TSMC February 2023 Revenue Report

[ta152h]

Can we assume that you are not an engineer?

Actually my degree is in engineering, but not electrical.

So maybe I am missing something, because all the overclocking records belong to Raptor Lake. Before that, Alder Lake. Even before I7, it was Skylake (well, or actually Bulldozer). So, maybe you can explain to me why AMD processors don't scale with power as well as Intel's do? AMD sure tried this time around, with Ryzen 4, and did better, but ultimately the latest I7 iteration further improved the clock speeds Intel was able to reach.

So, how is it Intel always has higher clock speeds, both stock and in overclocking? Architecture could play a role, but when Intel consistently beats anything using TSMC, it's a bit of a pattern. It always seems to be the same pattern, AMD processors on TSMC use less power at most clock speeds, but as you add more, they don't scale to the same degree. And eventually fall flat on their faces, while Intel's hit overclocking records. It's also very well known that AMD processors don't overclock nearly as well as Intel's. So, maybe you can explain how this is so, but TSMC has a better performance node. Really, if it's something I'm not getting, I'd like to know, so do expatiate on it. And it's not like AMD doesn't like clock speed for their CPUs. Again, they sure tried this time around to match Intel, just they couldn't. Purely architecture? Maybe, but seems unlikely.

 

[SailorBob]

Actually my degree is in engineering, but not electrical.

So maybe I am missing something, because all the overclocking records belong to Raptor Lake. Before that, Alder Lake. Even before I7, it was Skylake (well, or actually Bulldozer). So, maybe you can explain to me why AMD processors don't scale with power as well as Intel's do? AMD sure tried this time around, with Ryzen 4, and did better, but ultimately the latest I7 iteration further improved the clock speeds Intel was able to reach.

So, how is it Intel always has higher clock speeds, both stock and in overclocking? Architecture could play a role, but when Intel consistently beats anything using TSMC, it's a bit of a pattern. It always seems to be the same pattern, AMD processors on TSMC use less power at most clock speeds, but as you add more, they don't scale to the same degree. And eventually fall flat on their faces, while Intel's hit overclocking records. It's also very well known that AMD processors don't overclock nearly as well as Intel's. So, maybe you can explain how this is so, but TSMC has a better performance node. Really, if it's something I'm not getting, I'd like to know, so do expatiate on it. And it's not like AMD doesn't like clock speed for their CPUs. Again, they sure tried this time around to match Intel, just they couldn't. Purely architecture? Maybe, but seems unlikely.

Perf / watt makes a big difference in the two highest margin segments mobile and DC/HPC. AFAIK AMD's offerings are leading in both those segments, and on par in consumer desktop. I think most people looking for a laptop are less concerned with raw single thread performance and more with battery life and perf / watt since you have hard constraints on total system power usage in mobile.

 

[cliff]

TA152h: Somebody who can run spice with the foundry models can compare the performance on process by running a ring oscillator with the various FETs (different threshold devices). can also current starve them (fets in series with the supply) and see the frequencies at different currents. There are other tests that can check the gain, M1 to poly capacitance (Cgd has a multiplying effect), and metal interconnect capacitances on say, M4 to M6 or M5 to M7.

You shouldn't be gauging TSMC vs Intel's foundries capability based on TSMC's customers CPU architecture.

 

[ta152h]

TA152h: Somebody who can run spice with the foundry models can compare the performance on process by running a ring oscillator with the various FETs (different threshold devices). can also current starve them (fets in series with the supply) and see the frequencies at different currents. There are other tests that can check the gain, M1 to poly capacitance (Cgd has a multiplying effect), and metal interconnect capacitances on say, M4 to M6 or M5 to M7.

You shouldn't be gauging TSMC vs Intel's foundries capability based on TSMC's customers CPU architecture.

Except I already mentioned that, didn't I? I said it could be architectural, I mean, we only have to look at the Pentium 4 and Bulldozer to see that.

But, when it's one processor after another, it sure does give credence to Intel consistently having the best process for absolute clock speed. Especially since that's always their focus, and not TSMC's.

Further, if it's architectural, is AMD really that bad at designing products? Since their IPC is very slightly worse, overall, they clearly are not going for a "brainiac" design, as we used to call it, like Apple seems to be. That's why I wouldn't compare it to the M2 or M3 line.

It's not 100%, but it's very suggestive. Minimally, there is absolutely no supporting evidence to suggest TSMC has every passed Intel in this metric. Maximally, Intel has always led. Either way, their processors are the highest clocked in the world (except possibly IBM's mainframe chips, but they use much more exotic methods for cooling/power in their standard configuration), and suggests they shouldn't be underestimated as a competitor.

 

[ta152h]

Perf / watt makes a big difference in the two highest margin segments mobile and DC/HPC. AFAIK AMD's offerings are leading in both those segments, and on par in consumer desktop. I think most people looking for a laptop are less concerned with raw single thread performance and more with battery life and perf / watt since you have hard constraints on total system power usage in mobile.

It's not that I disagree with you, but this is out of the context of what we were talking about.

Of course performance per watt matters a lot, especially in servers and laptops. And it seems clear TSMC's 5nm is better overall, by a good bit. But, I was just stating that even being "one node behind", Intel's process allows for the highest clock processors in the world, and probably exceeds 5nm in absolute performance. And that's not easy, or trivial.

But, to be honest with you, I agree with you. My typical rig is a 10 watt Tremont desktop with no fan, because I hate noise and it's plenty fast for what I do on the computer. So, no disagreement at all that absolute performance is the end all, it's not. It does show some capability though, and that's what I was trying to point out.