3 ms·

Nice writeup overall, but I think it falls victim to the "trusting the number on the tin" trap. If you look purely at the names TSMC and Intel assigned to their

Nice writeup overall, but I think it falls victim to the "trusting the number on the tin" trap. If you look purely at the names TSMC and Intel assigned to their process nodes in the 2010s, it'd be easy to look at the timeline and think "TSMC got smaller transistors than Intel, and so Intel fell behind. Ergo, the core architecture didn't matter." In reality, TSMC only caught up to Intel's process in 2016. The playing field remained pretty even over the ensuing 4 years. The level playing field exposed Intel's stagnant core architecture, breaking their datacenter hegemony. I think the author gets the big picture right - the process node sets the bar. But, I do wonder if the author glosses over architecture too easily. Sure, history tells us that it only matters when the "process playing field" is level. But, times are changing. There are fewer players than ever in the cutting-edge process arena, yet the cost of implementing those new processes marches forward at its usual exponential pace. Major world powers are more interested than ever in keeping their domestic fabs at the forefront of the field. In this environment, can anyone safely assume they'll be able to lean into a long-term process advantage?

adrian_b 2y ago
The quality of the CMOS manufacturing process is the main factor that determines the multi-threaded performance of a CPU when it is constrained by a power limit. The design of the CPU matters for this quality criterion only in the unlikely scenario when the design team was not competent and they have made a far from optimal design. When AMD has transitioned to TSMC in 2019, they have instantly got a much better MT performance per watt than Intel, so the TSMC "7 nm" was already better than anything Intel had in 2019. At that time, "7 nm" was already an old process for TSMC, having been used by Apple since 2018. Therefore, at the latest in 2018 TSMC was well in advance over Intel. So what you said, that "TSMC only caught up to Intel's process in 2016. The playing field remained pretty even over the ensuing 4 years." is only partially correct. No later than 2018, i.e. no later than 2 years after 2016, "the playing field" was no longer even. Perhaps TSMC was already above Intel long before 2018, but there are no older products in direct competition that would allow us to judge whose CMOS process was better.
- ryukoposting 2y ago
  > The design of the CPU matters for this quality criterion only in the unlikely scenario when the design team was not competent and they have made a far from optimal design. I think this is an overstatement in a couple of ways: > unlikely scenario We're still less than a decade removed from the final descendants of Bulldozer. We can't directly compare it to Zen because of the process differences, but Bulldozer sucked in several ways that weren't related to its 28nm process. > made a far from optimal design AMD and Intel continue to release improved x86 architectures. Alder Lake introduced heterogeneous cores to x86. Zen 5's dual decode path is less visible, but an innovation in its own right. If there is a long-term local max in x86 design, we haven't found it yet. That's not to say flagrantly bad design is impossible (again, Bulldozer) - it's only to say that there are indeed two avenues to making a CPU better. Even if one avenue (process) sets the hypothetical ceiling for the other (architecture), we haven't found that ceiling very often for any particular process node. Incremental improvements are a routine part of any uarch, even when the process remains the same.

adrian_b 2y ago

The quality of the CMOS manufacturing process is the main factor that determines the multi-threaded performance of a CPU when it is constrained by a power limit. The design of the CPU matters for this quality criterion only in the unlikely scenario when the design team was not competent and they have made a far from optimal design. When AMD has transitioned to TSMC in 2019, they have instantly got a much better MT performance per watt than Intel, so the TSMC "7 nm" was already better than anything Intel had in 2019. At that time, "7 nm" was already an old process for TSMC, having been used by Apple since 2018. Therefore, at the latest in 2018 TSMC was well in advance over Intel. So what you said, that "TSMC only caught up to Intel's process in 2016. The playing field remained pretty even over the ensuing 4 years." is only partially correct. No later than 2018, i.e. no later than 2 years after 2016, "the playing field" was no longer even. Perhaps TSMC was already above Intel long before 2018, but there are no older products in direct competition that would allow us to judge whose CMOS process was better.

ryukoposting 2y ago
> The design of the CPU matters for this quality criterion only in the unlikely scenario when the design team was not competent and they have made a far from optimal design. I think this is an overstatement in a couple of ways: > unlikely scenario We're still less than a decade removed from the final descendants of Bulldozer. We can't directly compare it to Zen because of the process differences, but Bulldozer sucked in several ways that weren't related to its 28nm process. > made a far from optimal design AMD and Intel continue to release improved x86 architectures. Alder Lake introduced heterogeneous cores to x86. Zen 5's dual decode path is less visible, but an innovation in its own right. If there is a long-term local max in x86 design, we haven't found it yet. That's not to say flagrantly bad design is impossible (again, Bulldozer) - it's only to say that there are indeed two avenues to making a CPU better. Even if one avenue (process) sets the hypothetical ceiling for the other (architecture), we haven't found that ceiling very often for any particular process node. Incremental improvements are a routine part of any uarch, even when the process remains the same.