: India and Europe both heavily subsidize health care and education...
Why not let foreign governments subsidize health care? If an American corporation does not want to pay for health care, let the burden fall on some agent that does. It's simply moving the cost from one agent that does not want to bear the risk to another that does.
You lament about the American deficit, but what about the deficits of countries that finance health care and other social programs? If you look at some real statistics, you'll see that the US debt/GDP ratio is among the lowest in the world (hint - go to economist.com).
First of all 32-bit chips can access 4 GB of memory, not 2 GB. And second, he's got it backwards. Apple's machines CAN do 64-bit calculations, but they can't do 64-bit addressing.
I think you are the one mistaken.
32 bit chips can address 4GB of memory. On typical operating systems, a portion of that is dedicated to the OS leaving the user with much less than 4GB (The upper 2GB for Windows and I believe 1 GB for Linux).
Apple's new machines (IBM PPC970 processor) can do both 64 bit arithmetic and can address 64-bits of virtual memory.
As others have mentioned, there's no great benefit in changing from 32-bit pointers to 64-bit. The optimizations for the G5 that you see in today's apps are from handling complex 64-bit calculations. In this regard Mac OS X is fully up to speed.
Except that your pointers double in size which puts more pressure on your data caches and memory bandwidth which causes you to lose performance. There is also opportunity cost. 64-bit datapaths in the processor are wider in the critical path of the chip and will slow down your top frequency.
The performance increases are slowing down because we're reaching the end of life for 130nm technology. We can't squeeze much more speed out of it. I don't see any 130nm IBM chips running faster than 2.2GHz...
You're quoting numbers from 130nm x86 processors. Make no mistake, Intel and AMD will have 90nm processor offerings that will have the same benefits that IBM is claiming, and dare I say - much higher performance and performance/watt. That 4GHz P4 you're quoting will be on a 90nm process but it certainly won't be 150W.
No one except for Intel is currently sampling/shipping 90nm products.
Thanks for the insight; I wasn't aware of the SNL limit. But is it specific to semiconductors, or does it also apply to optical/quantum computers?
It applies to the abstract model of a switch that is gating the flow of particles from either side of an energy barrier. This would apply to optical computing as well. In the end, we can't scale smaller because Boltzmann's constant does not scale.
Quantum computing operates on a different set of criteria, but you still need to maintain coherent quantum states which requires energy.
There appear to be deep relationships between information and energy. Do a google search on information physics.
No, you are calculating incorrectly. Debt as a percentage of GDP is smaller than the majority of countries. You are using accumulated debt to GDP, when you should use the debt service payment to GDP ratio.
Over $300 billion is "lost" every year in interest payments. That is $300 billion that we would have to spend elsewhere or get a tax cut.... Debt is an unnecessary drain on national resources which should be eliminated at the first opportunity
Unless that debt is used to finance growth that exceeds the debt repayment. What if the money loaned was used to contribute OVER $300b in economic growth? That's why we use debt-service/GDP ratio. If we grow the GDP (aka - economy) more than the debt service, we come out ahead.
Governments run defecits all of the time. Contrary to popular belief, the sky is not falling. Provided that markets for goods and capital remain open, flexible and unfettered, market forces will be enough to "defuse" the deficit "incrementally". Meaning, if the US tries to accrue too much debt, interest rates will rise (implying a more risky investment) and foreign investors will not be as likely to provide capital. It's a self regulating system.
Debt has been used to finance things like highways and schools. Things like infrastructure, scientific research (a category under which a mission to the Moon would fit) and education are INVESTMENTS, which use debt in order to finance future growth. This is in fact how we got out of the great depression -- huge public works projects put people to work and then war bonds appeared to finance the war.
We should not use defecit financing when doing so would cause "too much" growth which would lead to inflation. When the economy is weak, the government spends more to stimulate growth - it's called fiscal stimulus.
Interest rates are low. Foreigners are more than happy to finance the US because they know it's a safe haven. Since when is the US going to default? It's not like we're Russia or Mexico or Brazil.
Our credit is good, so why shouldn't we borrow? Why is debt automatically considered a bad thing?
Running defecits can be good sometimes. Government spending can boost domestic economic activity and create wealth. As long as debt isn't a large percentage of GDP (it isn't) we're fine. Using debt to finance growth benefits everybody.
