For one thing, it makes it harder for me to understand when it will be necessary for new processors to use exotic materials. Since I've learned about the naming, and that processors have not actually reached the 5nm physical limit for silicon, it feels wrong to perpetuate it.
Apologies, I assumed it was mostly common knowledge among people who keep up with tech news. I began seeing stories about it at least a few years ago so I figured many of us had heard at this point.
Yes, anyone who cares about chips has known this for years.
It's the same as software engineers who don't know what SIMD is or think the CPU in their laptop or smartphone works like a MIPS they studied in college. When something like this is only one click away: https://en.wikichip.org/wiki/intel/microarchitectures/sunny_...
I've been curious, if this is true, why hasn't intel moved the goalpost for 7nm and declared victory, rather than the constant black eye of not having achieved that "marketing" number for years?
Disappointed that Intel had to change their naming because otherwise they'd lose out in marketing honestly. Corporations can't be truthful because a competitor won't be.
Back in the day everyone was concerned with the number of bit their gaming console had. The Nintendo 64 was the last to use bits as a primary marketing feature.
>The Nintendo 64 was the last to use bits as a primary marketing feature.
That's not true. There are some new "retro" consoles whose primary marketing feature is that they are 8 bit. For example the RES Plus' primary marking feature is that it plays 8 bit games.
Completely different situation. Those retro consoles are merely referencing the outdated but, for the old consoles, common classification, to indicate compatibility. They are not part of the race that once existed (i.e. you will hardly find a retro console claiming to be "better" because it can emulate 16bit consoles instead of 8bit ones).
Why? The average consumer doesn't care whether the semiconductor in their smartphones are 7nm or 5nm, they care about speed and power efficiency. Anyone who does care about feature size/density would know this and wouldn't be misled. Besides, they themselves don't care about feature sizes either, they only really care about other performance characteristics (eg. density, power consumption, switching speed), not some superficial metric like "nm".
Because you're advertising a product in a way that's deceitful, obviously. If I sell you a CPU and I say "yeah it's 4.5ghz", and it turns out it runs at 1.6ghz, it should be illegal to say "ghz is just a brand, it's marketing".
Semiconductor processes aren't a feature that's generally advertised to end users. I've never heard a mobile phone or computer manufacturer describing their devices as having a "5 nm processor", for example. The clients who are seeing these descriptions (like, purchasing managers in electronics manufacturers) are generally aware of these issues.
> The term "5 nm" is simply a commercial name for a generation of a certain size and its technology, and does not represent any geometry of the transistor.
And I learnt something new: it seems like MTr/mm² is the metric in use now. I wonder if marketing will find a way to co-opt that?
The problem with "density" is that designers use power optimized transistors for some things and speed optimized transistors for other things, on the same manufacturing process. Also, the quality of standard cell libraries matters, kinda like the quality of the compiler affects the SPEC score of a chip.
But EUV is a real thing and cleanly divides the haves and the have nots. While Intel 7 may have similar density to TSMC 7, there is a huge difference in the process and cost of making them. Intel will not go beyond that without EUV, and neither will anyone else.
I wonder which EUV steppers they bought from ASML, the 100 millions ones or the new half of billion ones (don't think those EUV steppers are available yet).
It seems, from random stuff I read on internet, that even with current ASML EUV steppers, there is a lot of transistor size reduction possible. And with the high-NA improvement, we can expect even more.
This is very good news for the 2/3 next generations of CPU/GPUs. We can "reasonably" expect "really" more, namely significantly faster/less power hungry of what we have now. But it seems we are missing foundries to scale for the worldwide demand, and many video game AAA studios (which will have the decency to provide builds for linux based OSes, meh). On the mobile front maybe some new radio modulation beyond 5G in order to properly use chips based on those new processes.
Sure but at some other point in time Intel had a denser process and TSMC caught up despite Intel also improving, so that fact alone doesn't give you a lot of certainty about the future.
> Or is there any objective metrics to determine which one is better, other than MTr/mm2?
There are many, many objective metrics. Many are interdependent, many are tunable or selectable within certain ranges, many are in opposition, most are highly secretive, and all of them influence he type of device being made, the circuit design, logic, in different ways.
