Apple chip partner TSMC declares Moore's Law is not dead
Moore's Law, the name given to the principle that the speed and capability of computers is expected to double every two years due to advances in microchip technology, is not dead according to Apple A-series chip foundry TSMC, which reasons it is very much still relevant today for a number of reasons.
First outlined in a paper by Intel CEO Gordon Moore in 1965, Moore's Law is a prediction that the number of transistors in a dense integrated circuit will double roughly every year, before revising it to every two years in 1975. While the rule has been relatively accurate over the decades, difficulties with production has led to some believing it isn't sustainable.
Over the years, the cost of a die shrink has gone up in terms of researching techniques, the materials required for the task, and the lower yields of smaller processes, to the point where physics is making it extremely hard to accomplish. The difficulties led to Intel effectively giving up on keeping to Moore's Law, switching the cadence of its chip development from "Tick-Tock" to "Tick-Tock-Tock."
According to a blog post by TSMC, the firm behind the Apple-designed A-series chips used in iOS devices, Moore's Law "is not dead." Head of global marketing Godfrey Cheng reasons industry observers believing in the demise of the Law as taking the rule to mean the performance of the chip must double.
"Since the 2000's, compute performance has largely increased not through the improvement of the transistor clock speed but rather through both silicon architecture innovation and the threading or parallelization of computing workloads," Cheng writes. In that case, performance boosts are not from clock speed boosts but rather by "throwing more transistors at a compute problem," and that Moore's Law is about density of transistors.
For the "elephant in the room" over the death of the Law over die shrinks, Cheng references the announced N5P node for 2021 that enhances the existing N5 node, which uses a 5-nanometer process while providing 7% higher performance compared to N5.
For reference, the current A-series chips use a 7-nanometer process, but Apple is believed to be exploring the use of a 5-nanometer version for the 2020 iPhone's "A14" chip.
An example of TSMC's Chip on Wafer on Substrate technology, using the world's largest silicon interposer with space for two 600mm-squared processors and 8 HBM memory devices.
System-level density is also a factor in improving performance, with limitations in memory caching starting to be a problem. Locating memory physically closer to the cores offers improved latency and a higher bandwidth of data for processing, which will aid fields like AI and autonomous vehicles.
"As a car drives down the roads at highway speeds, every millisecond counts for safety," Cheng adds. "Locating memory close to the edge processing core is vital to reduce latency."
TSMC's advanced packaging techniques offering "tight integration of logic cores with memory" is also cited, with the line between a semiconductor and a system cited as "blurry as the new advanced packaging techniques are silicon wafer based." These techniques include the use of complete systems with a silicon-based interposer or fan-out-based chiplet integration, the ability to stack chips on wafers or wafers on top of wafers prior to integration, all to increase density.
Cheng moves on to reiterate "Moore's Law is about increasing density. Beyond the system level density achieved through advanced packaging, TSMC will continue to grow density at the transistor level. There are many paths available to TSMC for future transistor density improvements."
These improvements can include the use of transistors made of "two-dimensional materials instead of silicon as the channel," described as "we are raiding the periodic table." The stacking of multiple layers of transistors in a "Monolithic 3D Integrated Circuit" could allow a CPU to be placed on aGPU with memory layers sandwiched between.
"Moore's law is not dead," insists Cheng, as "there are many ways to continue to increase density."
First outlined in a paper by Intel CEO Gordon Moore in 1965, Moore's Law is a prediction that the number of transistors in a dense integrated circuit will double roughly every year, before revising it to every two years in 1975. While the rule has been relatively accurate over the decades, difficulties with production has led to some believing it isn't sustainable.
Over the years, the cost of a die shrink has gone up in terms of researching techniques, the materials required for the task, and the lower yields of smaller processes, to the point where physics is making it extremely hard to accomplish. The difficulties led to Intel effectively giving up on keeping to Moore's Law, switching the cadence of its chip development from "Tick-Tock" to "Tick-Tock-Tock."
According to a blog post by TSMC, the firm behind the Apple-designed A-series chips used in iOS devices, Moore's Law "is not dead." Head of global marketing Godfrey Cheng reasons industry observers believing in the demise of the Law as taking the rule to mean the performance of the chip must double.
"Since the 2000's, compute performance has largely increased not through the improvement of the transistor clock speed but rather through both silicon architecture innovation and the threading or parallelization of computing workloads," Cheng writes. In that case, performance boosts are not from clock speed boosts but rather by "throwing more transistors at a compute problem," and that Moore's Law is about density of transistors.
For the "elephant in the room" over the death of the Law over die shrinks, Cheng references the announced N5P node for 2021 that enhances the existing N5 node, which uses a 5-nanometer process while providing 7% higher performance compared to N5.
For reference, the current A-series chips use a 7-nanometer process, but Apple is believed to be exploring the use of a 5-nanometer version for the 2020 iPhone's "A14" chip.
An example of TSMC's Chip on Wafer on Substrate technology, using the world's largest silicon interposer with space for two 600mm-squared processors and 8 HBM memory devices.
