Graphene Transistors @ 100 gigahertz

[JumpingJack]

i will be impressed when a pll can generate a clean frequency at 1THz.
oh wait. http://insidetech.monster.com/news/a...us-to-1000-ghz

IBM does great research but the actual implementation is very expensive which is where intel wins. cost per wafer would be ridiculous with graphene.

IBM is the top technology company in terms of generating fundamental research. My point, really, is that most of it is purely academic -- meaning, much of these 'major breakthrough' announcements, like this one, is most likely not going to be productized (useful), some will, most won't... with IBM it is all about prestige, a lot of show (good, science, show) and no substance.

[JumpingJack]

I said performance NOT Clock speed.

and we ALL KNOW by now that clock speed is only part of the equation for performance. [performance = IPC * clock speed)

Rhetorical since that was my point ... design can be geared to clock speed to sacrifice IPC, and vice versa, or for a given design with good IPC, clock speed can be improved with better transistors.

Power6 was a good example of IBM balancing that approach... Prescott is a good example of clock speed gone awry.

[nn_step]

Rhetorical since that was my point ... design can be geared to clock speed to sacrifice IPC, and vice versa, or for a given design with good IPC, clock speed can be improved with better transistors.

Power6 was a good example of IBM balancing that approach... Prescott is a good example of clock speed gone awry.

unfortunately as clock speed increases IPC GOES DOWN, due to starvation [caches only can help so much] largely because memory latency is holding performance back. http://www.behardware.com/articles/6...-duo-test.html (a legacy read but important for this discussion)

In fact if you drop a 100Ghz version of any processor on the market into a current system, the performance increase would be less than 10%

[zalbard]

In fact if you drop a 100Ghz version of any processor on the market into a current system, the performance increase would be less than 10%

Into a 3 channel x58? Doubt that, it will be a lot more, it has a ton of unused bandwidth, since running the RAM at 1200MHz and 2000MHz barely makes any difference.

[Chumbucket843]

with a physics solver you need oodles of bandwidth because it doesnt cache well. ideally you would need 100GB/s on an i7 to keep it fed but caches and registers reduce the need fairly well. they just use a lot of die space.
the orange system is dual gainsetown and the blue is dual harpertown.
http://it.anandtech.com/IT/showdoc.aspx?i=3536&p=13

[nn_step]

Into a 3 channel x58? Doubt that, it will be a lot more, it has a ton of unused bandwidth, since running the RAM at 1200MHz and 2000MHz barely makes any difference.

Bandwidth IS NOT the Issue; the problem is LATENCY. Assuming you have the world's BEST memory (10.5ns); you are still going to have to wait more than a thousand clock cycles waiting for it to show up on average. Even double the largest current x86 cache [8MB] is not going to satisfy a 100Ghz Processor.
A basic introduction to the subject www.cis.upenn.edu/~cis501/lectures/04_caches.pdf

[Particle]

It's not that cut and dry. That paper you linked to helps to demonstrate that. Keep recently used memory in cache. Fetch chunks around requested addresses. That will reduce average latency below the memory's latency, so the 10.5ns average you're suggesting isn't likely to be correct for real workloads.

[nn_step]

It's not that cut and dry. That paper you linked to helps to demonstrate that. Keep recently used memory in cache. Fetch chunks around requested addresses. That will reduce average latency below the memory's latency, so the 10.5ns average you're suggesting isn't likely to be correct for real workloads.

actually it is going to be correct for the majority of applications. Largely because extremely few applications utilize current generation prefetch instructions. [Since by default MOST applications are legacy] Since data dominates all major applications, hardware prefetch logic will range in effectiveness from 2% to 90% [So a couple orders of magnitude difference in the speed up by clock speed increases]

You are correct in saying that it isn't cut and dry; I freely admit that, but for most applications [Those that lack prefetch instructions and historically have bad performance] the potential performance increase by a 100Ghz version of Gulftown or Istanbul is going to be less than 10% ideally.

