Moore's law doesn't talk about performance...but about transistor/mm2 (as far as I know).Quote:
Originally Posted by HKPolice
Printable View
Moore's law doesn't talk about performance...but about transistor/mm2 (as far as I know).Quote:
Originally Posted by HKPolice
moore's law says nothing about performance.Quote:
Originally Posted by HKPolice
the problem isnt the fundamentals. there is plenty of room for performance improvements but the industry if focused on cmos so much that other ideas are usually not worth investing in until every bit of conventional cmos transistor performance is milked out.Quote:
Ever since the Core2Duo architecture, all processors are incremental upgrades with diminishing returns. Frankly I'm disappointed at Intel but when going up against the laws of physics, you're always going to lose, especially when AMD has its head up its own ass. :(
frankly, if you blindly trust the laws of physics you have already lost.
As far as you know is correct, Moore's Law is not actually a law, it is just a silly observation. Gordon Moore plotted the transistor density (or costs) vs a time scale and the slope derived Moore's Law. Nothing fundamentally drives Moore's Law in a physics related concept, it is purely economical.Quote:
Originally Posted by prava
The economics of going smaller drives companies to go smaller, and his observation was it takes roughly 2 years for an iteration of doubling the transistor density. Moore never once wrote, said, or pulbished any statement that device performance will double every two years, just that the complexity (or number of transistors) will increase as a consequence of utilizing the transistor density to drive better devices.
People have forgotten that Fugger broke 5ghz on air using a 45nm i7, with 32nm he could hit the cap... actually that would be epic, "First man to reach the SandyBridge Cap on Air".Quote:
Originally Posted by rado992
No that's a challenge if ever I saw one.
Moore's Law <<<<<<<<< Murphy's Law, lolQuote:
Originally Posted by JumpingJack
the only way i read it was total transistors, not even per mm2.
i remember nvidia preaching about gpu advancements across time being faster than cpus, my response was gpus now consume about 2-10x the power than they did 10-15. while cpus have had a much slower advancement in tdp, since no consumer buys a 4P gaming rig, but can easily buy 4x 200W gpus.
just a silly observation? then i guess the same could be said for any law.Quote:
Originally Posted by JumpingJack
and no you can not decouple physics from moore's law. physics is fundamental to understanding the natural world. MOS transistors would be a part of that set.
Moore's law has nothing to do with the physics problem, and Moore's law is not a scientific law. Physical laws will be the boundary and cause for the end of Moore's Law, it is the economics that drives mores law not physics. A scientific law withstands the test of time and is enduring and is testable. Moore's Law is not enduring, it will end. It has to.Quote:
Originally Posted by Chumbucket843
Any natural law, be it the law of conservation of momentum or energy (Thermodynamics) achieve that status through testable anti-hypothesis and observations and, beyond the emerical, are rooted in some physical relation ship between cause and effect. This is not true of Moore's Law.
LONG LONG before C2D there were "just" incremental performance upgrades, not doublings/triplings.Quote:
Originally Posted by HKPolice
For the sake of argument, let's presently ignore moving from single to multiple cores, and look at the performance difference, for either camp, just in terms of IPC of the core itself. If you do that, you don't see doublings, etc. in IPC, EVER.
People forget that if you get 10-15% every time there is an arch change, those changes add up, and FAST.
1 x 1.15 x 1.15 x 1.15 x 1.15 = 1.749
Over 4 iterations, that's 75% faster IPC than the base you started out with, and it continues to grow. Then you toss in the ability to add more cores and up the frequency b/c of process shrinks, and you really have some huge jumps in performance.
People were spoiled with the IPC jump experienced with C2D just b/c P4 sucked so bad (can you saw WORSE IPC than P3?).
Moore's "law" is an observation, not a mathematical proof.Quote:
Originally Posted by Chumbucket843
There is no physics to it.
lol hahaha moores law...when will ppl realize that it is not of real consequence? its not a law its a marketing point.
interesting post...Quote:
Originally Posted by ajaidev
wait a minute isn't BD using fma instructions?Quote:
Originally Posted by -Sweeper_
FMAC, other thing...Quote:
Originally Posted by god_43
Actually it is FMA, but without software support it useless.. just as AVX.Quote:
Originally Posted by -Sweeper_
And as we see software hardly useses any of the new insturctions (SSSE3, SSE4.x) when not both cpus support it, we'll see the same for the first FMA implementation in BD. At least both sopport AVX so that will get some use.
More specifically, without compiler support, they are both useless.
I don't think FMA was planned to be in the first version of Intel's AVX, so I'm not entirely sure what the situation is. But this info is like third hand from back when I wanted to know about the Tock after the first i7's.Quote:
Originally Posted by god_43
If it's in Bulldozer, then I guess Intel decided to use it first time round, if not then maybe it's because AMD decided to match Intel for functionality.
