3 ms·
The Next 5 Years of Chip Technology
- peterburkimsher 9y agotl;dr - Existing technology will be pushed to the limits first. There are lots of new technologies currently being researched, but it's not obvious which will take over. New ideas will "scale over the next 10 years", not over the next 5 years. Moore's law is not ending because of physical limits. It's slowing down because we don't know which option to pursue. We are currently at 10 nm [1]. After that is 7 nm (2018), then 5 nm (2020), then 3 nm. FinFETs will be used at 5 nm. Another technology might come at 3 nm, or later. 3D NAND memory is currently stacked at 48 layers. It will be increased to 96, 128, maybe even 256 layers. The challenge is to keep thin wafers flat when stacking them on top of each other. DRAM will be around for a while. There are new ideas such as MRAM and XPoint, but DRAM is faster and lasts longer. Personal addition: Existing technology will get us 2TB SDXC cards. After that we'll need a new industry standard for larger cards. [1] https://en.wikipedia.org/wiki/10_nanometer https://en.wikipedia.org/wiki/10_nanometer
- jacquesm 9y ago> Existing technology will get us 2TB SDXC cards. Which is an incredible thing in and of itself.
- narrator 9y agoThat schedule is optimistic given recent trends. Intel has been stuck at 14nm process technology for almost 4 years now and everyone else has caught up.
- peterburkimsher 9y agoThe market's pretty harsh to those who don't catch up. Apple's already started making their own A-series chips. Intel will most likely get their act together soon, or pivot out of the market and move into services like IBM did. In "The Next 5 Years" though, I think Intel will still be around, and hopefully catch up with the silicon.
- scentoni 9y agoMoore's law is slowing because of the exponential increase in R&D and manufacturing costs to push to the next generation.
- astrodust 9y agoTransistor counts are still climbing. These used to be more strongly correlated with performance, but these days it's different. GPUs have seen exceptional gains generation to generation, ARM is absolutely on fire, and AMD just dropped a bomb with their Epyc server chips. Don't think it's dead yet. It just had a hangover from Intel getting drunk on success with the Core series of chips.
- deepnotderp 9y agoNo love for monolithic 3D? :/
- deepnotderp 9y agoSemi industry: "A 5x denser memory cell with no process changes is "too risky" " Also semi industry: "Yeah go ahead and pour half the periodic table with ~5nm layers each on die for MRAM"
- deepnotderp 9y agoAlso this fanatic focus on the FEOL needs to stop, ~80% of the power consumption, even in an OoO CPU, is in the BEOL.
- mjevans 9y agoCould you elaborate about FEOL and BEOL? Do those correspond to the most base layers of structure creation and modification on a chip and then the larger structured interconnect layers that get added later respectively?
- deepnotderp 9y agoFEOL = Transistors, so all the fuss about FinFETs, GAA-FETs, carbon nanotube FETs, etc. BEOL = Interconnects. You know, the stuff that's RC quadratic and ~80% of the power consumption.
- snaky 9y agoMaybe the future of interconnects is mechanical after all. https://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-21.html https://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-...