Did you pay for your house in full or did you go into debt? With interest rates so low (a reflection of foreigners willing to finance US debt), the government should take advantage.
I made a few typos in the previous post and I just wanted to make things a little clearer.
The paper uses an abstract model of a switching device (2 wells separated by a potential barrier between them). This generality of the model allows them to consider just about any kind of actual solid state switching device.
In order to maintain a coherent switch (that is to say that we can tell the state "on" from the state "off" after accounting for Heisenberg Uncertainty), there is a minimum energy required for the barrier height.
For width > 5nm, barrier_height = SNL energy For width 5nm, barrier_height increases as (1/w)^2 where w is the barrier width.
I am assuming T = 300K, but the analysis is still valid for cryogenic temperatures of near T = 0.
In the past, transistor scaling allowed us to benefit from both increased device density (transistors/mm^2) and increased switching speed (MHz).
The point of the paper was that this double benefit will disappear when geometries start getting smaller than a certain value. After this point, you will need to trade off device density vs. switching speed.
There are fundamental thermodynamic limits to how much thermal energy can be removed from a solid per unit time.
These are hard physical limits. You can think of the SNL energy as the minimum energy required to flip a bit which is around 0.017eV at 300K. How many bits do you want to process per second? How much power can you deliver to acheive this without burning out your device in the process?
The reason that this calculation is material independent is that there is an additional constraint of power-density on the shape of the energy barrier.
As you mentioned, the tunneling probability is a function of width, barrier height and effective mass of the tunneling particle. We are trying to construct a switch where we can control the flow of the particle from one side of the barrier to the other. In the "on" state, there is no energy barrier, the electron can move freely, and in the "off" state, the barrier is erected. We need to control the tunneling probability such that we can distinguish on from off.
Consider that the Shannon-Von-Neumann-Landauer (SNL) limit for the smallest energy required to process a bit is k_b * T * ln(2) ~= 0.017eV where k_b is the Boltzmann constant and T is temperature. For width > 5nm, this holds as a good approximation for the minimum height of the barrier to maintain a coherent switch. For a 5nm the energy increases as (1/w)^2 where w is the barrier width.
This is a LOT of power when summed over the entire chip area.
They invoke power density arugments that say that it is impractical to have 5-10 MEGAWatt! / cm^2 power density. The rate at which this thermal energy can be removed from a solid is limited -- and THIS is the reason why we can't scale smaller.
Fundamentally, we are power limited.
I am not a physicist, but I do design microprocessors for a living and I did study semiconductor physics in school.
I don't think I mentioned anything about stealing in my post.
The subject was on the tendency for people to skip commercials using technology. If you call that stealing then you and I disagree on the defintion of the word.
In lieu of commercials, I was offering suggestions of alternative streams of revenue for content-creators. Clearly, there is a demand for content and people are willing to pay - just look at dvd-sales, pay-channel subscriptions etc...
I am merely suggesting that advertisers should't force the consumer to watch commercials that they clearly don't want to watch. Find some other innovative way to get those revenues. I'm not paid to find solutions for Hollywood, so I'm not going to tell you about any solutions in my short post here, but I expect some people are working on this, and now they have a large incentive to do so.
Forcing consumers to watch commercials that they don't have any interest in watching does not add value to the content - it subtracts value.
Innovation the the heart of industry. If you don't innovate, your competitor will and you'll lose market share. This of course, does not work in the case of monopolies or trade cartels of which some may argue Hollywood fits into.
...but then again, if the only money was in the DVD release, why do TV at all
Nobody is going to shell out hard earned dollars on DVD box sets of content they have never seen or know nothing about. You could think of the television shows as advertisements for the DVDs. Perhaps this will cause the quality of shows to improve because if the show sucks, nobody is going to buy the dvd. This is a pretty strong incentive.
Or perhaps this will lead to the pay-per-view system dominating the ratings. This has worked for HBO quite well.
The Tivo/DVR watchers are skipping the commercials because for the most part they are annoying. This should be seen as a strong feedback signal to the advertisers that their methods do more to annoy than to inform.
Perhaps Hollywood isn't entitled to the gravy train that has been going on for the past 40 years or so and they might have to *gasp* INNOVATE, like everybody else to maintain a healthy profitable business.
