http://www.anandtech.com/show/10183/intels-tick-tock-seemingly-dead-becomes-process-architecture-optimization
What do you guys think about intel abandoning tick tock?
This will probably be the end of Moores law which was true for 50 years.
Why are they doing it? do they just want to jew out as much money as possible? or is it technical difficulties?
Probably technical difficulties. We're rapidly approaching the point that what moore's observation (transistors on a chip) was originally about can't continue.
That's not to say that technology won't keep increasing exponentially (totally unknown).
>>53665965
Yeah i guess they had to abandon silicon at some point i just didnt thought it was going to be so soon especially since they were already developing 5 and 7nm for some years.
Also this means the next cpu cycles will deliver even less performance improvement
>>53665918
I'll give you a brief synopsis of what has happened over the last few years at intel:
>Head Jew: We need to out pace the goys and maintain process dominance. We can't let anyone produce and sell chips with the same margins we have!
>Engineer Jew: Oy! But to do that we would need immense R&D spending to shrink our BEOL further than every other foundry on a competing node.
>Finance Jew: Fuggedaboudit, it'll pay for itself in da end. Trust me, I know a thing or two about returns. After all my entire family returned from da camps.
>Head Jew: Just make it happen, no matter what!
Some time passes, and intel transitions from 22nm Trigate to 14nm Trigate
>Engineer Jew: Its teeeribble, just teeerible. Its like a second shoah. The poly pitch is too tight, our yields are just teeerible. We're losing money on every chip. We won't be able to launch desktop Broadwell on time. Even Core M binnings are a mess.
>Finance Jew: Look deez things happen. Fuggedaboudit. Things will turn around.
>Engineer Jew: Its TEEEERIBBLE! It will take years to work out da kinks!
>Head Jew: We need to spin this into a positive so the goyim don't find out. Contact market and get them drawing up some slides for a press release.
>pic related
Intel is becoming AMD see:
http://wccftech.com/intel-broadwell-ep-xeon-e5-2600-v4-march-launch/
>A 5.1 GHz CPU with a TDP too high for any air cooling
>A 10 core desktop CPU
>A 22 core server CPU
>>53666439
So their not allowed to release any powerful CPUs? Is that what you're saying?
>>53666211
Kek
but "Process-Architecture-Optimization" really sounds like something quickly made up so normies will think "o wow le new process n stuff such improvement"
>>53666463
>more cores = powerful
AMD has already disproven that
>>53666211
wow, you anti-semites really are nuts, aren't you?
>>53665918
>This will probably be the end of Moores law which was true for 50 years.
Moore's law was a decision. Slower progress doesn't mean no progress.
>>53665918
>10^-4
>10^-2
>0
>10^2
>0
Who the fuck made this retard graph?
>>53666526
No, you're saying Intel can't even THINK about releasing a CPU that can only be water cooled, or have a lot of cores or have a high TDP. No matter the market.
>Muh single core!
>Muh low power consumption!
How the fuck is AMD a better processor? More cores doesn't necessarily mean better performance. Do your research faggots AMD is built by Jews because obviously they don't know what the fuck they're doing. Intel has much better benchmark performance with integrated graphics, while AMD runs hot with high power usage which causes them to become more unstable with high temperatures. Check this out here, the best of AMD and Intel:
http://cpuboss.com/cpus/Intel-Core-i7-6700K-vs-AMD-FX-8350
>>53665965
Moore's law was a stagnation policy introduced so that they did not have to do anything in the innovation department.
Hence why we have the same fucking instruction set for ten fucking years.
>>53666872
>not have to do anything in the innovation department
well they did have to come up with a smaller manufacturing process every two years
>>53665918
stop posting day old stuff slowpoke
>>53666536
>0
Holy shit you're right.
Can't be unseen.
