Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/what-happens-when-shrink-ends.17649/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

What Happens When Shrink Ends?

Arthur Hanson

Well-known member
Shrink will be coming to an end shortly. Will layering or stacking be next? Will integrating memory(automata)be next? Speed increases? Totally new architectures? What companies will prosper? What companies and ecosystems will win when the current ecosystem morphs? Any thoughts on this and especially TSM's plans would be appreciated. Will the entire ecosystem change with 5G and thin client computing? Will Rust become a major game changer? Thanks
 
Shrink will be coming to an end shortly. Will layering or stacking be next? Will integrating memory(automata)be next? Speed increases? Totally new architectures? What companies will prosper? What companies and ecosystems will win when the current ecosystem morphs? Any thoughts on this and especially TSM's plans would be appreciated. Will the entire ecosystem change with 5G and thin client computing? Will Rust become a major game changer? Thanks

I think the answer is all of the above but I would not say shrink is ending. That has been said many times before throughout my 40 year career and has been proven wrong every time. The semiconductor industry is filled with brilliant people who solve very complex problems. Do not bet against us.

The big game changer is AI. Not only will AI drive leading edge wafer sales, AI will also help the semiconductor industry address challenges in design and manufacture, absolutely.
 
Conventional shrinking has a path with over decade to go. So no shrinking is not "ending shortly" even if the industry doesn't find a cheat or beyond CMOS technology to get even smaller. As for layering multiple device layers that is a method to further increase the rate at which density scales. NAND did it, logic is going to do it with CFET soonish per IMEC roadmaps, and DRAM is looking for ways to do it. As for PPW improvements just look at how much gas was in the tank for intel 14 and 10nm when those nodes needed extended lives. If the industry ever reached a device scaling wall, then I'm sure the industry will be able to milk at least 20% more from their latest nodes. Probably more if all hands went to improving that metric as opposed to working on denser devices. As for integrating memory that is already happening for workloads where that is useful. With SRAM scaling coming to a halt until things like 2D mats, CFET, and BSPDNs become a thing; you might see a large amount of interest in leading edge companies directly bonding cache to compute dies (V-cache) or on package cache dies (broadwell, Sony handheld gaming systems, and RDNA3).
 
I think the answer is all of the above but I would not say shrink is ending. That has been said many times before throughout my 40 year career and has been proven wrong every time. The semiconductor industry is filled with brilliant people who solve very complex problems. Do not bet against us.

The big game changer is AI. Not only will AI drive leading edge wafer sales, AI will also help the semiconductor industry address challenges in design and manufacture, absolutely.
Do you see Rust being a major game changer or just another step to something even better?
 
Do you see Rust being a major game changer or just another step to something even better?
Huh? Computer languages which reinvent the wheel simply grease the wheels of busy-work replacing stuff that mostly exists. Universities teach programming language design so every new generation does that.

It has a few interesting features that may make the new busywork incrementally better than the old, but mostly the same game.

Julia is more interesting for new stuff than Rust with better handling of data and algorithm abstractions.

Meanwhile ML/AI is an exploding game changer mostly using Python and C++ (PyTorch) compiled into CUDA, and you can't get more ordinary than those these days. It is not the tools, it is what you do with them.
 
We will run into economic limits on shrink before physics limits.

In any case, look at the car industry... for a long time, from the early 1900s until the 1970s-1980s prefermance was about horsepower, but where it got harder to squeeze more horsepower out of an engine and consumer preferences changed to things like safety and fuel efficiency. Same thing will happen in computing, where CPU becomes less important and the integration of features like security, AI, and power efficiency become more important. This is already happening, but it will happen more so.
 
We will run into economic limits on shrink before physics limits.
Since that's how so many technologies are, I think this is a reasonable prediction.
In any case, look at the car industry... for a long time, from the early 1900s until the 1970s-1980s performance was about horsepower, but where it got harder to squeeze more horsepower out of an engine and consumer preferences changed to things like safety and fuel efficiency. Same thing will happen in computing, where CPU becomes less important and the integration of features like security, AI, and power efficiency become more important. This is already happening, but it will happen more so.
You've lost me with the automotive analogy. Many modern vehicles, both cars and light trucks, have an absurd amount of horsepower and performance. For example, the Tesla Model 3 dual-motor performance model, arguably a semi-frumpy family sedan, runs the 1/4 mile in 11.6 sec with a trap speed of 115mph. That's nutty performance for any general purpose street use. A BMW M340i x-drive, another mid-to-high end 4-door sedan, but gasoline powered, does the 1/4 mile in 12.2/112mph. Again, silly performance for mid-range family sedans, and equals or exceeds that of high-end Corvettes and Porsche 911s from less than 20 years ago. Current performance model Corvettes and Porsches can do the quarter mile in less than 11 seconds, which when I was growing up I think required a parachute for stopping at the local drag strip.

