CPU design

[Xpl1c1t]

I always been curious as to how CPUs are designed and the steps it takes to put together a new core.
So i guess some of my questions are:
Can cores be manufactured at any micron size, are the cores just blueprint designs that are imput into computers and printed on silicon?
Is each transistor in a CPU been put where it is for specific reasons? Or do they say, "yea, we need allot of transistors on this part, yep"?
Why the fug does it take so long to design a chip? What is the average ammount of time it takes to design a CPU.

Answer all this and ill give you a cookie! :gift::beer:

 

[labrat25]

i don't know about the rest of the problems but i can answer your first...

the micro size relates to the wavelength of light they use to print the CPU, the shorter the wavelegth the small the size but it gets harder and harder to work with shorter wavelegths of light.. the advantages are lower power consumption and more dies per wafer (higher profit margin)... right now they're pushing into the UV

 

[pillage2001]

Is each transistor in a CPU been put where it is for specific reasons? Or do they say, "yea, we need allot of transistors on this part, yep"?

Transistors are used to form gates and gates are used to form the logic in one CPU.


Why the fug does it take so long to design a chip? What is the average ammount of time it takes to design a CPU.

Designing the chip is one thing, mapping them into a IC layout is another thing. If designing takes 1 day, mapping would take 2 to 4 times as long.

I once had to design a 8X8 SRAM and it took 16 hours to design and conduct testing while the IC layout took 3-4 days. My partner was on the virge of going into the psychitric ward after doign the IC layout for 3 days and it failed.

 

[ai42]

Processor design (GPU, etc) are all done on computers these days.

Honestly, Intel may actually the most raw processing power of any company and not because they are testing chips. Their datacenter I'm sure is quite massive.

You are not going to pay a electrical/computer engineer 50k+ a year to sit on a CAD design and place one transister here, and the next transistor there etc. That would litteraly take years to do with 30+ million transistors. So what they do is sit down and figure out main logic portions of the processor. Decided how to get the CPU to do a given instruction. Then they setup like 5,000+ guidelines for the computer to follow and tell the computer to have fun finding a transistor arrangment that will do all of that. Once that is done they go back and optimize it and add extra data pathways just in case one of the lithography lines dosen't connect right. And finally rearrange them for the best thermal spread across the processor. Obviously all of this simulation and design takes massive amounts of processing time thus the need for a large datacenter.

Now once the design is done they "tape it out" and send if off to the fabs to be made. And once at the fab they will make a few hundred tests and examine them closely to make sure they are working up to spec (which rarely happens). Then they go back with the reports from the fab and modify the deisgn for maximmum yield.

Supposedly the next gen Centrino Donthan has run into fab issues and is being redesigned, and that is the reason for the delay.

 

[AWhackWhiteBoy]

Nvidias engine room

thats how GPUs,CPUs,etc are made.

 

[RaynorWolfcastle]

Originally posted by: ai42
Processor design (GPU, etc) are all done on computers these days.

Honestly, Intel may actually the most raw processing power of any company and not because they are testing chips. Their datacenter I'm sure is quite massive.

You are not going to pay a electrical/computer engineer 50k+ a year to sit on a CAD design and place one transister here, and the next transistor there etc. That would litteraly take years to do with 30+ million transistors. So what they do is sit down and figure out main logic portions of the processor. Decided how to get the CPU to do a given instruction. Then they setup like 5,000+ guidelines for the computer to follow and tell the computer to have fun finding a transistor arrangment that will do all of that. Once that is done they go back and optimize it and add extra data pathways just in case one of the lithography lines dosen't connect right. And finally rearrange them for the best thermal spread across the processor. Obviously all of this simulation and design takes massive amounts of processing time thus the need for a large datacenter.

Now once the design is done they "tape it out" and send if off to the fabs to be made. And once at the fab they will make a few hundred tests and examine them closely to make sure they are working up to spec (which rarely happens). Then they go back with the reports from the fab and modify the deisgn for maximmum yield.

Supposedly the next gen Centrino Donthan has run into fab issues and is being redesigned, and that is the reason for the delay.

Your description sounds pretty accurate from what I know, but AFAIK some critical pathways are still designed by hand. This is, in part why it takes so long for CPUs to jump to the next generation. I'm pretty sure that it takes 2+ years to design/validate a modern CPU. Also, I think that GPU design (which uses a lot more parallelism), is almost entirely based on HDL code. I'm sure you can find a fairly detailed description of how the design is done somewhere on the net.