Ummm...There are plenty of Wi-Fi alternatives. Intel by far does not have a monopoly on the Wi-Fi market. They only have a monopoly on the name Centrino.
Blockbuster wants to drive the sale price of DVDs down to $5 and abandon the rental business altogether. The publishers are reluctant.
Where did you hear that?
Considering that Blockbuster makes about 40% of it's revenue on late fees, I don't think that they want the consumer to hold onto a permanent copy for less than they can get for their late fees.
If Netflix could improve their distribution system in such a way that you could get your movies faster...Blockbuster is toast.
By the time the HD-DVD standard is widespread, hard drive capacities will be able to completely backup your "old" dvd collection.
Don't you just love exponential density improvements?
Video content is not expanding at an exponential rate. Therefore, at some point in time in the near future, your hard drive will be able to contain just about every movie around.
They already have dvd players with enough memory buffering to make the layer change seemless. I believe the buffer sizes are around 2 Megabytes. This is found in high end dvd players.
It's a well known fact that outsourcing of call centers is occuring because it is cheaper to hire the overseas labor. This saves the company money and reduces prices here in America.
You saying you are personally willing to pay more for a product if it is supported by better trained technical support services. You may be willing to do that, but the majority of people aren't. The majority make the tradeoff of cheaper computers vs. lower quality call service support.
This quality tradeoff occurs all of the time in the PC business. In the commodity PC business the lowest cost producer wins the most market share. It has always been about price and Dell is clearly the lowest cost producer in the PC business. That's why PC's are butt ugly, keyboards are made flimsy plastic, notebook computers from Dell have cheap keyboards and low-grade materials that cause the casing to bend and break easily. Everything in a PC is designed to be as cheap as possible while still being functional. Compare that to the more expensive Macs which are clearly higher quality and tell me what product has the highest market share?
Sure, there are niche players that make high end, high quality PCs, but they cater to a very tiny minority.
Actually, the original intent of Transmeta was to produce a chip that had performance on-par or better than comparable Intel/AMD offerings.
When the project failed to do that (quite badly), then the marketeers refocussed the company message to start talking about 'low-power' and efficiency. This deflects the critics who do not understand computer architecture and things like power-efficiency.
Yeah, it's a neat research project and having Linus work there didn't hurt PR at all, but the performance just isn't there. The truth is that the architecture sucks rocks. Any statically scheduled CPU is going to have major problems competing with dynamically scheduled x86 processors. You can have low power and efficiency with the same performance as Transmeta using standard Intel/AMD CPUs and just scaling down the voltage (and hence frequency). In fact, Intel and AMD have ULV (ultra-low-voltage) CPUs that do just that.
This is a metric that could be used. How much performance can you get per watt?
For supercomputing, I would imagine that something like SPECfp/watt or SPECrate/watt would be a decent metric.
If your limitation is a finite power budget, then you pick the most highest perf/watt CPUs.
P4 3.2 EE = 18.44 SPECfp/Watt (80 watts) Crusoe = ?? No performance numbers published, but I'll bet you it's lower
Building larger caches (which can be made low power) is a good way to acheive high power/performance efficiency. Crippling your performance with a VLIW seems like a bad choice. You can voltage scale down a Pentium-M to the power levels of a Crusoe and easily get 2-3x the performance.
When you have 2 completely different instruction sets, you CANNOT remove the compiler as a variable.
I'm sorry, it's just not possible in the same sense of the statement (2+2 = 5) is not possible.
You have the x86 gcc backend and the PPC gcc backend code. They are completely different beasts. Many key optimizations for performance happen in the back-end.
So what is "fair" in this respect? I think that SPEC has some very reasonable rules with regards to comparing computers of different instruction sets. You are allowed to use the best available compiler on the market to compile completely visible source code. You can see if the compiler was tuned *just* for the benchmark by examining the generated code. In fact many people have called shenanigans on Sun for some of their floating point optimizations which are IMHO quite suspect. All of this is in the open for critics to see and evaluate.
If the only compiler on the market is one that produces crappy code for your system, then that's a fault of choosing a system that only has inferior compilers to choose from. It's all part of the choice you make in investing in one system or another.
You may disagree, but what would you suggest as a better test? I'd love to see more compute-intensive application source code benchmarked outside of SPEC.
Slashdot Mirror
Poincare found that after intensely studying a problem and then walking away and "forgetting" about it for a while led to new insights about the problem during the "forgetting" period.
