I recently got into a conversation in comp.arch concerning how fast signals propagate and how far they can travel in microprocessor wiring.
Some of the posters seem to think that wafer-scale traces/wires are a lot slower than PWB-scale and system-scale traces/wires because they are RC transmission lines, not LC.
I did a few crude simulations and it seems to me that the RC slows down the risetime on single edges and cuts the amplitude way down on high frequency clock signals, but I can't see any reason to think that the propagation would be a lot slower than the usual 60%-80% of C rule of thumb.
I am familiar with normal board and system level transmission lines such as ECL, stripline, coax, etc., but have never done any work with chip-scale electronics. Does anyone here know how fast and how far one can move a signal across a die? Thanks!
I can't answer your question directly, but I recently asked a very knowledgeable Xilinx employee (on comp.arch.fpga) whether their latest FPGAs used termination techniques for any of their internal signals, and his response was "yes". So, if this is true (I have no reason to doubt this particluar individual) then this means that their internal signal edge rates are fast enough and their path lengths are long enough to warrant the cost and complexity of on-chip termination (and I'm not talking about their I/O termination features).
Guy Mac I have been doing a web searches on this, and found these:
_The Future of Wires_
formatting link
_The Wire_
formatting link
Parasitic Extraction and Performance Estimation from Physical Structure
formatting link
formatting link
I also found this in an abstract (the actual paper isn't online):
A 3Gb/s/wire Global On-Chip Bus with Near Velocity-of-Light Latency
"We successfully show the practical feasibility of a purely electrical global on-chip communication link with near velocity- of-light delay. The implemented high-speed link comprises a 5mm long, fully shielded, repeaterless, on-chip global bus reaching 3Gb/s/wire in a standard 0.18 µm CMOS process. Transmission-line- style interconnects are achieved by routing signal wires in the thicker top metal M6 layer and utilizing a metal M4 ground return plane to realize near velocity-of-light data transmission. The nominal wire delay is measured to 52.8ps corresponding to 32% of the velocity of light in vacuum.
Prop delay versus distance is a serious issue with layout on a Xilinx FPGA. If the chip is 1 cm square, and you can easily get a ns of routing delay, that's c/30 right there.
The RLC transmission occurs when the series loss component is too high to "ignore" also adds delay to the propagation time. The classic wave equation and simpliciation all use transmission constant, with rsitive looss it is a complex ni,ber, hance larger for the same L and c and hence slower...
Better look at it as an RLC transmission line or as a lossy LC line. See if you can then calculate the speed, in the same way one alculates for a coax and then a PCB trace with ground plane. Then do some measurements to see how the theory matches practice.
You can get more bandwidth on the fatwire levels (high in the stack), basically because the capacitance per unit length is the same, and the resistance goes down. Of course, the number of wires available goes down too, and you have to get through all the lower wiring levels to get up to the fatwires. This leads to wireability problems.
In something like a highly multicore processor, you need lots and lots of fast wires. There's a current DARPA program called TELL (for terabit electrical links at low power, or something like that). TELL is all about finding out how far electrical links can be pushed (and therefore at what point you have to go to optical links or accept reduced performance). Since wiring capacitance tends to be independent of size scaling, it's really hard to get below 2 pF/cm, so to save power, links need to use very low voltage, at which point you get into nasty problems with noise, drift, offset voltages, and crosstalk.
This gets especially difficult starting with the 32 nm node (iirc again), because the threshold voltages of the FETS are hard to control...it turns out that you have to worry about statistical variations in the number of dopant atoms in the FET channel. A 30-nm cube of silicon, doped to 10**20 per cc, contains 2700 dopant atoms. If a chip has a billion transistors, you'll have lots of 6-sigma outliers, which will be off by +- 300 atoms, or 11% of nominal, which causes a nasty threshold voltage shift. Smaller devices get worse fast.
The wire guys think they can overcome these problems, and maybe they can...but my money says that we'll see on-chip optical signalling in the next 10 years. (I also have to try keeping my management convinced.) ;)
Forgive me for a rather nasty question. What is this phoney push for more junk on a piece of silicon to support what used to be relatively simple applications? What was wrong with KISS? In a personal computer, one does not need 2^10 core CPUs, or even dual core; any CPU speed over 1Ghz is wasted, and for 99+% uses Win98Se is more than good enough. Now, if one gets into graphics (read: games, design PCBs or other complex artwork), then more speed becomes useful and Win2K becomes a better choice. Oh, you say, we "need" dual (or quad) core for graphics. What the hell is that large graphics chip on the fancy video card for? Boat anchor? In fact, what good was the MMX instruction set for, since the sound card already supported those functions. On a cell phone guess what - its purpose is to send and receive calls, period. Want to do something else like portable music - players have been around for over 10 years that do that; they just get smaller and store more. Etc etc and etc (courtesy of Yul Brynner in the King and I).
You are assuming that every computer is a PC. Some computers are not PCs. Some computers are routers with 48 separate gigabit ethernet ports and a requirement to inspect every packet coming in from every port. Some are part of large web server farms that handle all of the searches on Google or all of the bids on eBay. Some are doing simulations of complex phycical systems. Some are doing the rendering for the next Jurassic Park movie. And some are at Supernews, handling your posts and the post of millions of others.