Now should an application utilize prefetch instructions aggressively, and aggressively optimize Code and Data segments for optimal prefetch; then they will surely see a considerable increase [Assuming they can hit a cache hit rate approaching 100%]

[saaya]

As someone who does research with graphene, I http://www.xtremesystems.org/forums/...&p=4231463must point out that these devices have no intention of ever working in a digital environment. The on/off ratio of graphene devices is pathetically low due to a lack of band gap, or in the case of bilayer graphene a very tiny band gap during an applied field. The large challenge of making a reliable digital circuit that does not waste power out of graphene is still up in the air.

nice, thx for the inside knowledge!
what do you think about using graphene traces for silicon transistors?

Megahurtz war is over, begun the core war has.

oh and stop bashing ibm guys, its not their fault journalists keep picking up their tech anouncements and try to turn it into something of relevance to our daily life and end up with some bold and outlandish claims and suggestions...

[initialised]

what do you think about using graphene traces for silicon transistors?

I doubt that graphene could cope with current BEOL processes but changing to graphene wouldn't be much more painful than changing from Al to Cu was. Plus carbon is a horrible contaminant in Si.

[flutie98]

nice, thx for the inside knowledge!
what do you think about using graphene traces for silicon transistors?




oh and stop bashing ibm guys, its not their fault journalists keep picking up their tech anouncements and try to turn it into something of relevance to our daily life and end up with some bold and outlandish claims and suggestions...

Its being looked at, however there will be no real benefit over Cu until below ~13nm line width when the Cu resistivity skyrockets.

[Particle]

I would think silver traces would be the next logical step wouldn't they?

[flutie98]

I would think silver traces would be the next logical step wouldn't they?

Ag does have a lower resistivity than Cu, but only by about 2%, its just not an economic solution. There are many other metallurgical issues as well.

[Otis11]

Its being looked at, however there will be no real benefit over Cu until below ~13nm line width when the Cu resistivity skyrockets.

Well, if we're good to 13nm, we should be ok for the next ~18 months... still got 28 and 22nm shrinks to go for GPU and CPUs.

But do memory also use copper? Cuz they're already at 25nm coming out, so we could hit that limit pretty quick there...

[Particle]

I don't see cost being a huge barrier considering how much a wafer costs already. $390 could cover a whole 300mm wafer in 1mm of pure silver. I will defer to you in terms of metallurgical issues. I'm not familiar with that.

[flutie98]

Well, if we're good to 13nm, we should be ok for the next ~18 months... still got 28 and 22nm shrinks to go for GPU and CPUs.

But do memory also use copper? Cuz they're already at 25nm coming out, so we could hit that limit pretty quick there...

Yep DRAM also uses Cu interconnects, However my feeling is that a lithographic wall will be hit before a Cu resistivity issue is. EUV Litho has been in the R&D phase for years, hopefully it will be ready when we need it.

To comment further on Ag, just looked at the price difference, roughly 5x per gram, it would add up. That being said it might happen some day.

[Otis11]

Yep DRAM also uses Cu interconnects, However my feeling is that a lithographic wall will be hit before a Cu resistivity issue is. EUV Litho has been in the R&D phase for years, hopefully it will be ready when we need it.

To comment further on Ag, just looked at the price difference, roughly 5x per gram, it would add up. That being said it might happen some day.

Well, on the performance side... I don't think any of us would flinch too much at paying an extra $30 for a 22nm chip...

So then we need to totally rework the entire manufacturing process to get smaller?

If so, that could take some time...

[>HyperlogiK<]

IBM is the king of marketing hype about their process technologies. The only thing they can do is come up with press releases about how great their technology is. Meanwhile, Intel continues to actually churn out real processors based on their tech. How many years ago did IBM claim to have HK-MG technology? How many processors have IBM shipped with this tech? Until it comes to market, I think IBM is just blowing smoke. Don't believe the hype.

Although Intel has generally held the lead in process technology they haven't always done so. In fact IBM/AMD were ahead in 2004 with their 90nm SOI process. Also the Common Platform alliance will probably have comparable expertise and more RnD budget than Intel does, but equally, IBM/Samsung/Sony didn't exactly manage to set the world on fire with Cell.

But right now, and for some time to come I agree that Intel=Win.

[flutie98]

Well, on the performance side... I don't think any of us would flinch too much at paying an extra $30 for a 22nm chip...