... take a break from the problem :) The latent memories of the problem interact with your unrelated activity in the subconsious in novel and unexpected ways that may lead to better insights into the problem.
He recommended that his students take breaks to go think about something else entirely.
For a more in-depth analysis of this idea I would recommend reading Jacques Hadamard's "The Psychology of Invention in the
Mathematical Field"
The next time you suffer from a block and you just can't get anywhere
Users of EDA software care about performance. Time to market is EVERYTHING the highly competitive ASIC markets. Just about everybody is moving to x86 due to it's superior performance - The 64-bit x86 chips from AMD are only going to accelerate this move.
You're describing SOEMT (switch on event multithreading). When the process blocks on an "event" like a cache miss, switch to another context.
The machine you were thinking about was the TERA machine which switched on cache misses.
This is NOT hyperthreading. Hyperthreading is SMT. The switching and interleaving is much more fine-grained. Thread switch heuristics and resource sharing algorithms exist in many places in the pipeline.
Why is revenue used as a metric? Why is revenue per share important? Companies can issue arbitrarily many shares so this metric is useless. Employees are only as good as their value to the bottom line.
Profit should be the metric. The whole reason to go into business in the first place is to make a profit.
I am reminded of that Saturday Night Live skit where you had a bank that would make change for you. You could give them a dollar and they'd give you back 4 quarters. How do they make profit? Volume...
I could setup a business that had billions in revenue and 0 profit.
Let's compare profit:
Microsoft: 13.10B
IBM: 12.93B
It's closer than you think - and Microsoft is still ahead. Microsoft is a cash machine. They accumulate money at a rate of about 1 billion a month. They can't re-invest it because what is a more profitable use than Microsoft's existing (and paid for) business?
They make so much money that they literally don't know what to do with it.
Perhaps this is why they are considered to be the largest technology company around.
It would be slower for most things and faster for some things.
With 64 bits you need to make the critical path of the processor wider (the ALUs, the register file etc...). The makes your circuits slower. You have the lost opportunity cost of frequency -- You could have made the processor run at a higher frequency with a 32 datapath - For those of you with a background in computer arithmetics, think about the carry-chain in an adder.
64 bit addressing means that your pointers now take up twice as much space as your old 32 bit pointers. This exerts more pressure on your memory system and makes your L1 and L2 caches appear smaller. It also makes your memory bandwidth appear narrower since you need to ship more data across the bus between the processor and DRAM.
The benefit of x86-64 is that it provides more registers. This means that there are fewer loads and stores which are used to spill and fill stack variables. This reduces the number of instructions required. It doesn't necessarily speed up the processor because stack variables would live in the store-forwarding buffer and would bypass the cache lookup, but you do have fewer instructions which means less pressure on your instruction cache and lower instruction-fetch bandwidth requirements.
There is a benefit for applications that naturally use 64-bit data types. This is a minor effect because many apps do not need the full 64 bit arithmetic.
The other big benefit is that you can now address a 64-bit virtual address space which is convenient for accessing large data structures.
In a commodity based business, the lowest cost producer wins.
A commodity is something that is NOT unique. It is easily replaceable and has almost no differentiation from one producer compared to another.
Given that all of these PC makers are basically selling the same machine (commodity), the one with the lowest cost structure will have the highest profit.
Dell wins hands down. They have literally only about 2 hours of inventory (meaning that a computer will live in a warehouse for about 2 hours on average). This is an amazing feat for a company of this size.
The size of Dell also works to it's advantage. It's called economies of scale - the larger you are - the cheaper it is for you to produce your goods.
With Dell being so large, they can act like Walmart and squeeze their suppliers (Intel, Taiwanese mobo makers, Korean/Chinese DRAM makers etc...) for sweetheart deals on components.
Let's say you have the same machine from Dell, HP and Gateway and Dell offers the lowest price, what do you think the majority of consumers will buy?
PCs are commodities - there is very little that can differentiate one producer from another. The minute one producer offers an advantage on their PC, 3 nanoseconds later all of their competitors have copied them and offer the same thing. The only way to differntiate youself is through price. Cheaper wins.
A few things you're forgetting: India and Europe both heavily subsidize health care and education...
