perceived performance comes in one flavour. actual performance comes in two.
before i get started, what you need to understand is that the hotter a microprocessor is, the slower it works. the reason is twofold: firstly, for whatever laws of physics, heat somehow interrupts the integrity of an electrical signal(which all aspects on electronics rely on) and rendering it useless and unable to perform 'work'(imagine tripping every other step you made - you wouldn't get very far, very fast now would you?). secondly, the transistors and electronic pathways get physically damaged by heat. in an efffort to make their parts last, designers will make their parts perform SLOWER(this is called throttling) as they hit certain temperature thresholds. when they perform slower, they produce less heat(its a negative feedback loop for you biology people
).
performance comes as a result of architectural designs. the first thing you should know is that exact details of microprocessor architectural are FAR BEYOND my head...you need to be damn smart and damn well educated to understand it. basically, this decides the amount of transistors used in a chip and HOW they are used in the chip. the first thing you need to understand is that GPUs and CPUs are VERY different beasts. the GPU has many, many copies of the exact same hardware inside of it, whereas the CPU has only one of each type of hardware, but those are much larger and very quick and 'powerful.' basically, compare it to this: i could use 1200 labourers to build a dirt bridge to cross a river, or i could use a team composed of a lumberjack, a range-sighter, and a surpervisor to figure out that i could simply jsut chop down a large tree which would gain me access to the other side. hopefully, that helps you see how the chips are built inside. the sum of the internal specialized components, and the interface(largely determined by the internal components) to things such as memory compose the architecture. the optimization, or sometimes complete overhaul, of the architecture(through error-and-trial and application of theoretical concepts) produces the performance increase from generation to generation.
secondly, performance comes as a result of transistor shrinking(each size is referred to as a 'process' - such as the 40nm process). why does this make things run faster? i'll admit, i don't understand all the nuances, but one huge factor is the fact that smaller transistors require less power to run and therefore produce less heat(refer to the above blurb on heat), allowing it to run at higher speeds without the penalties a larger process would face. moving on, like i said, i'm not* entirely sure about this, but i believe smaller transistors can fir closer together, meaning that the electronic signal has a shorter distance to travel before it can 'get to work.'
thirdly, the PERCEIVED performance increase. this has NOTHING to do with design in the slightest. its entirely an effort to milk the most out of the customer(okay, i'm being VERY har dhere, but from the perspective of the customer, its exactly what is happening)...i'm talking about harvesting. what is harvesting? well, we need to go back to the beginning(AGAIN). remember that i said microprocessors were very small things? turns out that small things are very hard to make. so hard in fact, that getting 50% of the chips you;re making actually come out and work perfectly is a fairly acceptable yield. what about the chips that come out wrong? the thing is, even though that the chip may not work properly, it may still work in some degree. alright....so the chip only runs at half-speed, i can't use it as my top-of-the-line product, obviously. what do i do? i HARVEST it and sell it as a lower tier product even though it was originally meant to be a top-tier product. now, through the combination of top-tier products and middle-tier products, i can sell 85% of each batch. thats a HUGE bonus and can mean a successful generation of products or a financial flop. its hard to understand, but i'll give you a few examples below. anyways, how is this a 'rip-off' to the consumer? you know that GTX 260 you just bought? well, its just a HARVESTED GTX 280...YET, when you go buy a GTX 280, you're paying a premium, even though NVIDIA can still make a profit if it sold it at GTX 260 prices.
exmaples of harvested products:
HD 4850 is a harvested HD 4870(theres a little more to the story... a hd 4850 uses a different, slower type of memory than the hd 4870)
GTX 260 is a harvested GTX 280
HOWEVER, a HD 4870 is NOT a harvested HD4890. the 4890 uses a slightly modified version of the RV770 core(it uses the RV790 core).
ps, a little bit of the info could be wrong....i'm fairly intoxicated atm.
EDIT:
http://www.anandtech.com/video/showdoc.aspx?i=3469you should read that....its an article about the development behind the RV770. the GPU(not the card...its important to keep that distinction) that wrecked havoc with NVIDIA this generation. as a general note, i recommend reading the articles written by ANAND himself if you want to know more about hardware...he's a great writer and knows his shit.
ps, i'm not related to Anandtech in anyway, i just like their articles. i hope this doesn't seem like i'm advertising.
EDIT2:
also, you should know that when you shrink the transistors in a given architecture, you generally must REDESIGN the architecture to be use the smaller transistors. generally, this doesn't SEEM to be much of a problem. HOWEVER, for the gt200(or w/e it is called...i've heard so many different names this generation for nvidia), NVIDIA successfully shrunk it from 65nm(or something rather) to 55nm. ALTHOUGH, when they tried to shrink it down to 40nm....to put it simply...it was a CLUSTERFUCK. no 40nm gt200 parts(at least, last time i checked) are released or planned. for those interested, it was the second blow to NVIDIA this generation. it meant that NVIDIA couldn't improve yields*, NOR could it release a proper gtx 200 part for laptops**.
*i didn't explain this above, so i'll explain it here. chips are made on a ROUND piece of silicon(called a wafer). the circle is then cut up into a grid(the pieces on the edge - not square/rectangular - are useless). the smaller your process, the smaller your die will be(if you want it to be, sometimes you'll keep it the same size and just stuff the extra space with more transistors), meaning that you will fit more onto a given wafer...with the end result of earning more profit.
**the gt200 core simply produces too much heat to be used a laptop chassis. you may say that you've seen gtx260m parts out there, HOWEVER, in reality, they're actually G92 cards. ps, nvidia has gotten alot of crap about this...its ANOTHER renaming for them, lol.
Post has been edited 3 time(s), last time on Aug 28 2009, 8:55 pm by Sauceover.
None.