System-level density is also a factor in improving performance, with limitations in memory caching starting to be a problem. Locating memory physically closer to the cores offers improved latency and a higher bandwidth of data for processing, which will aid fields like AI and autonomous vehicles.
"As a car drives down the roads at highway speeds, every millisecond counts for safety," Cheng adds. "Locating memory close to the edge processing core is vital to reduce latency."
TSMC's advanced packaging techniques offering "tight integration of logic cores with memory" is also cited, with the line between a semiconductor and a system cited as "blurry as the new advanced packaging techniques are silicon wafer based." These techniques include the use of complete systems with a silicon-based interposer or fan-out-based chiplet integration, the ability to stack chips on wafers or wafers on top of wafers prior to integration, all to increase density.
Cheng moves on to reiterate "Moore's Law is about increasing density. Beyond the system level density achieved through advanced packaging, TSMC will continue to grow density at the transistor level. There are many paths available to TSMC for future transistor density improvements."
These improvements can include the use of transistors made of "two-dimensional materials instead of silicon as the channel," described as "we are raiding the periodic table." The stacking of multiple layers of transistors in a "Monolithic 3D Integrated Circuit" could allow a CPU to be placed on aGPU with memory layers sandwiched between.
"Moore's law is not dead," insists Cheng, as "there are many ways to continue to increase density."
Comments
It’s not the same as Moore’s Law, but will achieve performance enhancements for years. Processors will be more modular and stackable.
Getting all the different parts (cpu, gpu, memory, wireless, etc) to integrate smoothly is going to be a challenge, but getting rid of the bus between them is going to be huge. Power management is going to be super efficient but controlling temperature is going to be crazy complicated.
I’m still impatiently waiting on biological computers though (that I read about 20 years ago) and it looks like I’ll continue to do so.
Come on people, my brain is in dire need of a memory upgrade... ; )
I get what he’s saying though... processors (etc) will continue to grow in power, even if Moore’s Law is dead (dying).
http://browser.geekbench.com/mac-benchmarks
https://browser.geekbench.com/ios-benchmarks
Macs have increased about 10x in 10 years as opposed to 32x if they doubled every 2 years. It's closer to doubling every 3 years just now, although some gains go into the integrated graphics chips.
It's going to be very hard to improve going forward. The marketing rep from TSMC talks about density but that's not all there is to it:
https://www.tsmc.com/english/newsEvents/blog_article_20190814.htm
It's possible to stack chips in layers, which can increase density per volume but not necessarily power efficiency. The key factor has always been performance-per-watt. A mobile chip can be stacked to 32 layers but if it uses 100W of power, it can't fit in a smartphone.
The following site says that shrinking won't be economically viable after 2021, I guess that will be 2-3nm, which will have the usual production cycle until 2024:
https://arstechnica.com/gadgets/2016/07/itrs-roadmap-2021-moores-law/
Beyond that, it will all be about stacking chips and cooling. But I don't think it will matter, another 4x over where we are now would mean a 32-core Macbook Pro with a 12TFLOP GPU. The demand for more performance won't be there any more outside of a few specialist fields. For those fields of work, 3D stacking and higher power draw will work just fine and can be hosted in a server to bring cloud computing costs down.
What would make an improvement is if the heat that a CPU generates could be turned back into power for the CPU rather than trying to just discard it or even better, use materials that don't heat up as much. We might see practical uses of quantum computing, even if they could be used as coprocessors. Maybe there will be advances in optical computing. These are things that have been and will be researched for years before we see anything practical.
I don't see an issue with Moore's Law coming to an end, the needs of the end users will be satisfied.
CPU and GPU power have, for past 5-10 years increasingly moved past the requirements of the average user typing letters, browsing the web, etc...
In fact, even the 13 year old laptop I use for my financial stuff runs just fine. And, Apple continues to sell products with CPUs that are nearly 3 years old (A10) and 5 years old (A8).
Yes, certain datacenters are hungry for more and more power. But I spent 20 years with one who profited very nicely from avoiding bleeding edge and always running 2nd generation equipment. And, the average consumer is seeing less and less benefit from stronger, faster CPUs because existing hardware already exceeds their needs.
Instead, for the average consumer, computing advances will come mostly from:
-- Communictions improved speed and power from 5G
-- Data storage and retrieval improvements as solid state devices proliferate.
-- Battery performance improvements
-- Tighter, more complete integration of devices - particularly in IoT.
Perhaps as an example: My friend's 3 year old piece of junk laptop from HP had become almost unusable as it got slower and slower (mostly from thrashing around dealing with multiple updates). Yesterday I swapped out her old 7200rpm HDD for a $40 SSD and it's back to running just fine.
Or is Cheng really this stupid?
Writing better code or more clever caching techniques isn’t EXACTLY the same as increasing transistor density. Ok. The two things aren’t related in any way beyond the end result on execution time.
This guy is effectively saying there really isn’t any difference between 1,000 slaves with shovels and a man driving a diesel powered strip mining machine - after all they both make holes in the ground.