In the news recently was Intel's stock falling 20% because its 7nm processes are being delayed by 6+ months. The argument seems to be that Intel is already behind (since they're using 10nm processes while TSMC is on 7nm processes, soon to be 3nm processes), so it's only going to get worse for Intel. The entire chipmaking process isn't something I understand very well however, so I'm hoping someone can enlighten me.

Why exactly is it important to be on 3nm processes instead of 5nm or 7nm? This article says:

There are a few major underlying technologies that dictate the potency of any chip. The most fundamental rule of processors still holds true: The densest process nodes, provided they have decent power, performance, and area (PPA) characteristics, will often win the battle if paired with a solid microarchitecture.

I'm guessing this means that 3nm processes are "better" because for whatever reason I don't understand, it fundamentally improves the PPA characteristics. In other words, although you can improve while on 10nm processes, you'll never catch someone on 3nm processes - sort of like how a 40-year old can run faster with training, but will never catch someone who's 25 years old.

But if this is the case, then TSMC's processes are just the best in market, and they'll be leaving everyone else in the rear-view mirror. That hasn't happened though - the same article notes that Intel's processors still leads in certain categories, such as overclocking and gaming performance. Somehow Intel has managed to remain competitive with TSMC in spite of poorer technology. Even more surprisingly, SMIC manages to rake in billions of dollars in revenue in spite of using only 14nm processes. Perhaps one could say "therefore their revenues are going down", but they aren't, and that's in spite of Intel/TSMC/Samsung being on 10nm and smaller processes for years. Can someone explain what's going on?

I'm sure there are other applicable tags than just "Intel" and "fabrication", but I genuinely don't know what they are - if you do feel free to edit them into the question.

  • 2
    \$\begingroup\$ It's not all about the feature size. For example, MIPS' first commercial CPU was the R2000. At the time, Motorola and Intel had the best FABs and MIPS wasn't allowed to use them. Nor even 2nd best. The MIPS R2000 was instead designed for low-tech brokered FABing from 3rd tier (or worse) FABs. Old stuff. Ancient stuff. More than an order of magnitude worse, in fact. Yet they scared the daylights out of Intel and forced rapid progress at Intel on RISC. It's possible to do a lot more with the same die space. But you have to be creative. It's a new situation today. Not sure how much applies now. \$\endgroup\$ – jonk Aug 12 at 8:43
  • \$\begingroup\$ your runner analogy is bad, and thus excellent: I know quite a few 50 year olds who will leave me in the dust when competing in any running activity. They have 40 years of training. I don't. Intel has 50 years of experience in building high-performance CPUs and was a market leader for the majority of that time. Just making your gate more efficient (your lungs younger) doesn't make your design smarter (you better trained/skilled in running technique). \$\endgroup\$ – Marcus Müller Aug 12 at 8:46
  • 1
    \$\begingroup\$ I'm guessing this means that 3nm processes are "better" because for whatever reason I don't understand A smaller process node means that chips can be made that use less power but have increased performance. Compare a modern smartphone with a PC from about 15 years ago. 15 years ago we needed a big box with many chips, today we can squeeze the same performance in a hand-held device running on a battery. That would have been impossible without nano-meter size processes technologies. \$\endgroup\$ – Bimpelrekkie Aug 12 at 8:56
  • \$\begingroup\$ There's a whole lot of aspects involved in making the best CPU. From late 1990s to early 2000s, it was all about increasing the clock frequency, until that wasn't feasible any longer. Then they went 64 bit and started added multi cores instead. All of this happened in parallel with reducing transistor sizes, as well as more and more advanced ways of instruction pipelining, pre-fetch caching and branch prediction. \$\endgroup\$ – Lundin Aug 12 at 8:59
  • 3
    \$\begingroup\$ "The argument seems to be that Intel is already behind (since they're using 10nm processes while TSMC is on 7nm processes)" This is false. These numbers are not equivalent, and are basically useless marketing. In reality, Intel's 10 nm is actually denser than TSMC's 7 nm procede. You are right that Intel is falling behind (their 10 nm (equivalent to TSMC 7 nm) is barely working while TSMC is almost ready for thier 5 nm (equivalent to Intel 7 nm)), but it's not because their marketing number is bigger. \$\endgroup\$ – BrtH Aug 12 at 9:59

Somehow Intel has managed to remain competitive with TSMC in spite of poorer technology

Intel has a lot of experience in designing CPUs. A "stupid" design using the fastest technology can be worse (performance wise) than a "smart" design using somewhat older (and slower) technology.

TSMC only does chip-making they do not design, they are a foundry , so whatever chips are made by TSMC are not designed by them. The chips are designed by Qualcomm, Broadcomm and many other "fabless" semiconductor companies.