Why not let foreign governments subsidize health care? If an American corporation does not want to pay for health care, let the burden fall on some agent that does. It's simply moving the cost from one agent that does not want to bear the risk to another that does.
You lament about the American deficit, but what about the deficits of countries that finance health care and other social programs? If you look at some real statistics, you'll see that the US debt/GDP ratio is among the lowest in the world (hint - go to economist.com).
First of all 32-bit chips can access 4 GB of memory, not 2 GB. And second, he's got it backwards. Apple's machines CAN do 64-bit calculations, but they can't do 64-bit addressing.
I think you are the one mistaken.
32 bit chips can address 4GB of memory. On typical operating systems, a portion of that is dedicated to the OS leaving the user with much less than 4GB (The upper 2GB for Windows and I believe 1 GB for Linux).
Apple's new machines (IBM PPC970 processor) can do both 64 bit arithmetic and can address 64-bits of virtual memory.
As others have mentioned, there's no great benefit in changing from 32-bit pointers to 64-bit. The optimizations for the G5 that you see in today's apps are from handling complex 64-bit calculations. In this regard Mac OS X is fully up to speed.
Except that your pointers double in size which puts more pressure on your data caches and memory bandwidth which causes you to lose performance. There is also opportunity cost. 64-bit datapaths in the processor are wider in the critical path of the chip and will slow down your top frequency.
The performance increases are slowing down because we're reaching the end of life for 130nm technology. We can't squeeze much more speed out of it. I don't see any 130nm IBM chips running faster than 2.2GHz...
You're quoting numbers from 130nm x86 processors. Make no mistake, Intel and AMD will have 90nm processor offerings that will have the same benefits that IBM is claiming, and dare I say - much higher performance and performance/watt. That 4GHz P4 you're quoting will be on a 90nm process but it certainly won't be 150W.
No one except for Intel is currently sampling/shipping 90nm products.
Thanks for the insight; I wasn't aware of the SNL limit. But is it specific to semiconductors, or does it also apply to optical/quantum computers?
It applies to the abstract model of a switch that is gating the flow of particles from either side of an energy barrier. This would apply to optical computing as well. In the end, we can't scale smaller because Boltzmann's constant does not scale.
Quantum computing operates on a different set of criteria, but you still need to maintain coherent quantum states which requires energy.
There appear to be deep relationships between information and energy. Do a google search on information physics.
So 70% isn't a large percentage?
... Debt is an unnecessary drain on national resources which should be eliminated at the first opportunity
No, you are calculating incorrectly. Debt as a percentage of GDP is smaller than the majority of countries. You are using accumulated debt to GDP, when you should use the debt service payment to GDP ratio.
Over $300 billion is "lost" every year in interest payments. That is $300 billion that we would have to spend elsewhere or get a tax cut.
Unless that debt is used to finance growth that exceeds the debt repayment. What if the money loaned was used to contribute OVER $300b in economic growth? That's why we use debt-service/GDP ratio. If we grow the GDP (aka - economy) more than the debt service, we come out ahead.
Governments run defecits all of the time. Contrary to popular belief, the sky is not falling. Provided that markets for goods and capital remain open, flexible and unfettered, market forces will be enough to "defuse" the deficit "incrementally". Meaning, if the US tries to accrue too much debt, interest rates will rise (implying a more risky investment) and foreign investors will not be as likely to provide capital. It's a self regulating system.
Debt has been used to finance things like highways and schools. Things like infrastructure, scientific research (a category under which a mission to the Moon would fit) and education are INVESTMENTS, which use debt in order to finance future growth. This is in fact how we got out of the great depression -- huge public works projects put people to work and then war bonds appeared to finance the war.
We should not use defecit financing when doing so would cause "too much" growth which would lead to inflation. When the economy is weak, the government spends more to stimulate growth - it's called fiscal stimulus.
Interest rates are low. Foreigners are more than happy to finance the US because they know it's a safe haven. Since when is the US going to default? It's not like we're Russia or Mexico or Brazil.
Our credit is good, so why shouldn't we borrow? Why is debt automatically considered a bad thing?
Running defecits can be good sometimes. Government spending can boost domestic economic activity and create wealth. As long as debt isn't a large percentage of GDP (it isn't) we're fine. Using debt to finance growth benefits everybody.
Did you pay for your house in full or did you go into debt? With interest rates so low (a reflection of foreigners willing to finance US debt), the government should take advantage.