It's not a phoney push; there are plenty of applications where there will simple never be enough CPU horsepower available to solve them... at least not using technology that resembles anything like what we have today.
Ever for "simple" problems such as PCs running word processors, web browsers, etc., CPU horsepower helps a great deal: The average person today can walk up to pretty much any modern PC and get it to do "useful" things, whereas 20 years ago even if you were a skilled, e.g., IBM PC user you couldn't just walk to an Apple II or Commodore 64 and do "useful" things without a fair amount of instruction.
That may be for you, but not everyone feels the same. 20 years ago I'm sure I could have found someone saying that any CPU speed over 10MHz was wasted, no needed more than 640KB (hmm... there's an infamous quote!), and for 99+% uses DOS worked fine.
The signal processing in a modern cell phone is quite impressive -- hundreds of MIPS go into it.
For one thing, IBM (where I work) doesn't even make PCs anymore.
Depends on your problem set. I have a 14-processor Opteron cluster in my lab, used only by me, for electromagnetic simulation and device design. I put it together for a song (about $12 altogether), but some of the problems I use it for can't be solved on a machine with less than
30 GB of RAM. Google probably isn't going to try to run on a single uniprocessor box of any speed whatsoever.
I'm glad you're happy with what you have. I use computers of varying ages too...my office machines are 4 and 10 years old respectively, and they're both dual-processor SMPs, because I push them pretty hard sometimes. I've been writing multithreaded code since OS/2 2.0 came out in 1992. I also like using old apps--for instance, Wordperfect 5.1+ for DOS *flies* on a modern machine.
On the other hand, there are enough customers for the fastest machines (who know very well what they need) to keep me in beer and skittles, anyway. I like doing things that are useful and fun. Why do you do what you do?
It's ironic that most of the compute power in the world goes to gaming. The most compute-intensive thing we do, in fact the only compute-intensive thing we do, is fpga p+r. Design-rule checking the most complex pc board we make takes about 5 seconds on a standard-performance PC. The rest of what we do is dominated by our DSL rate.
Even Spice usually runs fast. I guess em simulation could be slow, but we rarely do that, thank Goodness.
Intel must be running scared; some day pc's will be good enough and become as exciting as toasters, and $5 Taiwanese cpu's will be powerful enough.
You can bet that they do nothing but encourage Microsoft to build OSes that consume vast quantities of CPU power performing, e.g., animation, transparency, 3D effects, etc. The IT "industry" also seems to encourage greatly increasing PC resource usage -- it's very common today that computers in larger companies run on-access virus scanners and create backups of every single file as soon as it's re-saved.
Intel does have a pretty smart pricing strategy -- the prices of their new CPUs drop almost like clockwork, to keep them competitive with offerings from, e.g., AMD, VIA, etc. AMD was doing well for awhile with their Athlon CPUs, but Intel came back with cheap dual-core units and now AMD is back to being not much better than an "also ran."
There is plenty of emphasis these days on "performance per MIP," which is a good thing for just about everyone -- large data server farms care about power consumptions just as much as battery-powered laptop users.
Joel Koltner snipped-for-privacy@yahoo.com posted to sci.electronics.design:
Actually the question is more an ongoing issue of what is currently affordable. A typical new desktop has more (and faster) memory, disk, and compute power than a 1960's (or even a 1970's) supercomputer.
Let's see, 1987, i could make all three of them (as well as some flavors of unit workstations) stand up, beg, roll over, and do most anything i wanted of it.
Maybe, if you like being infected with various forms of malware.
What part of that do you not understand is ridiculous hyperbole.
Specialized applications need to have specialized hardware & firmware to handle the particular needs. Routers and webservers are very good examples of tossing software at a problem just because a "fancy PC" has quad core, 1+Ghz FSB, etc. They really should have specialized hardware & firmware; the more in hardware, the more robust (read: harder for hackers to crack and/or overwhelm). Complex simulations points to vector processing and the "latest and greatest" PC ain't nowhere close to that. Real fancy graphics for a movie points to one or more high end (video) graphics cards perhaps controlled by a PC-like multiprocessor. Etc.
A long time ago, the PC/XT came out and not too long afterwards, there were database programs, spreadsheet programs, and word processor programs. I know that at least one of those spreadsheet programs was used by an accountant to set up and run a company with 4 or five seperate divisions. He started not knowing anything about computers and with no seperate help, in two months had a complete system, with reports, running flawlessly. Word processing? One WP simulated the Wang that was an "industry standard", so a user needed no re-training moving between the two systems. Others were also built on previous WP practices. So, in many cases, a "fair amount of instruction" was not needed.
** Most definitely not home number crunching..
** I run three OSes: DOS/Win3.11 for files that were generated back in the DOS daze as well as doing projects that almost nothing else will do; Win98Se for 99+% of my offline and online work (totally immune to some of the more current hacks, and user base too small to be a target), and Win2K for the "fancy" stuff like CorelDraw, Spice, PCB work and on occasion, multi-million digit software work. My other computer is an older P2-266 running DOS almost 100% of the time, but it also supports the other 2 OSes. Used mainly to run an old A/D board for datalogging via custom programs written in BASIC and compiled for use.
ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here.
All logos and trade names are the property of their respective owners.