>>53666536
the graph still makes sense because no cpu is directly on the 0 line just very close
>>53665918
> vertical axis is FLOPS/$1k
> mouse brain is 10^11 FLOPS/$1k
> human brain is 10^15 FLOPS/$1k
> capacity equivalent to cost efficiency???
Ray Kurzweil, not even once.
>>53666872
No, it was an observation about how transistor density was growing that proved to be true longterm.
Instruction sets are mostly meaningless. They don't really need to be updated very much, and desu they're too big already in x86
>>53667512
I just downloaded one of the first pictures google gave me without looking too closely but that picture really is amazing
>>53665918
why do you change your cpu every couple of years?
most people I know would be ok with a core2duo + facebook icon + other social + wiki + google (optional) + youtube + netflix
why do this with double xeon 16gb ram?
at some point people would stop the tech hype and settle for the current cpu
singularity when
>>53667150
>y scale discontinuous
Sure makes sense
Moore's law would be mathematically impossible to uphold in a finite universe.
Exponential growth doesn't exist in reality. At best, you can only ever get sigmoidal growth.
>>53665918
Current lithography has reached its limit.
They have to switch to completely different machines called EUV (extreme ultra violet).
EUV is an engineering nightmare.
The short wavelength can't pass through glass or even air.
So everything has to be placed inside a vacuum chamber, and they have to use mirrors instead of glass lenses.
Mirrors suck becasue the lose a lot of energy, which means the light source (a plasma) has to very powerful .
The powerful plasma in turn can easily damage all the delicate equipment.
Only one company (ASML) can even produce EUV machines.
So they have zero competition, which doesn't help progress either.
>>53668063
Better battery life.
>>53666526
There's a reason those 16/32 core server cpus have a low clock speed.
>>53667544
>it was an observation about how transistor density was growing that proved to be true longterm.
That's true.
But it was also a self-fulfilling prophecy.
Because the delivery schedules were adjusted to meet the expectations of Moore's law.
It became a goal to aim for, rather than just a natural progress of technology.
>>53665918
Can someone red pill me on tick tock?
>>53668182
mirrors aren't normally bad for optics at all, it's just that EUV requires dielectric mirrors (i.e., dozens of precision sub-wavelength thickness stacked layers) to work at all, and even then they absorb a big portion of incident light. that in turn means that the mirrors have to be water cooled to not distort and melt, which again means more cost.
like all things, costs should decrease with scaling, but the minimum infrastructure costs will still be higher than the good old days.
multi-patterning and EUV are scary from a cost front, and I don't even see how things can be moved past that to soft X-rays, which will tend to scatter high-energy electrons all through the photoresist and substrate.
>>53668242
Intels way of saying 5% boost of speed per generation by doing nothing on every tock
https://en.wikipedia.org/wiki/Tick-Tock_model
>>53668182
And to add to that, even if EUV machines were up and running in Intel fabs right now, we're still hitting limits of silicon. A silicon atom itself is about 0.2 nm wide. The next die shrink is already 10 nm. That's just 50 Silicon ATOMS wide.
Soon after, we'll hit the limits of stability. It won't matter if we could precisely move a single atom at a time; the circuit would still be unstable.
>>53668283
nehalem, SB, and haswell were all pretty decent improvements, it's just skylake that's been unmitigated dog shit.
>>53668309
> reminder that minimum feature sizes are still several-fold the process node name, because reasons.
>>53668309
quantum computers when?
>>53668242
They could easily combine both into a single product refresh.
But this way they can sell you a new CPU every year instead of once every two years.
>>53668309
>>53668281
>>53668182
A question what kind of cpu we will be using like 20 years from now on ?
>>53668370
> implying western civilization will still exist in a recognizable form in 2036.
>>53668314
I'll give you nehalem and sandy but haswell was dog shit as well when compared to sandys boost.
>>53668370
Likely graphene, gallium arsenide, or optical chips, or some mix of the three.
The most extreme option would be diamond based chips, but then theres the issue of manufacturing that much substrate to sate the world's demands, and dealing with DeBeers getting all uppity.