FMA comes with haswell, according to intel it would increase coresize to much, so they wait for 22nm.
Quote:
Originally Posted by Hornet331
FMA support is already wide spread in processors and code compilers. C99 supports FMA through the fma standard math library function, nvidia gtx2XX and new gpu's support it, amd's 5XXX gpu's support it, the cell processor, fujistsu SPARC VI, itanium.... BD will support FMA4 and Haswell will support FMA3. :up:Quote:
Originally Posted by mstp2009
also, windows 7 sp1 will support AVX as does linux kernal 2.6.30 :D
What is this "fma standard math library"? The C99 extension to math.h?
Too bad hardly anyone cares about C99 really, partly thanks to Microsoft. They do not support it at all, and have no plans to. Intel doesn't support C99 either, only the major features, same with GCC although their support is way better it isn't still 100 % compatible.
I hope Intel will implement FMA in Haswell in the right way, keeping a throughput of two instruction per cycle (FMA+FMA or FMA+MUL/ADD). One FMA has no advantage over separated MUL & ADD (except a bit better accuracy), but some serious disandvatages.
hah... all that moores law talk... :D
everybody sees something different in it :D
i always get a laugh out of presentations that claim moores law is dead, moores law has been surpassed or moores law cant continue because they did something... :D
depending on how you look at it, moores law has been dead for quite a while, as manufacturing and r&d costs have increased a lot, so you dont "just" get double the transistors every 18 months... if youd spend enough you could probably get 4 times the transistors every 18 months, but thats not the point...
another thing is that intel has been working on a 24months shrink cycle for the past years, and for amd you can add a few weeks on top of that...
anyhow... i think its silly taking this thing so seriously... i mean the original conclusion he made was that transistor density doubles "roughly" every 18 months... "roughly"... i doubt that he meant it to be a serious rule or guideline, and im surprised so many people see it as such and argue over semantics :D
I think you meant to say, "for amd you can add several months on top of that."Quote:
Originally Posted by saaya
Their 32nm looks like it will ship about ~18 months after Intel's, so they've been building quite a few months up each cycle.
Yes, a 4-operand form. Intel is supporting a 3-operand form starting with Haswell.Quote:
Originally Posted by god_43
well the cycles have slowed down overall... i dont think amd has fallen back even more... im not sure...Quote:
Originally Posted by terrace215
when amd went 45nm intel went 32nm...
by the time amd goes 32nm, intel will go 22nm... right?
so the cycles have slowed down cause fabs are getting more expensive, so you have to run them for longer to really get any ROI, 45nm lasted a long time and 32nm will last a while as well...
How do you know? Not trying to be antagonistic, just a question.Quote:
Originally Posted by JumpingJack
I think you mean that the sale and use of extreme editions is a blip on the balance sheet. It's more of a prestige thing than serious sales. Don't be too quick to dismiss "mainstream" overclocking as irrelevant.
And even if it is true that only an abolutely tiny amount of users overclock their systems, these are the guys that generally work in the field of IT or have some sort of connection to it. Meaning that they are more often than not the gateway to regular consumers. For example, I am convinced that the great overlockability of C2D was an important reason for it's huge success. And not only because enthousiasts were pushing C2D to their relations. I believe that everyone started overclocking his C2D at home because it was just so easy to get a 1Ghz incrase for free. Doing so on A64 was not that easy.
Intel is tired of seeing the higher binned CPU's gathering dust as everyone is buying low end and overclocking the snot out of it. This is especially worrying for Intel as the crowd that overclocks is generally the only group that would buy a high end CPU in the first place. So they have to force these K type CPU's onto them, better yet, make them get the more expensive socket with quad DDR3 channels at the end of 2011 to be able to change the bus speed and the multi's.
So we are stuck with a blunt tool called multi overclocking for "mainstream" SB (read: everyone in 2011, and probably over 90% of the market after 2011 as well). What is there to tweak other than Vcore and multi? It's like trying to make a painting using a pneumatic drill.
There seems to be no technical reason for this limitation to multi clocking either. If there was a technical reason to it why would they allow changing the bus speed on S2011 (I'm going by what Terrace215 said earlier in this thread btw)? SB on S2011 is the same architecture with quad channel DDR3 after all, right?
My point is Intel invented an artificial way to milk overclockers (which actually seems to contradict your statement that the overclocking market is tiny). They can do this if they want to, it's their company, but the real reason is that they limit the bus speed clocking just because mobo manufacturers will probably find a way to keep "mainstream" SB overclocking alive if the BCLK can be adjusted. Going the multi way gives Intel absolute control of what we can do. It's just wrong to destroy our hobby with pragmatic thinking like that. It dumbs down overclocking and quite frankly, takes the fun out of a place like XS. I'm directing this at everyone that thinks it's all fine and dandy what is happening now btw. Not at you JumpingJack. :)