Apparently horsepower, even silly horsepower, still sells.

And now we have 16 core client PC CPUs. And the Apple M2 Ultra has about 200GB/sec of memory bandwidth. I doubt very much that computing vendors at any level are going to focus on metrics (like security) that few non-technical buyers can understand, when performance attributes like core counts and clock speeds have figures of merit anyone easily understands, and are easy to sell. IMO, the practical chips will be the ones designed for in-house use in data centers, like Amazon's Graviton chips.
 
Apparently horsepower, even silly horsepower, still sells.
Cruising the autobahn at 200 klicks consumes around 60kW (80hp) sustained in a streamlined sedan. The M340i of course is built in the USA, it would be insane to sustain autobahn speeds in that shape, but they put similar engines in it. Tesla's "cheap" upcoming vehicle looks likely to be around 110 to 120 kW peak, much more sedate than current models.
 
We will run into economic limits on shrink before physics limits.
So long as we beat the physics, we will have economic solutions. Moore's law was originally stated about costs per package (economics) and the economics continues to incentivize solving the physics. It may not make the devices smaller, but it will make them cheaper. Just like Dennard scaling, Moore's law borrows shrinking only while convenient, then the money will flow into other levers.
 
Cruising the autobahn at 200 klicks consumes around 60kW (80hp) sustained in a streamlined sedan. The M340i of course is built in the USA, it would be insane to sustain autobahn speeds in that shape, but they put similar engines in it. Tesla's "cheap" upcoming vehicle looks likely to be around 110 to 120 kW peak, much more sedate than current models.
Not that it's all that relevant, but the M340i is built in Germany, India, and I thought I read Mexico in a recent BMW CCA magazine, except for the x-drive version I mentioned, which I think is only built in Germany. While BMW largest plant by unit volume in the world is in South Carolina, I think it only builds SUVs (excuse me, SAVs).
 
Since that's how so many technologies are, I think this is a reasonable prediction.

You've lost me with the automotive analogy. Many modern vehicles, both cars and light trucks, have an absurd amount of horsepower and performance. For example, the Tesla Model 3 dual-motor performance model, arguably a semi-frumpy family sedan, runs the 1/4 mile in 11.6 sec with a trap speed of 115mph. That's nutty performance for any general purpose street use. A BMW M340i x-drive, another mid-to-high end 4-door sedan, but gasoline powered, does the 1/4 mile in 12.2/112mph. Again, silly performance for mid-range family sedans, and equals or exceeds that of high-end Corvettes and Porsche 911s from less than 20 years ago. Current performance model Corvettes and Porsches can do the quarter mile in less than 11 seconds, which when I was growing up I think required a parachute for stopping at the local drag strip.

Apparently horsepower, even silly horsepower, still sells.

And now we have 16 core client PC CPUs. And the Apple M2 Ultra has about 200GB/sec of memory bandwidth. I doubt very much that computing vendors at any level are going to focus on metrics (like security) that few non-technical buyers can understand, when performance attributes like core counts and clock speeds have figures of merit anyone easily understands, and are easy to sell. IMO, the practical chips will be the ones designed for in-house use in data centers, like Amazon's Graviton chips.
Recent increases in horsepower required a complete change in vehicle architecture from ICE to EV. From the 70s to early 2000s engine size grew smaller and horsepower was flat.

In 30 years maybe there will be a complete change in computing architecture to quantum, until then performance gains will be more incremental for CPU. We will continue to see performance gains in things like GPU, memory bandwidth, and power efficiency.
 
Recent increases in horsepower required a complete change in vehicle architecture from ICE to EV. From the 70s to early 2000s engine size grew smaller and horsepower was flat.
Huh? Not true at all. What country are you referring to?

In 30 years maybe there will be a complete change in computing architecture to quantum, until then performance gains will be more incremental for CPU. We will continue to see performance gains in things like GPU, memory bandwidth, and power efficiency.
That's why SoCs use specialized offloads for everything from video codecs to compression, encryption, communications protocols, whatever. There's lots of potential in whatever. You get 10-100x in performance versus software on a CPU. Even Intel is going to on-die offload blocks in Sapphire Rapids, which historically Intel has despised (especially Gelsinger, ironically). Heterogeneous computing, compute in memory, compute in storage, and compute in networks are also viable enhancements. There's so much that can be done architecturally that waiting for quantum to get practical isn't necessary.

As for quantum computing, it is at least decades away for replacing general purpose CPUs. At my age I probably won't be around to see it.
 
In some ways, shrinking ended a long ago. There were the good old days when one could simply rely on a Dennard scaling. Then diffraction kicks in, used OPC to mitigate...etc etc.
Then clock speed limits kick in, strained silicon is introduced in Fab and wide superscalar is introduced from the design side. This is how shrinking works in the real world. Cannot really avoid diminishing returns on shrink.