>>53668352
Quantum computers aren't the answer.
They are only good for certain tasks.
Optical computing might be an answer.
I don't think it will do much for processing power per core, but it might make multiple cores work together more efficiently and improve data bus speeds.
>>53668182
That sounds pretty complicated. Out of pure curiosity, how complex were the older ways of doing it and has technology advanced to a point where one of the particularly old ways is now possible to DIY?
>>53668370
My money is on optical.
It can be integrated with tradition silicon chips and there are already products on the market.
Many people say graphene, but I don't believe in that for the same reason we're still using piston engines instead of (theoretically superior) rotor designs.
The investment needed to get graphene on the same level as silicon is enormous, and in the meantime you won't sell anything.
Silicon just has too much of a head start.
>>53668398
>>53668399
>Optical
how exactly does optical do anything better than low-power interconnects?
before you even get to the lack of the known optical equivalent of a transistor, how would they even have advantageous density given that semiconductors already have transistor and wiring sizes well below the wavelength of visible and most UV light?
The problem is not with Intel (or AMD for that matter). The problem is with Computer Science. I own an i7-4770k but some family members are happily working with E6600s. There is no discernible difference between 95% of our daily computing experience. Programmers are failing to utilize the full potential of modern processors.
>>53668399
I dont get optical computing cpus are already smaller than the wavelength of light. So chips would need to be way bigger or less complex.
Yeah i get that you achieve a higher bandwidth with light but i cant imagine that it would make such a huge improvement.
>>53668472
https://www.youtube.com/watch?v=Nv2TxiwAquM
>>53668472
The old way required a huge lens, it alone costing millions.
It also requires extremely accurate measurements to ensure the wafer is lines up correctly and to compensate for an uneven surface.
But I think very low resolution, say several millimeters instead of nano meters, could be possible.
Essentially it's just photography.
Why do people pick haswell over broadwell? If broadwell is newer?
>>53668574
>The old way required a huge lens
Is that due to size of the wafer being worked on/amount of chips being made at a time or for other reasons?
>>53668499
>>53668513
Two main advantages: higher bandwidth and less heat generated.
I don't think the actual computations will be done optically, just things like data paths, memory controllers, etc. - those make up quite a large portion of your electronics.
So called "optical" chips are really hybrids.
>>53668499
Well, not full optical. Partial optical (light assisted transistor switching has been recently developed) could definitely work though. Replacing the metal clock grid alone could result in a simultaneous performance increase and power consumption decrease.
Light has a much much higher propagation velocity than an electronic signal in wire, so clocking would be easier to do, and can be done faster.
>>53668617
The lens + the immersion lithography is to allow the large mask image to be successfully shrunk down to the design size.
>>53668575
because there were literally only 2 desktop Broadwell CPUs released, and the Xeon Broadwells aren't even fully released yet (EP/E5-v4 unveiled in a week supposedly?).
Broadwell, Skylake and (probably) Kaby Lake (will) have been pretty disappointing releases.
>>53666858
>http://cpuboss.com/cpus/Intel-Core-i7-6700K-vs-AMD-FX-8350
>let's compare a 2012 processor to a late 2015 die shrunk processor, they'll surely be comparable.
>>53668563
That's nothing like making the actual transistors yourself, which I think is what he was asking.
>>53668314
>hasles
>decent
lmao
>>53668660
> light assisted transistor switching
only work in optical-cavity sized devices as far as I know, so nowhere near feasible for most needs at present
>replacing the metal clock grid
similarly, I think only the large-scale clock distribution has been seriously discussed, with regional electrical distribution trees still doing most of the heavy lifting.
optical waveguides just have minimum bend radii that only go so small
>Light has a much much higher propagation velocity than an electronic signal in wire, so clocking would be easier to do, and can be done faster.
index of refraction of glass is ~1.5, and ~3.5 for silicon, so I'm not sure what you're basing this off of.
