Will Moore's law put an end to CPU and GPU segmentation?

Question

Let's take Intel for example. Intel has 2, 4, 6 and 8 core mainstream CPUs. The 2 core CPUs are cheap and the 8 core ones are expensive. That's how they segment their products and make money. However, as transistors shrink from 10 to 7 to 5 to 3 nano-meters, how much longer can will companies be able to segment their products this way?

AMD just released a 16 core mainstream CPU. I don't think Intel and AMD can keep relying on die shrinks and adding more cores, so how are they going to segment their products?

Answer 1

Names like 7nm, 10nm, 5nm, whatever are all pseudonyms but the mainstream media and wall street hasn't figured this out. The actual measurements of the transistors, gates, etc vary. The only aspect about this that's true is thing keep getting smaller, but even these shrinks will hit their limit. FYI, Intel's 10nm process is equivalent to Samsung's and TSMC's 7nm process. Even the actual measurements of the transistors and whatnot vary between one 14nm process to another. Intel's original 14nm process that was used for Skylake is not exactly the same as the 14nm++ process used for Coffee lake 1.



It looks like these companies will be able to pull off a few more die shrinks before quantum tunneling become unavoidable. TSMC screwed up their 20nm process but moved on quickly. Intel has been stubborn about trying to get their 4 year late 10nm process out and they've lost a lot of money on it. Intel is already beginning to build their so called "7nm" production facilities in Hillsboro Oregon (30 minute drive from Portland). These die shrinks will be more difficult, more expensive, and they'll take longer to pull off.



The solution is to make chipletts like AMD is doing with Zen, and what Intel plans to do with the Dedicated GPU they are going to release next year. They'll make small packages of multicore chips and higher end models will have more of these chipletts on the substrate. The 12-core and 16-core Zen 2 package have 2x chipletts and a larger 14nm I/O controller chip. This is how AMD also packages Epyc and Threadripper.



Intel is still doing with Coffee Lake what has always been done. They'll cut hundreds of chips from a wafer, and the most functional ones are destined to become a Xeon or an i7. The ones that leak or need too much voltage to remain stable will become lower end models like a Core i3 or Celeron. This is where that whole "Silicon Lottery" thing comes from.



"AMD just released a 16 core mainstream CPU."

It's not out yet and the 16-core part won't be out until September. At best, AMD has announced it, not released it.

Answer 2

For example, what AMD is doing with its CPUs to obtain so many cores is instead of making a bunch of monolithic dies with either 2,4,6, or 8, etc. cores on them. They are making smaller dies, called chiplets, with a maximum 4 cores on them, and then just adding a bunch of them together into one package. So a 4 core processor might only use one chiplet, while an 8 core would use 2 of them. Anything in between, let's say a 6-core would be an 8-core with 2 out of the 8 disabled. This way if they get a partial failure on one of the chiplet dies, and only 2 cores out of the 4 are working perfectly, they don't have to throw the whole thing away, and they can continue to use those 2 good cores, and you'll get yourself a 6 core in a two-chiplet design. It's not only tremendously cost-effective, but it also lets them get higher performance out of the good cores, as the bad cores will be disabled, and they won't be receiving any power, while the good cores will get all of the power to themselves, thus achieving higher single-core performance, while staying within the same power limits.



Also node shrinks are now getting more efficient without requiring an actual node shrink. For example, Intel has been stuck on 14nm for 5 years now, so what it's been doing is improving the same 14nm node over and over again. The first improvement was called 14nm+, the next one was 14nm++, 14nm+++, etc. Slight improvements in packing transistors together better, etc. That's probably the way they're going to be working from now on, more improvements within the nodes from generation to generation, rather than an actual new node every generation.

Answer 3

Not necessarily. While CPU's usually have one die, AMD is using two with space for a third die on a single CPU package. Also there is nothing stopping the use of multiple CPU's as in servers or designing larger sockets for a single package containing multiple cpus inside.

Answer 4

Moore's "law" isn't a law.



CPU's and GPUs are hitting the limits at how small they can be shrunk. Intel's been working on the 10 nm chips for ages now.



Also, adding cores isn't really making computers faster in many cases. Programs have to be specifically designed to use more than one CPU core. Games and productivity programs are getting better at using more than one core, but there's always a limit to how many cores would be worth it to support.

Answer 5

"Core" in CPUs is not precisely defined. AMDs exact definition differs from Intel.

Further, number of cores does not define what occurs in a core. In other words, there are two core CPUs stronger than four core ones when the cores are not the same. This is the same issue as frequency of clock. What occurs in one cycle is critical.

Both Intel and AMD can turn cores on and off and set and lock or unlock frequencies and turbo.

There are limits to the die shrinking and new technology is being experimented with that is totally different.



AMD - Threadripper 2990WX 3 GHz CPU cores 32 The number of threads 64 Passmark 23252 and US$1617

250 Watt thermal design, 68C max temp 0.012 micron

Intel Core i9-9980XE @ 3.00GHz CPU cores 18 The number of threads 36 Passmark 29473 and US$1985

165 watt thermal design 84C max temp

Intel Core i9-9960X @ 3.10GHz CPU cores 16 The number of threads 32 Passmark 28735 and US$1700

165 watt thermal design 85C max temp

There are server Intel Xeon with more cores. Xeon 8176 CPU cores 28 The number of threads 56

Intel is 0.014 micron



Moore's law is more of an observation.

https://gist.github.com/emartin59/0345adc1a60ad58433bb9b24113f490b



As the circuit lines and spaces shrink they start getting physical issues

AMD is moving to 7nm and Intel is releasing their 10nm and then to 7nm

https://wccftech.com/intel-10nm-ice-lake-sunny-cove-14nm-comet-lake-amd-ryzen-3000-cpu-z-benchmark-leak/

They're up to 100 million transistors per sq mm.