Are ASIC's the endgame?

[Meatball]

Considering BFL has said their going to be running 65nm parts and the big CPU vendors are pushing out 22nm stuff already, it's not a stretch to imagine ASIC's getting down to 22nm or smaller.  That's give you even more power efficiency.

[cedivad]

Considering BFL has said their going to be running 65nm parts and the big CPU vendors are pushing out 22nm stuff already, it's not a stretch to imagine ASIC's getting down to 22nm or smaller.  That's give you even more power efficiency.

We will see a 22nm ASIC only in 10 years or if bitcoin becomes mainstream before that.

[bcpokey]

Considering BFL has said their going to be running 65nm parts and the big CPU vendors are pushing out 22nm stuff already, it's not a stretch to imagine ASIC's getting down to 22nm or smaller.  That's give you even more power efficiency.

We will see a 22nm ASIC only in 10 years or if bitcoin becomes mainstream before that.

Yes. Let's not forget that the only company doing 22nm is Intel (amd has not been able to get a 22nm process for their CPUs, and I believe Intel boxes them out of fabs for 22nm as well). And the last Intel 65nm CPU was the Intel® Core™2 Extreme Processor QX6700 which came out in Nov-07. 5 years to go from 65nm -> 22nm for a Multi-BILLION dollar microprocessing giant. Admittedly once it has been done, it becomes easier to do again, but the cutting edge of ASIC technology is ~7 years behind.

[Gabi]

AMD makes 28nm GPUs 

[Jaw3bmasters]

So it's safe to assume the ASIC tech will progress like Pentium......just keeps getting smaller for multi-cores.....

[abeaulieu]

So it's safe to assume the ASIC tech will progress like Pentium......just keeps getting smaller for multi-cores.....

It's already more advanced than than most of the pentium series of processors. I believe they didn't even get to 65nm until the Pentium 4.

The problem with this "things are going to keep getting smaller" theory is that we're approaching the laws of physics with transistor sizes this small. And we're already quite limited to clock rate because of physics.

[BitBlitz]

The problem with this "things are going to keep getting smaller" theory is that we're approaching the laws of physics with transistor sizes this small. And we're already quite limited to clock rate because of physics.

Bah.  We've been "approaching the laws of physics" for decades.  About 20 years ago, I heard declarations that silicon will never break 250MHz.  The frequency limit has more to do with line lengths, turns, and signal phasing than transistor sizes.  I've already heard about silicon switching tested around 500GHz.  Doing that with complex data paths and keeping the signals phased correctly is where the magic is.

[ElectricMucus]

The problem with this "things are going to keep getting smaller" theory is that we're approaching the laws of physics with transistor sizes this small. And we're already quite limited to clock rate because of physics.

Bah.  We've been "approaching the laws of physics" for decades.  About 20 years ago, I heard declarations that silicon will never break 250MHz.  The frequency limit has more to do with line lengths, turns, and signal phasing than transistor sizes.  I've already heard about silicon switching tested around 500GHz.  Doing that with complex data paths and keeping the signals phased correctly is where the magic is.

Well we the Van der Walls radius of a silicon atom is 220pm, so it is reasonable to assume that any transistor must be at least 6 times that. (3 atoms two for each junction) that's 1.32 nm. Then you probably cannot 'dope' a single atom so that figure is double again to 2.64nm.
If you then consider that any semiconductor must include a certain doping ratio that figure (the minimal number of atoms per junction) becomes larger and larger so we are not that far away and it's reasonable to assume that transistors can only get one magnitude smaller than they are currently at best.

[rini17]

If someone finds someway to reduce the work by for example 20 bits it would be a million times faster.
That trick could work for all current technologies, cpu/gpa/fpga/asic.
But keep in mind, the trick would be something many cryptographers haven't found yet. So personally I would bet on quantum computers.

This is actually very good comment and such thing is already researched: http://en.wikipedia.org/wiki/SHA-2#Cryptanalysis_and_validation . Even if it does not endanger full SHA hash function yet, it is really possible someone soon figures out how to use these "meet-in-the-middle preimage attacks" to reduce the work by at least few bits. I'm sure there are people hard thinking about it already.

EDIT: And it does not mean it will necessarily work for all technologies. If it will need gobs of parallel computation, FPGA/ASICs are way to go. If it will rely on accessing terabytes of memory (think rainbow tables), CPU will suffice.

[BitBlitz]

Well we the Van der Walls radius of a silicon atom is 220pm, so it is reasonable to assume that any transistor must be at least 6 times that. (3 atoms two for each junction) that's 1.32 nm. Then you probably cannot 'dope' a single atom so that figure is double again to 2.64nm.
If you then consider that any semiconductor must include a certain doping ratio that figure (the minimal number of atoms per junction) becomes larger and larger so we are not that far away and it's reasonable to assume that transistors can only get one magnitude smaller than they are currently at best.

Before you go too far with that, 2 and 3nm transistors have been demonstrated.  You still want to say our noses are against the silicon performance wall?   

[ElectricMucus]

Well we the Van der Walls radius of a silicon atom is 220pm, so it is reasonable to assume that any transistor must be at least 6 times that. (3 atoms two for each junction) that's 1.32 nm. Then you probably cannot 'dope' a single atom so that figure is double again to 2.64nm.
If you then consider that any semiconductor must include a certain doping ratio that figure (the minimal number of atoms per junction) becomes larger and larger so we are not that far away and it's reasonable to assume that transistors can only get one magnitude smaller than they are currently at best.

Before you go too far with that, 2 and 3nm transistors have been demonstrated.  You still want to say our noses are against the silicon performance wall?  

Yes.
IIRC that weren't 2 to 3 nm transistors but transistors with a resolution of 2 to 3 nm (that's ~6 atoms). It's close to what is physically possible but not the end of the line.
That is I always assumed the nm figure in microelectronics refers to the feature size, alas the average size of an element, the transistor, correct me if I'm wrong. In that case the demonstrated transistor would even be about at the end of the line.

However, current tech approaching it's limit doesn't need to mean progress must stop here. We can for one produce larger chips or even wafer scale designs, stack them for "3D chips" and finally something like "holographic computers" where the properties of the quantum states of individual atoms are used for useful computation. The next limit would be the plank units but that's a long way.