I made a few typos in the previous post and I just wanted to make things a little clearer.
The paper uses an abstract model of a switching device (2 wells separated by a potential barrier between them). This generality of the model allows them to consider just about any kind of actual solid state switching device.
In order to maintain a coherent switch (that is to say that we can tell the state "on" from the state "off" after accounting for Heisenberg Uncertainty), there is a minimum energy required for the barrier height.
For width > 5nm, barrier_height = SNL energy
For width 5nm, barrier_height increases as (1/w)^2 where w is the barrier width.
I am assuming T = 300K, but the analysis is still valid for cryogenic temperatures of near T = 0.
In the past, transistor scaling allowed us to benefit from both increased device density (transistors/mm^2) and increased switching speed (MHz).
The point of the paper was that this double benefit will disappear when geometries start getting smaller than a certain value. After this point, you will need to trade off device density vs. switching speed.
There are fundamental thermodynamic limits to how much thermal energy can be removed from a solid per unit time.
These are hard physical limits. You can think of the SNL energy as the minimum energy required to flip a bit which is around 0.017eV at 300K. How many bits do you want to process per second? How much power can you deliver to acheive this without burning out your device in the process?
The reason that this calculation is material independent is that there is an additional constraint of power-density on the shape of the energy barrier.
As you mentioned, the tunneling probability is a function of width, barrier height and effective mass of the tunneling particle. We are trying to construct a switch where we can control the flow of the particle from one side of the barrier to the other. In the "on" state, there is no energy barrier, the electron can move freely, and in the "off" state, the barrier is erected. We need to control the tunneling probability such that we can distinguish on from off.
Consider that the Shannon-Von-Neumann-Landauer (SNL) limit for the smallest energy required to process a bit is k_b * T * ln(2) ~= 0.017eV where k_b is the Boltzmann constant and T is temperature. For width > 5nm, this holds as a good approximation for the minimum height of the barrier to maintain a coherent switch. For a 5nm the energy increases as (1/w)^2 where w is the barrier width.
This is a LOT of power when summed over the entire chip area.
They invoke power density arugments that say that it is impractical to have 5-10 MEGAWatt! / cm^2 power density. The rate at which this thermal energy can be removed from a solid is limited -- and THIS is the reason why we can't scale smaller.
Fundamentally, we are power limited.
I am not a physicist, but I do design microprocessors for a living and I did study semiconductor physics in school.
I don't think I mentioned anything about stealing in my post.
The subject was on the tendency for people to skip commercials using technology. If you call that stealing then you and I disagree on the defintion of the word.
In lieu of commercials, I was offering suggestions of alternative streams of revenue for content-creators. Clearly, there is a demand for content and people are willing to pay - just look at dvd-sales, pay-channel subscriptions etc...
I am merely suggesting that advertisers should't force the consumer to watch commercials that they clearly don't want to watch. Find some other innovative way to get those revenues. I'm not paid to find solutions for Hollywood, so I'm not going to tell you about any solutions in my short post here, but I expect some people are working on this, and now they have a large incentive to do so.
Forcing consumers to watch commercials that they don't have any interest in watching does not add value to the content - it subtracts value.
Innovation the the heart of industry. If you don't innovate, your competitor will and you'll lose market share. This of course, does not work in the case of monopolies or trade cartels of which some may argue Hollywood fits into.
...but then again, if the only money was in the DVD release, why do TV at all
Nobody is going to shell out hard earned dollars on DVD box sets of content they have never seen or know nothing about. You could think of the television shows as advertisements for the DVDs. Perhaps this will cause the quality of shows to improve because if the show sucks, nobody is going to buy the dvd. This is a pretty strong incentive.
Or perhaps this will lead to the pay-per-view system dominating the ratings. This has worked for HBO quite well.
The Tivo/DVR watchers are skipping the commercials because for the most part they are annoying. This should be seen as a strong feedback signal to the advertisers that their methods do more to annoy than to inform.
Perhaps Hollywood isn't entitled to the gravy train that has been going on for the past 40 years or so and they might have to *gasp* INNOVATE, like everybody else to maintain a healthy profitable business.
Ummm...There are plenty of Wi-Fi alternatives. Intel by far does not have a monopoly on the Wi-Fi market. They only have a monopoly on the name Centrino.