>>53668668
A 3570k would spank the 8350
>>53668617
No, it's all about precision.
Which determines how small you can make the transistors.
The main factors determining precision are:
- wavelength.
- refractive index of the medium.
- numerical aperture (aka 'f-number') of the optical system.
The bigger the numerical aperture (smaller f-number), the sharper a lens can be.
This is becasue of an effect called diffraction, which happens any time light passes through an opening.
But lenses with wide apertures also require a lot of corrections to achieve that sharpness, which generally means combining many lens elements.
So you end up with a lens that is both thick and tall, containing a whole lot of precision cut glass.
>>53668563
Wait, that guy works for Google? I remember reading about that guy's computer back when I was in 8th grade in 2007. That really wasn't what I was thinking of though as the CPU for that was made entirely out of discreet components. I was wondering if technology has advanced to the point where a DIY computer more advanced than something from the early 70s was now possible.
>>53668660
Thanks, I'm probably going to spend the rest of my night reading about that now.
>>53668182
This. Most engineers estimated that our current FINFET tech could only go to 150 nm, but we've stretched it to 14nm some with some serious hacking.
Instead of investing in completely new technology, we invested in getting our current tech better. Now it's starting to bite us in the ass
>>53668888
[witnessed]
>>53668398
>>53668476
Another potential candidate is directed self assembly.
Basically mixing chemicals that form patterns by themselves. (and with a little help from an EUV machine).
The potential is enormous, but it's still very early days.
>>53668984
That sounds pretty sci fi
>>53668888
They've invested a lot in EUV.
It just took the engineers a lot longer than expected, and all but one company gave up.
>>53668841
I know it will but lets keep things similar.
>>53669074
Well, to be honest, I'd say that a skylake vs 8350 comparison is more relevant because Intel isn't really selling anything from 2012 anymore, but AMD is. I could walk into my local Micro Center and pick up a 8350 or a Skylake cpu, but I can't buy an Ivy Bridge anymore.
Compare old parts with old parts when they're both off the market.
>>53668839
>only work in optical-cavity sized devices as far as I know, so nowhere near feasible for most needs at present
Alright, i'll give you that
>similarly, I think only the large-scale clock distribution has been seriously discussed, with regional electrical distribution trees still doing most of the heavy lifting.
optical waveguides just have minimum bend radii that only go so small
>index of refraction of glass is ~1.5, and ~3.5 for silicon, so I'm not sure what you're basing this off of.
At these scales optics dont behave like they do on the macro scale like we're familiar with. Quantum effects come into play, and light acts really fucking weird. Manipulating light on this scale would ultimately have to be done entirely with this in mind.
As for the index of refraction, we're less refracting and reflecting light and more along the lines of phasing it around corners and through thin walls. This could best be done in some sort of localized vacuum, which, considering we're switching to EUV anyways, would not be that hard to establish then seal the clock pathways in hard vacuum.
Of course I could be entirely fucking wrong about it, I'm just going on the bits and pieces of an old article involving the subject i remember reading a while ago.
>>53669224
Hmmm, yeah I got you on that point.
>>53668984
they should use a million-monkeys method
throw shit together
get a super high camera
use image recognition to detect anything that makes a usable chip
>>53666999
Senpai
Those trips
>>53666211
This is pretty much 100% complete truth.
Agile? ok
>>53665918
Well, they're still waiting on fab machines for lab 42, last I saw they had most of their core lithography stuff, but still have issues with the reflect masks for the ultra UV beams.
It's more that we're falling so far behind the ITRS in terms of next gen fabrication, that there will be a lag period.
As an aside, characterising quantum as calculations per second as OP image does is fairly nonsensical.
>>53668503
This is pretty much it.
Single core speeds aren't going to increase much in the foreseeable future. The cost of having ridiculous amounts of parallel power available will continue to drop, though, but we're still not very good at utilizing it for general purpose tasks.
