Performance =/= clockspeed, and that hasn't been the case since like the late 90s. In the modern context performance has varying definitions. Most often it seems to entail clocks or on-current @x volts, @y leakage, and/or @iso power. It is also worth remembering performance is also not a uniform number either. Back when intel was lapping TSMC on process tech, TSMC's nodes gave competitive performance per watt characteristics at the bottom end of the power envelope.
I do agree that to the wider foundry market density seems to be more valued than performance, although I suspect there are limits to how far folks would go for density. For example take intel's claims for 10nm vs 14++. 10nm was WAY denser than 14++ (probably pretty close to 3x given the CPP relaxation that happened on 14++). Even if 10nm performed/yielded where intel originally would have wanted it too, it would have still had worse PP characteristics than 14++. This then leads to a dilema, would someone like a Qualcomm then chose to tape out a snapdragon on 14++ or 10nm? IMO, that depends on the product, but if we assume that the Freq/V curve maintained the same pattern even at the bottom of the curve, and that QCOM could get the functionality they wanted at a die cost that was acceptable on 14++; then I think they would have picked 14++ over 10nm. As for DC (the part of the foundry market that has been growing the fastest of the past few years), density is a secondary concern (doubly so now that everyone is moving in the direction of disag). Does it matter how expensive H100 is to produce? At the prices they are being sold for, not really. Does it matter how much power it draws? Absolutely! If for example the tops/W somehow got worse than A100, customers would rather buy more A100s than fewer H100s.
Of course. Nothing is ever as simple as a single number when we are talking process technology. Unfortunately these F/V curves and seeing the size of the std cell libs are about as good as we can get for comparing nodes without doing taredowns and spending weeks discussing all of the aspects where node A or B is better in xyz categories.
ICF got iced right at the end of BKs tenure. I find it hard to imagine that BS was planning a resurrection while simultaneously telling investors intel will begin outsourcing key products.
Are you surprised N5 is so much larger than N3 too? The former 7nm was supposed to be launching products back during the infancy of N5. Intel leadership even talked about it as a better than their 5"nm" node. Three years later and intel still claims a 2x boost, but is now competing with a node that is 1.6x the maximum theoretical density.
How is this still not the case. It presumably has all of the features i4 lacked. Unless all nodes have now been obsoleted by N3E, I fail to see how intel 3 is "unsuitable as a foundry node" just because the HD lib is half as dense as N3E (assuming techinsight's projections come to pass). Does that mean I expect intel 3 to take the world by storm? Not in the slightest. But I think that is a TTM problem rather than intel should have made i4/3 more aggressive problem.
I'm not sure what you're talking about with performance =/= clock speed. It never has been the case, NEVER. Performance is, and always has been, IPC x clock speed. I'm sure you must already know that, so I'm going to assume you mean clock speeds with with different voltages, which you allude to, which makes sense to me, but makes little sense in the context of that chart. It seems obvious to me the chart is referring to something approximating higher clock speeds. Intel nodes can hit the highest clock speeds, and the power curve is less malicious at top end. You can't get an AMD processor to hit the same clocks as Intel, at any the highest speeds, as the appreciation of clock speed to power use gets catastrophic. Even moreso than Intel's which clearly lose efficiency pretty fast too. They can't overclock as high, and they aren't clocked at the same speed stock. So, if you want to get into something more complicated than the chart implies, you're arguing with the chart. But, clearly it means Intel chips will be able to clock the highest. Will they be able to clock the highest at every voltage, for example, power draw? No, but it's kind of understood by most what the chart is implying. At the higher ranges, Intel chips should clock the highest. At the lowest voltages, I'd expect that not to be true, but that's not what the chart is really talking about. At least I assume it is not.
We don't have power usages, so I didn't even bother talking about them. They aren't within this context because of that. But, looking at a GPU, for example, yes, you do care about how much it costs, always. You saw NVIDIA complaining about it. And here's the gist of it, which you even alluded to without doing it explicitly, if I can pop more transistors in there, and achieve the same performance as something with fewer, but running at higher clock speeds, do I win? In a CPU, nope! In a GPU, almost certainly, because as you mention, performance isn't everywhere the same, neither is power usage. In almost all cases, if I can clock my GPU lower, and get the same performance, I will almost certainly get better efficiency, since efficiency generally gets worse and worse as you go up in clock speed, particularly as you reach for the highest clocks available. And, power use matters a lot.
What is this nonsense about using quotes around something I never said? If you're going to quote me, quote something I said, not something obtusely manufactured and put into your words. I never said I3 was "unsuitable as a foundry node", I don't like absolutes enough to say something so definitive. But, it's clearly less attractive, vis-a-vis the competition than 18A apparently will be. And if you don't believe that, just pay attention to the messaging now. I3 isn't getting much talk or interest, apparently, whereas Intel is now talking much more about 18A. Do you think that's completely unrelated to its relative merits and just timing and other aspects? I don't, I also don't think it's entirely based on what is described in these charts, but it's got something to do with it.
Clearly, I3 does not have all the features I4 lacked, high density being a primary one. It's certainly got many improvements, and also took EUV usage to a much higher level, and hopefully will be less expensive to make, but saying it has all the features I4 lacked is simplistic. Could one say, for example, I3 lacks PowerVIA? Sure can, but then, what defines lacking? It's an idiotic conversation to have. But, what both lack is, competitive density. Power use? I guess we'll have to see what that curve looks like. If it's very good, well, my feeling is that will be very important to a lot of people, if the power use within their context is favorable. So, the verdict isn't in yet, but I find them calling something I3 with such relative poor density a bit misleading. And let's remember, the 3nm, or i7, etc... was a way of expressing size. And surely, I3 doesn't measure up (yeah, I'll slap myself for you, for using such a lousy pun). But, nowadays they are a bit more liberal with their meanings, so I guess it's not as damning as it would have been.