SMIC is also a foundry, just like TSMC. SMIC concentrates on older technology nodes that are "less interesting" to foundries like TSMC. A lot of chips are needed in mobile phones, cars etc. Not all of these need to be made in the latest-and-greatest fastest technology.

Realize that making a "simple" chip which can be made (or is already an existing design) in for example 14 nm (or much older, larger feature size technologies) is much, much cheaper than using the latest technology node. I mean: very often there is simply no need to use the newest technology node.

... up with TSMC if their processes are just the best?

Are they "the best"? What is "the best"? For a Soc powering the latest and greatest smartphone, indeed using the smallest manufacturing node might be essential.

But how about a chip used in a pocket calculator or microcontroller in an espresso machine? For these applications the chip needs to be cheap. A simple microcontroller, when bought in large quantities, can be had for a few cents per chip. I guarantee you that these chips will be made in an "ancient" technology, like 0.35 um CMOS (just my guess). The (at the time very expensive) equipment needed for making such chips has been written off long ago so you don't need to pay a premium price because expensive equipment is needed. That allows for chips that are extremely cheap. In many applications "cheap" is much better than "latest and greatest".

| improve this answer | |
  • \$\begingroup\$ The latest and greatest in MCUs is in the double nanometer digits. STM32H7 is 40nm and they do have their own fabs. The more cost-focused STM32F0 series is 180nm, Data taken from Wikipedia page on STM32. And those are not the very cheap MCUs mentioned in this answer, F0s start from something around 40-50 cents. \$\endgroup\$ – Jan Dorniak Aug 12 at 18:54
  • 3
    \$\begingroup\$ Also, one has to be careful when comparing process nodes. Intel's 10nm node is very close in density to TSMCs 7nm node - both processes have advantages in certain topologies, so sometimes Intel can do better with 10nm than TSMC with 7nm, and sometimes the inverse. Comparing nominal node size, Intel ostensibly seems like they're two nodes behind when they're really only one node behind, effectively. \$\endgroup\$ – J... Aug 12 at 19:50
  • \$\begingroup\$ TSMC is producing the direct competition to Intel CPUs, AMD Zen CPUs. And they are leaving Intel in the rear-view mirror. \$\endgroup\$ – Josef says Reinstate Monica Aug 13 at 7:39

A single number does not make a better product.

A better product does not automatically and immediately capture all the market.

A better process does not have infinite capacity.

A smaller process number in nanometers ("lambda") usually indicates that you can fit more transistors onto the chip, and they will individually consume less power. However, the power density has become a problem, including the "leakage" power from transistors that are not currently doing any work. As a result, modern high-density processors are made up of a large number of cores, and some of the cores have to be idle or run at less than maximum speed most of the time.

Using your larger number of transistors effectively to provide useful features and accelerate particular workloads is still the big challenge for designers.

But beyond the technical considerations, the whole field of product design and product marketing knows that the reasons humans buy products are more complicated and holistic. Much as people would like to reduce product quality to a single, emotionless, one-dimensional number, it doesn't work that way.

Intel's branding and marketing dominance has been formidable for the last forty or so years. Even if their products were spectacularly bad one year they would still have a large market share. Just as oil tanker takes a long time to turn round and an even longer time to sink, so does a dysfunctional mega-corporation.

Not all semiconductors need or benefit from the latest process. Here at Cirrus Logic (CRUS) we've only just got on to 22nm, but are doing very nicely because a large part of our chips are analogue: power amplifiers and similar.

| improve this answer | |
  • 1
    \$\begingroup\$ Many here are too young to remember the first Pentium 90s, which ran hot and had errors when doing floating point more than 6 decimal places deep. A quick investigation found a carry-over connection was missing. It could have been a death-nell for Intel, but they moved heaven and Earth to get fixed versions to market in the same year. It was soon forgotten about and Intel marched ahead. \$\endgroup\$ – user105652 Aug 12 at 9:51
  • \$\begingroup\$ @VTNCaGNtdDVNalUy I had one of those. Intel stonewalled for a very long time but eventually, very grudgingly, did the right thing. The box makers were caught in the middle. IIRC they denied it was likely to affect anyone, then offered a software fix that destroyed FPU performance (shades of 737 Max). \$\endgroup\$ – Spehro Pefhany Aug 12 at 14:49
  • \$\begingroup\$ @VTNCaGNtdDVNalUy, heck, many people here won't remember the Pentium 4 era: for almost six years, AMD's chips were faster, cheaper, and lower-power than their Intel counterparts, and Intel still maintained market dominance. \$\endgroup\$ – Mark Aug 12 at 23:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.