3 ms·

Somewhat related: I'm super curious if M1's advantages would melt away if Intel simply fabbed their current designs on TSMC's 5nm process, to have a level playi

by chubs 6y ago

Somewhat related: I'm super curious if M1's advantages would melt away if Intel simply fabbed their current designs on TSMC's 5nm process, to have a level playing field between Intel and Apple's chip designers.

lostlogin 6y ago
> if Intel simply fabbed their current designs on TSMC's 5nm process I think that ‘simply’ is the wrong word.
- skunkworker 6y ago
  This, porting an architecture to another process is a multi-year process. In a Reddit AMA a week and a half ago they mentioned that the Rocket Lake backport took 2 years (from 10nm -> 14nm) [1]. That's quite a quick schedule and also on a very mature process. "The project final concept closed in Q1 2019 and we are launching today, so your estimate is right. Took us 2 years for RKL" https://www.reddit.com/r/intel/comments/memyfv/ama_march_30th_830am_to_300_pm_pst_intel_11th/gstscr4/ https://www.reddit.com/r/intel/comments/memyfv/ama_march_30t...
  - monocasa 6y ago
    A backport is harder FWIW. You have to go looking for gates to cut from the design.
tumblewit 6y ago
There is still the problem of ARM being superior. You can’t put a “low power” x86 5nm chip in a phone right now and expect it to be on par with the iPhone. edit: also don’t forget intels fabs are actually better than tsmc when nodes are the same
- yoz-y 6y ago
  > also don’t forget intels fabs are actually better than tsmc when nodes are the same They haven’t been on the same node size for quite a while though. I wonder if this will be true in the future. Intel has a problem because even if they catch up and for some reason TSMC will be blocked at 5nm, TSMC will still have a multi year experience and will be able to improve the 5nm process.
58028641 6y ago
The M1 has greater instruction-level parallelism. That is what allows the M1 to run at a lower clock that Intel, but still run faster.
- mhh__ 6y ago
  And that parallelism in large part comes from them having a genuinely enormous transistor budget. ARM is easier to be wide on, but the way x86 cores are implemented now makes the decoding a coefficient of overhead rather than an absolute game ender.
klelatti 6y ago
Of course, you wouldn't fab the current design on TSMC 5nm as they'd have more transistors to play with. It's still not a level playing field as they'd be implementing different ISAs. If Intel acquired an ARM architecture license and then designed an ARMv8 CPU with identical constraints then you'd get a valid comparison. Although you'd likely have different IP in other areas which could skew the results. If you're looking for a comparison of the ISA's I think the decoding issues / execution width points to a likely ARM win still - only question is by how much.

by chubs 6y ago

lostlogin 6y ago

> if Intel simply fabbed their current designs on TSMC's 5nm process I think that ‘simply’ is the wrong word.

skunkworker 6y ago
This, porting an architecture to another process is a multi-year process. In a Reddit AMA a week and a half ago they mentioned that the Rocket Lake backport took 2 years (from 10nm -> 14nm) [1]. That's quite a quick schedule and also on a very mature process. "The project final concept closed in Q1 2019 and we are launching today, so your estimate is right. Took us 2 years for RKL" https://www.reddit.com/r/intel/comments/memyfv/ama_march_30th_830am_to_300_pm_pst_intel_11th/gstscr4/ https://www.reddit.com/r/intel/comments/memyfv/ama_march_30t...
- monocasa 6y ago
  A backport is harder FWIW. You have to go looking for gates to cut from the design.

tumblewit 6y ago

There is still the problem of ARM being superior. You can’t put a “low power” x86 5nm chip in a phone right now and expect it to be on par with the iPhone. edit: also don’t forget intels fabs are actually better than tsmc when nodes are the same

yoz-y 6y ago
> also don’t forget intels fabs are actually better than tsmc when nodes are the same They haven’t been on the same node size for quite a while though. I wonder if this will be true in the future. Intel has a problem because even if they catch up and for some reason TSMC will be blocked at 5nm, TSMC will still have a multi year experience and will be able to improve the 5nm process.

58028641 6y ago

The M1 has greater instruction-level parallelism. That is what allows the M1 to run at a lower clock that Intel, but still run faster.

mhh__ 6y ago
And that parallelism in large part comes from them having a genuinely enormous transistor budget. ARM is easier to be wide on, but the way x86 cores are implemented now makes the decoding a coefficient of overhead rather than an absolute game ender.

klelatti 6y ago

Of course, you wouldn't fab the current design on TSMC 5nm as they'd have more transistors to play with. It's still not a level playing field as they'd be implementing different ISAs. If Intel acquired an ARM architecture license and then designed an ARMv8 CPU with identical constraints then you'd get a valid comparison. Although you'd likely have different IP in other areas which could skew the results. If you're looking for a comparison of the ISA's I think the decoding issues / execution width points to a likely ARM win still - only question is by how much.