The M4 Max boasts a clock speed of 4.4 GHz, while the M1 Max clocks in at 3.2 GHz. This difference alone yields a 40% performance boost, which could have been achieved for free if Apple had permitted overclocking.
Furthermore, the M4 Max features 16 cores, compared to the M1 Max’s 10 cores. This disparity accounts for approximately 60% of the performance difference.
In essence, the M4 Max offers double the speed, achieved through both overclocking and the addition of six more cores. This approach mirrors Intel’s strategy of simply creating larger and hotter chips.
If the M4 Max had a clock speed of 3.2GHz and 10 cores, it would be comparable to the M1 Max.
Sure if we take at face value those factors would net a 2.2x increase vs the 2.3x we actually see (in this renderer at least, CB R24 shows a 2.55x increase from the M1 Max, ~800, to the M4 Max, ~2000, which still outpaces 2.2x by quite a fair margin). However, there are several things wrong with your analysis:
1) Max clocks are not the clocks used during MT workloads. The latter may have increased by 40%, they may not. Judging by the results below, they almost certainly didn't.*
2) The M1 Max had an 8+2 core configuration with the two efficiency cores "double boosted" for 2x core clocks. From the M2 onwards the Max got 4 efficiency cores with the same clocks (as did the M2 and M4 Pro, with the M3 Pro being a very different architecture). As such, the relevant metric is 50% more P-cores rather than 60% more cores.
3) As AMD has said in their description of Zen 5 vs Zen 5c, core architecture has to be designed specifically to enable much higher boost clocks. The main difference between the two Zen cores, in Strix Point at least (desktop/Halo/Fire rise also have full AVX-512), is the circuitry required to enable to higher clocks ... and even with those, the Zen 5 Strix Point mobile cores are no more overclockable than Apple Silicon is (especially not in practice more on that below). Also, not repeat point 1, overclocking for ST and overclocking for MT are very different paradigms. Further, overclocking for either requires substantial binning, only a small percentage of dies are fit for that purpose and even then only a few if any can get to 40% ST increases. Selling such binned chips is the business model of a chip-seller, not a device-seller like Apple. Finally it's not "for free". Even if possible, which is unlikely (again see below), would require massive energy for very little gain, especially as aforementioned given the likely design of Apple chips.
So rather than hypothesizing, let's take a look at actual data
using a larger version of a graph that I put in a different thread in response to a post under similar misapprehension that Apple's CPU design had stagnated. The below plots Redshift rendering results from CB R24 against Power. Inside the bubbles is efficiency (Pts per W). As a point of comparison to the M4 Pro/Max, we'll use the M2 Pro which had the same 4-e-core layout as all later Max chips and shared the same CPU as the M2 Max:
Comparing ST efficiency of the M4 Pro to the M2 Pro, we see that it has increased by 22%. Meanwhile the 8+4 M4 Pro with the same configuration as the M2 Pro has increased its MT efficiency by about 34%. TSMC rates N3E 18% performance at the same power vs N5, but the M2 Pro was on N5P was already 5% better performance. Again this is highly dependent on chip and where on the clock speed a particular core is already at so it shouldn't be taken exactly but Apple is very much outpacing gains TSMC is claiming for its nodes. Further it's important to note that the performance of the 14-core and 16-core variants of the M4 Pro/Max increase their performance linearly along with the power such that they all have the same efficiency.
To look at what happens when one tries to overclock a chip not designed for it, we can turn to the HX 370 AMD chips. Now the Strix Point CPU is not user overclockable, according to Techpowerup the multiplier is locked, nor do I think anyone, OEM or otherwise, overclocks its single core. BUT Notebookcheck used a version of the HX 370 whose MT performance was allowed to be overclocked past AMD's recommended TDP range ("54"W) to "65W" and "80W" which actually used close to 90W (rightmost AMD HX 370 point in quoted graph) and over 100W (not pictured) respectively for little to no gain in performance. At "54W" TDP (actually uses 74W) it gets a score of 1166, at "65W" (88W) it gets about 1200, at "80W" (109W) it still only gets 1216 (not pictured). Basically going from AMD's recommended TDP settings of "54W" (74W) to "80W" (109W), the chip burns ~50% more power for only 4% more performance and almost all of that is gained before it hits "65W" 88W. You can see why AMD lists the Strix Point HX 370's max TDP as "54W". This also is par for the course for Intel chips when pushing "bigger and hotter" chips as you put it, not linear increases as with Apple silicon, but exponential ones. Since Apple does not design their cores to be pushed like this, it is highly unlikely that user overclockable AS would have achieved noticeable MT performance gains.
In conclusion, we can see that in at least this one renderer's case, the M4 Max has improved efficiency over the M2 Max by about 40% in addition to having its performance increase by 2.55x over the M1 Max (and 2x over the M2 Pro/Max). Sadly I don't have efficiency figures for the M1 Max but the improvement is probably better than 40%. Of course not every renderer will show as good results, Redshift for CB R24 seems particularly well optimized for AS. Lastly, depending on the design of the core, overclocking does not necessarily yield better performance and certainly not for free.
*EDIT: Based on
Geekerwan and
Anandtech the all-core clocks appear to have gone up by just under 28% in the M4 Max since the M1 Max. Based on the P-core alone we would expect a 1.28x1.5 = 1.92 gain in performance while the actual CBR24 performance gain is 2.55 which gives a ratio of 2.55/1.92 of about 1.33 and the observed, measured efficiency gain in CBR24 is 34% for the M4 Max since the M2 Max. The efficiency gain may be lower/same/higher since the M1, but any discrepancy is likely due to not taking the E-cores or their improvement into account.