So then we need to totally rework the entire manufacturing process to get smaller?

If so, that could take some time...

Dont worry, the industry has been pouring money into R&D since the early 90's to devise ways around these future (now now) problems. Just too much money at stake not too. We will get it done.

[Otis11]

Dont worry, the industry has been pouring money into R&D since the early 90's to devise ways around these future (now now) problems. Just too much money at stake not too. We will get it done.

Oh, I have no doubt that they WILL get it done, the question is when...

And the interesting part is how!

Not worried about the tech race ever stopping... I'm betting my future career on it.

[JumpingJack]

Although Intel has generally held the lead in process technology they haven't always done so. In fact IBM/AMD were ahead in 2004 with their 90nm SOI process.

Based on what figure of merit?

[damha]

Dont worry, the industry has been pouring money into R&D since the early 90's to devise ways around these future (now now) problems. Just too much money at stake not too. We will get it done.

Necessity is the mother of all invention (motivation).

Your posts are very informative flutie

[saaya]

Dont worry, the industry has been pouring money into R&D since the early 90's to devise ways around these future (now now) problems. Just too much money at stake not too. We will get it done.

well you know, i never figured out why they poured all that money into eulv...
cause its not going to push the limits all that further... using electrons instead of photons instead would allow much higher accuracy, wouldnt it?

and thx for all your posts flutje, very interesting info indeed!

[flutie98]

well you know, i never figured out why they poured all that money into eulv...
cause its not going to push the limits all that further... using electrons instead of photons instead would allow much higher accuracy, wouldnt it?

and thx for all your posts flutje, very interesting info indeed!

Its not about accuracy (it is extremely important but not the figure of merit when comparing litho tech usually). The figure of merit is resolution, the defining parameter is wavelength of incident energy. So for current tech we use immersion based 193nm DUV light this combined with alot of tricks (resist trimming, double patterning etc) has gotten use to the 32nm node meaning the gate length is about 40-80nm depending on the tech. So obviously we gotta keep shrinking and that means we will have to keep shrinking the wavelength of incident radiation. EUV at 13.5nm is a great idea with alot of problems. These will be fixed and I bet we will se this working at the sub 22nm nodes. You mentioned ebeam litho and since an electrons wavelength is tied inversely to its kinetic energy @100kV the wavelength is in the pico-meter ranges. obviously much better than EUV or anything else. But as we talked about in a previous thread, right now ebeam litho is a serial process and is incredibly slow. Hopefully a multi beam system will eventually catch up to the throughput of modern optical systems, I shouldnt say hopefully, it will have to.

[saaya]

Its not about accuracy (it is extremely important but not the figure of merit when comparing litho tech usually). The figure of merit is resolution, the defining parameter is wavelength of incident energy. So for current tech we use immersion based 193nm DUV light this combined with alot of tricks (resist trimming, double patterning etc) has gotten use to the 32nm node meaning the gate length is about 40-80nm depending on the tech. So obviously we gotta keep shrinking and that means we will have to keep shrinking the wavelength of incident radiation. EUV at 13.5nm is a great idea with alot of problems. These will be fixed and I bet we will se this working at the sub 22nm nodes. You mentioned ebeam litho and since an electrons wavelength is tied inversely to its kinetic energy @100kV the wavelength is in the pico-meter ranges. obviously much better than EUV or anything else. But as we talked about in a previous thread, right now ebeam litho is a serial process and is incredibly slow. Hopefully a multi beam system will eventually catch up to the throughput of modern optical systems, I shouldnt say hopefully, it will have to.

oh that was you back then in the other thread as well? thx for your reply
well i know that ebeam atm is serial, well i heard of it, but is it really all that hard to make it parallel or create a bundled stream that works kinda like a inverted mask? it would require a lot of time and money... but thats what eulv did and still does as well... thats what i mean... wouldnt it make more sense to invest the money in a tech that, maybe costs a bit more and takes a bit longer, but then REALLY pushes the boundries further, than spending so much time and resources on having a few extra years and then you run into the limits of that tech already...

i guess back when eulv dev was started ebeam didnt look as feasible as it looks now?