Even the experts with all the data for a particular technology and with a particular device in mind can't necessarily tell at a glance whether a change would be worthwhile. And even the meaning of worthwhile varies for the end-user of the device depending where their workload is. Make the middle-performing transistor slightly faster and leakier? Great you can close timing at a few ps faster but the device uses more power and is hotter, good for some customers bad for others.
Even simple density is not that simple in the fine details. Density can be in opposition to leakage, switching speeds, wire RC and propagation delay, etc.
There isn't an easy way to compare and there never will be.
Slight (and somewhat related, even) throwback to when (desktop) CPUs' primarily marketed performance indicator was their core frequency (i.e. "how many MHz a CPU has"), which also got divorced from reality over time. First bus and core frequency got separated, with bus frequency growing significantly slower and getting advertised much less, then the designation got at times decoupled entirely from frequency. For instance a Cyrix 6x86 PR166+ ran at 133Mhz but exhibited "comparable performance" to competitors' 166MHz CPUs, or later the Athlon XP 1500+ that ran at 1333MHz.
>This doesn't really matter, but it seems wrong and unintuitive and I didn't know it until now and I DON'T LIKE IT [1]
>sigh I'm really annoyed that I didn't know this.
No wonder why I am upset at the quality of hardware discussions on HN. And if you want to start a spin against TSMC, which I am already observing this trend on both twitter and Reddit. ( Not yet on HN ). Dont. TSMC didn't started it. Blame ingredient marketing, blame the (cough GPU cough ) customer, market analysts, Marketing or Apple. ( Not really Apple, but media that follows and publish news about Apple ).
38 comments
[ 39.6 ms ] story [ 1534 ms ] threadIf you have a need to understand this, shouldn't you be close enough to the field to discern marketing speak from the underlying technical details?
It's the same as software engineers who don't know what SIMD is or think the CPU in their laptop or smartphone works like a MIPS they studied in college. When something like this is only one click away: https://en.wikichip.org/wiki/intel/microarchitectures/sunny_...
That's not true. There are some new "retro" consoles whose primary marketing feature is that they are 8 bit. For example the RES Plus' primary marking feature is that it plays 8 bit games.
https://old.reddit.com/r/askscience/comments/jwgdld/what_is_...
https://old.reddit.com/r/ECE/comments/jxb806/how_big_are_tra...
> The term "5 nm" is simply a commercial name for a generation of a certain size and its technology, and does not represent any geometry of the transistor.
And I learnt something new: it seems like MTr/mm² is the metric in use now. I wonder if marketing will find a way to co-opt that?
TSMC 5nm: 171 MTr/mm2 in 2021 (for apple A15)
Intel 10nm: 100.76 MTr/mm2 in 2021 (for 12th core-i)
Intel keep saying they will win in the future, while they seems forget TSMC will improve too. Comparing Intel 10nm with TSMC 7nm is outdated.
Or is there any objective metrics to determine which one is better, other than MTr/mm2?
[1] https://www.anandtech.com/show/16823/intel-accelerated-offen...
> Or is there any objective metrics to determine which one is better, other than MTr/mm2?
There are many, many objective metrics. Many are interdependent, many are tunable or selectable within certain ranges, many are in opposition, most are highly secretive, and all of them influence he type of device being made, the circuit design, logic, in different ways.
Even the experts with all the data for a particular technology and with a particular device in mind can't necessarily tell at a glance whether a change would be worthwhile. And even the meaning of worthwhile varies for the end-user of the device depending where their workload is. Make the middle-performing transistor slightly faster and leakier? Great you can close timing at a few ps faster but the device uses more power and is hotter, good for some customers bad for others.
Even simple density is not that simple in the fine details. Density can be in opposition to leakage, switching speeds, wire RC and propagation delay, etc.
There isn't an easy way to compare and there never will be.
>sigh I'm really annoyed that I didn't know this.
No wonder why I am upset at the quality of hardware discussions on HN. And if you want to start a spin against TSMC, which I am already observing this trend on both twitter and Reddit. ( Not yet on HN ). Dont. TSMC didn't started it. Blame ingredient marketing, blame the (cough GPU cough ) customer, market analysts, Marketing or Apple. ( Not really Apple, but media that follows and publish news about Apple ).
[1] https://twitter.com/Foone/status/1494565133799940096