So, your question is more like 'What happens if shrinking gets more and more difficult'. Now we're relying on new packaging to save new fab capacities, DTCO. And now traditional customers have become chip designers as well, to save precious transistors. If shrinking stops entirely(no more cost savings by manufacturing), then, in my opinion, companies who rule workloads, will design chips on their own. More like a world of ASIC for each customer mabe?
 
Huh? Not true at all. What country are you referring to?


That's why SoCs use specialized offloads for everything from video codecs to compression, encryption, communications protocols, whatever. There's lots of potential in whatever. You get 10-100x in performance versus software on a CPU. Even Intel is going to on-die offload blocks in Sapphire Rapids, which historically Intel has despised (especially Gelsinger, ironically). Heterogeneous computing, compute in memory, compute in storage, and compute in networks are also viable enhancements. There's so much that can be done architecturally that waiting for quantum to get practical isn't necessary.

As for quantum computing, it is at least decades away for replacing general purpose CPUs. At my age I probably won't be around to see it.

My dates are a bit off, horsepower dropped through the 70s and did not recover until the 90s. Engine size has been stagnant since the 70s. Unit performance per unit of energy has steadily increased.

100% agree with what you say about architectural advancements and hetrogenous computing, this is the path forward for the next couple of decades in my opinion, and this is what I mean by raw CPU power becoming less important while other features become more important.
 
So, your question is more like 'What happens if shrinking gets more and more difficult'. Now we're relying on new packaging to save new fab capacities, DTCO. And now traditional customers have become chip designers as well, to save precious transistors. If shrinking stops entirely(no more cost savings by manufacturing), then, in my opinion, companies who rule workloads, will design chips on their own. More like a world of ASIC for each customer mabe?
Moore's law was always about economics. Read the original, and keep in mind that in the mid-60s a package often contained multiple "chiplets" since monolithic integration was not yet dominant. Dennard's Law had not yet been formulated (though Carver Mead had a model out through 1990 at roughly the same time, it was not reduced to pithy formulas like "MOSFET scales equal power per unit area as it shrinks") so simple device scaling was not the start of Moore's law, and not the end of it either.

So the "is Moore's Law alive" really becomes: do useful functions keep getting cheaper, on an exponential of useful rate? It helps that the economic driver is trillions of dollars by now, meaning almost anything will get sucked in if it helps. So the answer is yes. Everything from better production quality (think bigger chips with fewer defects) to 3D to packaging, and yes even continuing innovation in device geometry, backside wiring, liquid cooling, new 2D materials, etc.

Moore's economics have mellowed and complexified but are far from dead.
 
I believe a breakthrough in Asynchronous designs will create a massive leap forward in compute efficiency and effectiveness - above and beyond shink.
 
I believe a breakthrough in Asynchronous designs will create a massive leap forward in compute efficiency and effectiveness - above and beyond shink.
Do you mean a breakthrough in design tools? Because there have been billions spent in CAD tool development for global clocks, which puts asynch circuits at a big disadvantage. Although I suppose chiplets make this problem easier too.
 
Do you mean a breakthrough in design tools? Because there have been billions spent in CAD tool development for global clocks, which puts asynch circuits at a big disadvantage. Although I suppose chiplets make this problem easier too.
I mean a breakthrough in design and architecture approaches.

Semiconductor (logic) designs are based on a number of reasonable and proven assumptions (and approaches) that will eventually shown to be less efficient and less effective than a different approach. 70+ years of remarkable progress in a largely singular direction has created an industry-wide hubris that cannot consider that there may be alternative methods to accomplish the same processing with only a fraction of the resources required.

My bold assertion is that the entire industry will be in for a huge disruption when compute can be delivered without the current levels of power and complexity that it presently believes are required. And that this disruption will be caused in part through the workable application of assymetrical and parallel logic implementations.
 
I mean a breakthrough in design and architecture approaches.
It has already happened, you are just looking in the wrong place. GPUs and now ML and Inferencing, video conversion, signal processing, sensory fusion - if you look at an A15 at least 3/4 of the area is occupied by highly parallel data flow accelerators.

Innovations rarely replace a well established incumbent. Mostly they simply run around it and leave it where it was, in the rear view mirror.
 
Shrink will be coming to an end shortly.

I think shrinking continues for another decade or so, but the economic advantages (leverage) of shrinking have clearly wound down. While often ignored, Moore's Law was about the economics driven by technological advances. Fixed costs have exploded during the last decade and variable costs have gone up quite a bit too. When combined with yield challenges that are particularly significant for larger die sizes, economic advantages have waned at best. However, as Daniel notes, AI is opening new vistas for innovation as are architectural improvements, stacking and chiplets.
 
Back
Top