4 ms·

You mean how many chip cycles he needs, right? How many he's got we all know: maybe 2. Then CMOS scaling is over. (Not that things will completely stop evolving

by _yosefk 10y ago

You mean how many chip cycles he needs, right? How many he's got we all know: maybe 2. Then CMOS scaling is over. (Not that things will completely stop evolving, but they'll surely slow down.)

vessenes 10y ago
In 20 years? I think he's got at least 4, maybe 5. Of course, I don't know how big those evolutions will be in terms of advances, but I'm just a silicon dilettante, nothing like an expert.
- Bartweiss 10y ago
  I think the dispute here is that there are two different standards in play. He has time for 4-5 cycles, but Intel only has concrete plans for ~2 cycles. After that the existing die structures can't be meaningfully reduced anymore, and everything is up in the air. Maybe we get some wild, big-but-fast optical chips, maybe we get a breakdown of Moore's law. Of course, 3d chips look fairly achievable and might eke out a few more cycles, but things will get interesting sometime before that 20 year deadline.
  - dogma1138 10y ago
    I don't think Intel plans for more than 2 fabrication cycles regardless of some gloom and doom omg quantum physics go away plox prophecy. 14nm FinFet GPU's are only coming in now, and they'll stay on 14nm for at least 3-4 years. 10nm GPU's will probably come out in 2019-2020, 7nm GPU's will come out in 2025 or so, and 5nm GPU's will come out in 2030, that's already ~15 years and considering that everyone is predicting that we'll have to stick to processes longer as the small we get the more time it takes to actually optimize them for good yields and performance you probably have 3-5 years there too. So even if we do not adopt any major process changes like switching to graphene, solid state photonics to replace the metal pathways and even optical transistors or w/e else people are cooking up to enable us to continue to improve our integrated circuits he can still be correct about the time frame. Also regardless of what happens or not he was talking about true VR immersion this means that allot has to change other than just how good and convincing the graphics are this also includes learning to develop VR applications and as VR is a very new media that alone can take 20 years.
    - Bartweiss 10y ago
      This is a good summary, thanks. And yeah, I only meant to speak to the argument in the comments over how many cycles we can "fit" in the next 20 years. I think a lot of VR's biggest hurdles for that time aren't dependent on processing power. Focus and resolution issues will side-step some of it (seriously, our eyes are terrible outside of their central focus region!) because there's simply no need to render everything flawlessly as on a TV. Beyond that, there are going to be all kinds of design hurdles. Right now, we have nausea-inducing roller coaster rides and people running into walls because we can't sync up the (already impressive) visual experience with inner-ear confusion and physical boundaries. Just working out what the most compelling designs are for our current VR tech is the work of years, and that's without doing it on the treadmill of ever-improving tech!
dogma1138 10y ago
well TSV have evolved greatly which means that 3D dies can come soon. 3D transistors already expanded the limit of how small parts can get before physics hits you in the face with a baseball bat. On top of that memristors and graphene could open a whole new world for chip designers. If we have only 2 cycles worth of process left we are doomed :)
- Certhas 10y ago
  > 3D transistors already expanded the limit of how small parts can get before physics hits you in the face with a baseball bat. Yes, but we are just starting 10nm now. Go two more nodes (if I understand correctly that is meant here by cycles) and you are at 5nm. https://en.wikipedia.org/wiki/5_nanometer https://en.wikipedia.org/wiki/5_nanometer to put that into perspective: "An Australian team announced that they fabricated a single functional transistor out of 7 atoms that measured 4 nm in length." Maybe we can make our way down to single atom transistors eventually, but who knows how long it will take to make that commercially viable.
  - hexane360 10y ago
    We already have a single atom transistor (https://en.wikipedia.org/wiki/Single-atom_transistor https://en.wikipedia.org/wiki/Single-atom_transistor), but you're right that it's far from commercially viable.

by _yosefk 10y ago

You mean how many chip cycles he needs, right? How many he's got we all know: maybe 2. Then CMOS scaling is over. (Not that things will completely stop evolving, but they'll surely slow down.)

vessenes 10y ago

In 20 years? I think he's got at least 4, maybe 5. Of course, I don't know how big those evolutions will be in terms of advances, but I'm just a silicon dilettante, nothing like an expert.

Bartweiss 10y ago
I think the dispute here is that there are two different standards in play. He has time for 4-5 cycles, but Intel only has concrete plans for ~2 cycles. After that the existing die structures can't be meaningfully reduced anymore, and everything is up in the air. Maybe we get some wild, big-but-fast optical chips, maybe we get a breakdown of Moore's law. Of course, 3d chips look fairly achievable and might eke out a few more cycles, but things will get interesting sometime before that 20 year deadline.
- dogma1138 10y ago
  I don't think Intel plans for more than 2 fabrication cycles regardless of some gloom and doom omg quantum physics go away plox prophecy. 14nm FinFet GPU's are only coming in now, and they'll stay on 14nm for at least 3-4 years. 10nm GPU's will probably come out in 2019-2020, 7nm GPU's will come out in 2025 or so, and 5nm GPU's will come out in 2030, that's already ~15 years and considering that everyone is predicting that we'll have to stick to processes longer as the small we get the more time it takes to actually optimize them for good yields and performance you probably have 3-5 years there too. So even if we do not adopt any major process changes like switching to graphene, solid state photonics to replace the metal pathways and even optical transistors or w/e else people are cooking up to enable us to continue to improve our integrated circuits he can still be correct about the time frame. Also regardless of what happens or not he was talking about true VR immersion this means that allot has to change other than just how good and convincing the graphics are this also includes learning to develop VR applications and as VR is a very new media that alone can take 20 years.
  - Bartweiss 10y ago
    This is a good summary, thanks. And yeah, I only meant to speak to the argument in the comments over how many cycles we can "fit" in the next 20 years. I think a lot of VR's biggest hurdles for that time aren't dependent on processing power. Focus and resolution issues will side-step some of it (seriously, our eyes are terrible outside of their central focus region!) because there's simply no need to render everything flawlessly as on a TV. Beyond that, there are going to be all kinds of design hurdles. Right now, we have nausea-inducing roller coaster rides and people running into walls because we can't sync up the (already impressive) visual experience with inner-ear confusion and physical boundaries. Just working out what the most compelling designs are for our current VR tech is the work of years, and that's without doing it on the treadmill of ever-improving tech!

dogma1138 10y ago

well TSV have evolved greatly which means that 3D dies can come soon. 3D transistors already expanded the limit of how small parts can get before physics hits you in the face with a baseball bat. On top of that memristors and graphene could open a whole new world for chip designers. If we have only 2 cycles worth of process left we are doomed :)

Certhas 10y ago
> 3D transistors already expanded the limit of how small parts can get before physics hits you in the face with a baseball bat. Yes, but we are just starting 10nm now. Go two more nodes (if I understand correctly that is meant here by cycles) and you are at 5nm. https://en.wikipedia.org/wiki/5_nanometer https://en.wikipedia.org/wiki/5_nanometer to put that into perspective: "An Australian team announced that they fabricated a single functional transistor out of 7 atoms that measured 4 nm in length." Maybe we can make our way down to single atom transistors eventually, but who knows how long it will take to make that commercially viable.
- hexane360 10y ago
  We already have a single atom transistor (https://en.wikipedia.org/wiki/Single-atom_transistor https://en.wikipedia.org/wiki/Single-atom_transistor), but you're right that it's far from commercially viable.