6 ms·
IBM demonstrates manufacturing process for creating 100GHZ graphene transistors
- pwhelan 17y agoWhen this comes to market, the Von Neumann bottleneck compensation techniques will become obsolete in a way. However I would love to see how this will speed up computing in throughput, because this clearly is a major achievement.
- thaumaturgy 17y agoI predict that no fewer than 16 meta-languages and API layers will be piled on top of existing development methodologies, and that all future operating system graphics will be rendered in fully textured 3D, and that development will consist of drag-and-drop objects. So, in short, the end-user experience won't change at all. (Sorry, I'm a little cranky at the moment. Too many slow applications for no bloody good reason.)
- soundsop 17y agoThe bandwidth and latency differences between on-chip memory and off-chip memory are multiple orders of magnitude. Graphene improvement over silicon is only about 1 order of magnitude, so it will not eliminate the von Neumann bottleneck alone.
- anigbrowl 17y agoA more detailed examination of these results here: http://arstechnica.com/science/2010/02/graphene-fets-promise-100-ghz-operation.ars http://arstechnica.com/science/2010/02/graphene-fets-promise...
- rms 17y ago100-GHz Transistors from Wafer-Scale Epitaxial Graphene http://www.sciencemag.org/cgi/content/abstract/sci;327/5966/662?maxtoshow=&HITS=10&hits=10&RESULTFORMAT=&fulltext=graphene&searchid=1&FIRSTINDEX=0&resourcetype=HWCIT http://www.sciencemag.org/cgi/content/abstract/sci;327/5966/...
- rauljara 17y agoSo, about 10 years until they make their way into pcs. Hopefully hard drives will be fast enough to keep up with these new processors by then. It'd be really frustrating to have a computer with a processor that's 50x as fast, but the computer only runs about 1.5x overall.
- bjelkeman-again 17y agoI don't think we will have "hard drives" then. The solid state disks are starting to come down in price and the speed is nice in comparison. But, yes, I feel your pain.
- goodgoblin 17y ago"'Roads?' Where we're going we don't need 'roads'."
- rms 17y agoThat should be long enough where we have memristor based storage -- all the storage capacity of hard drives with the speeds of super-fast fast RAM.
- sparky 17y agoIt should be pointed out, for the edification of those who are not VLSI people, that transistors with an f_t of 100GHz are not the same thing as transistors from which one can build a credible 100GHz processor. f_t is the input frequency at which the gain through the transistor (biased in the linear region) is unity (1). This has some relation to how fast the transistor can switch from 0 to 1 when biased in the saturation region (the way digital logic works), but it is not 1:1. Likewise, the cycle time of a processor is the delay of the longest path from one flip-flop or latch (storage element) to another flip-flop that is expected to take 1 clock cycle. The length of this path is typically expressed as a multiple of the propagation delay of an inverter driving 4 other inverters (delay grows with output load, or fanout) (FO4 delays http://en.wikipedia.org/wiki/FO4 http://en.wikipedia.org/wiki/FO4). Typically, high-performance processor microarchitectures are designed with a 5 (Pentium IV) - 25 (slower but more power-efficient) FO4 delay critical path. Once you account for clock skew, jitter, process/temperature variation, etc., your processor frequency is 1 to 2 orders of magnitude lower than your transistors' f_t. The authors of this paper are careful to point out that their graphene transistors' f_t is higher than CMOS transistors of the same gate length (240nm), but that is comparing with a ~12-year-old CMOS processes. Modern CMOS processes are faster than this (for instance, this paper http://wwwtw.vub.ac.be/elec/Papers%20on%20web/Papers/DLinten/%5BLinten%5Djournal.pdf http://wwwtw.vub.ac.be/elec/Papers%20on%20web/Papers/DLinten... assumes a 90nm RF CMOS process with an NMOS f_t of 150GHz). Graphene can certainly catch up once they figure out how to shrink the devices, but it will take more than productizing the current state-of-the-art to compete with CMOS.
- rjurney 17y agoI don't know what you just said, but I'm glad someone is on top of this stuff. Keep up the good work!
- sparky 17y agoHaha sorry about that. The short version is: 1) These graphene transistors are faster than CMOS transistors of the same size, but CMOS transistors of the same size haven't been used for most purposes in about a decade. Modern CMOS transistors (the kind used in your Core 2 Duo or Core i7) are roughly twice as fast. 2) 100GHz transistors => ~1-10 GHz processors, not 100 GHz processors.
- grinich 17y agoJust when you thought Moore's law was broken.
- ntoshev 17y agoThe longest path in a 100 GHz chip would be 3 mm or a signal travelling at the speed of light won't have enough time to reach the other endpoint.
- wmf 17y agoSee sparky's comment; 100 GHz transistors allow you to build a ~10 GHz chip. Even at 2 GHz, NetBurst struggled with wire delay back in 2000.
- andreyf 17y ago<3mm cores might be a great idea - as multicore programming evolves, this may make possible thousands or millions of cores running at very high clock speeds.
- wmf 17y agoTilera's "KILL" cores are probably a bit under 3 mm x 3 mm, although I don't think massive multicore is a good idea for many applications. http://gamma.cs.unc.edu/EDGE/SLIDES/agarwal.pdf http://gamma.cs.unc.edu/EDGE/SLIDES/agarwal.pdf
- andreyf 17y agoI don't think massive multicore is a good idea for many applications But hey now, that's the problem of the applications, isn't it? There's probably quite a bit of exciting potential with what you can do with massive multicore architecture - slide 6 of the presentation you linked to being a very convincing reason why.