Blockbuster wants to drive the sale price of DVDs down to $5 and abandon the rental business altogether. The publishers are reluctant.
Where did you hear that?
Considering that Blockbuster makes about 40% of it's revenue on late fees, I don't think that they want the consumer to hold onto a permanent copy for less than they can get for their late fees.
If Netflix could improve their distribution system in such a way that you could get your movies faster...Blockbuster is toast.
By the time the HD-DVD standard is widespread, hard drive capacities will be able to completely backup your "old" dvd collection.
Don't you just love exponential density improvements?
Video content is not expanding at an exponential rate. Therefore, at some point in time in the near future, your hard drive will be able to contain just about every movie around.
They already have dvd players with enough memory buffering to make the layer change seemless. I believe the buffer sizes are around 2 Megabytes. This is found in high end dvd players.
It's a well known fact that outsourcing of call centers is occuring because it is cheaper to hire the overseas labor. This saves the company money and reduces prices here in America.
You saying you are personally willing to pay more for a product if it is supported by better trained technical support services. You may be willing to do that, but the majority of people aren't. The majority make the tradeoff of cheaper computers vs. lower quality call service support.
This quality tradeoff occurs all of the time in the PC business. In the commodity PC business the lowest cost producer wins the most market share. It has always been about price and Dell is clearly the lowest cost producer in the PC business. That's why PC's are butt ugly, keyboards are made flimsy plastic, notebook computers from Dell have cheap keyboards and low-grade materials that cause the casing to bend and break easily. Everything in a PC is designed to be as cheap as possible while still being functional. Compare that to the more expensive Macs which are clearly higher quality and tell me what product has the highest market share?
Sure, there are niche players that make high end, high quality PCs, but they cater to a very tiny minority.
Why is the parent modded "funny".
Obsoleting their own product is EXACTLY what Intel's business plan is. Why else would someone buy a new chip if their current chip is good enough?
Actually, the original intent of Transmeta was to produce a chip that had performance on-par or better than comparable Intel/AMD offerings.
When the project failed to do that (quite badly), then the marketeers refocussed the company message to start talking about 'low-power' and efficiency. This deflects the critics who do not understand computer architecture and things like power-efficiency.
Yeah, it's a neat research project and having Linus work there didn't hurt PR at all, but the performance just isn't there. The truth is that the architecture sucks rocks. Any statically scheduled CPU is going to have major problems competing with dynamically scheduled x86 processors. You can have low power and efficiency with the same performance as Transmeta using standard Intel/AMD CPUs and just scaling down the voltage (and hence frequency). In fact, Intel and AMD have ULV (ultra-low-voltage) CPUs that do just that.
This is a metric that could be used. How much performance can you get per watt?
For supercomputing, I would imagine that something like SPECfp/watt or SPECrate/watt would be a decent metric.
If your limitation is a finite power budget, then you pick the most highest perf/watt CPUs.
P4 3.2 EE = 18.44 SPECfp/Watt (80 watts)
Crusoe = ?? No performance numbers published, but I'll bet you it's lower
Building larger caches (which can be made low power) is a good way to acheive high power/performance efficiency. Crippling your performance with a VLIW seems like a bad choice. You can voltage scale down a Pentium-M to the power levels of a Crusoe and easily get 2-3x the performance.
When you have 2 completely different instruction sets, you CANNOT remove the compiler as a variable.
I'm sorry, it's just not possible in the same sense of the statement (2+2 = 5) is not possible.
You have the x86 gcc backend and the PPC gcc backend code. They are completely different beasts. Many key optimizations for performance happen in the back-end.
So what is "fair" in this respect? I think that SPEC has some very reasonable rules with regards to comparing computers of different instruction sets. You are allowed to use the best available compiler on the market to compile completely visible source code. You can see if the compiler was tuned *just* for the benchmark by examining the generated code. In fact many people have called shenanigans on Sun for some of their floating point optimizations which are IMHO quite suspect. All of this is in the open for critics to see and evaluate.
If the only compiler on the market is one that produces crappy code for your system, then that's a fault of choosing a system that only has inferior compilers to choose from. It's all part of the choice you make in investing in one system or another.
You may disagree, but what would you suggest as a better test? I'd love to see more compute-intensive application source code benchmarked outside of SPEC.