Small CBR24 efficiency update adding Intel Arrow Lake (285H) + M4 base (previously was an estimate) + another two HX370 points for comparison with Arrow Lake:
(expand chart)
Usual caveats from previous posts apply. With Arrow Lake analysis, NotebookCheck finally posted Apple M4 (base) results. My previous M4 base estimate was nearly spot-on for single thread, well within random variance. For multithreaded perf/efficiency, I was estimating efficiency based on a lower performing M4 chip that had a score of around 950 but 10% better efficiency than the M4 MacBook Pro Notebookcheck reported both performance and power for - possibly within random variation especially considering the lower performance. So not bad.
As NotebookCheck says 16-core Arrow Lake is an odd duck especially in relation to the existence of 8-core Lunar Lake. Of course Lunar Lake keeps going to lower power levels with lower performance, has a smaller CPU, and has much better single threaded efficiency (as an aside it also apparently has a better NPU and newer GPU). Although given how reportedly expensive Lunar Lake is with its on-package memory, it isn't clear if it is more cost-effective for Intel than Arrow Lake. But Lunar Lake's 8-core CPU clearly can't perform MT tasks very well. It's not designed to, but regardless Arrow Lake with its larger core count can easily blow past it in both performance and efficiency and can go down to the power levels that Lunar Lake operates at (though not the lowest possible levels for Lunar Lake).
So how does Arrow Lake compare to its competitors? Well the Apple M4 is so far ahead of just about everyone in single threaded performance/efficiency that it’s not really worth comparing. There were two Intel Arrow Lake devices with slightly different efficiencies and I chose the better of the two. This had an identical efficiency to the Strix Point chip albeit at a higher power draw/performance which is a small win (slightly better node plays a part here). The Qualcomm chips get similar performance but at much lower power.
In multithreaded efficiency, there are two Arrow Lake data points (~46W, MSI, and ~34W, Zenbook). This 16-core/thread Arrow Lake is very similar in performance and efficiency to the 12-core/24-thread HX 370 Strix Point chip at the higher power draw in the MSI device. This sort of tracks with expectations even though the Arrow Lake is on a slightly better node it has 14-16 threads - 2 are on a low power island (and don’t participate in multicore workloads making it effectively a 14-core), 6 are P-cores, and the remaining 8 are E-cores, while if I remember correctly the uplift from hyperthreading in CB R24 is roughly 25% making the HX 370 roughly a 15-thread equivalent (though all of those are effectively P-cores in the context of multithreading given that "c"-cores should behave similarly enough to P-cores here). But something clearly is wrong with the Arrow Lake Zenbook at the lower power draw getting only 740 points, it should be doing better. The Arrow Lake CPU has exact same efficiency here as at the higher power draw, which shouldn’t happen. At the lower power draw, with the same number of cores, its efficiency should increase as clock speeds go down. But it doesn’t. Not sure why. The Zenbook oddly had the better single threaded efficiency between the two, so it isn't like the Zenbook is just worse all around. The AMD HX 370 and Qualcomm X1E-78 show expected MT performance curves as power increases. The Arrow Lake H-series can also go to higher power levels, but it remains to be seen what performance gains that yields.*
Higher power/performance Arrow Lake HX-CPUs with 24-cores/thread are expected as are larger 16-core/32-thread Strix Halo and Fire range AMD CPUs. Based on results so far, even when these chips match the performance of the Apple M-series, it'll inevitably be at higher power draws even with the larger core/thread counts. These larger core/thread counts would in theory allow for greater down clocking and efficiencies to achieve the same performance, but relative to the M4, Apple's microarchitecture advantages means that still doesn't bring them to parity even in a test that is very multithreaded like CB R24. The closest these x86 processors get in MT efficiency to an Apple processor occurs at ~37W the HX 370 just misses the base M4's performance/efficiency, but the base M4 achieves this with significantly fewer threads available and only 4 P-cores. For any task that is single-threaded or lightly threaded, Apple's advantage grows. And when Apple adds P-cores for chips that are designed for greater multithreaded throughput, the efficiency of M4 Pro and Max becomes even tougher for AMD/Intel to match.
*EDIT: The MSI TDP chart in the Arrow Lake analysis article gives us some clue as to the Zenbook’s behavior. While NBC didn't measure wall power at each TDP, we can see that as power decreases performance for the Intel Arrow Lake H chip drops off significantly faster than for the AMD chip assuming they are drawing similar power at the same TDP, which they are definitely aren't given the wall power results. Even so, given the Zenbook results, this is not what I was expecting from the performance/power curve, so hopefully they add more wall power measurements for Arrow Lake H at different performance settings. It’s possible Artie Lake H just doesn’t do well at low power for some reason.
(expand chart)
Usual caveats from previous posts apply. With Arrow Lake analysis, NotebookCheck finally posted Apple M4 (base) results. My previous M4 base estimate was nearly spot-on for single thread, well within random variance. For multithreaded perf/efficiency, I was estimating efficiency based on a lower performing M4 chip that had a score of around 950 but 10% better efficiency than the M4 MacBook Pro Notebookcheck reported both performance and power for - possibly within random variation especially considering the lower performance. So not bad.
As NotebookCheck says 16-core Arrow Lake is an odd duck especially in relation to the existence of 8-core Lunar Lake. Of course Lunar Lake keeps going to lower power levels with lower performance, has a smaller CPU, and has much better single threaded efficiency (as an aside it also apparently has a better NPU and newer GPU). Although given how reportedly expensive Lunar Lake is with its on-package memory, it isn't clear if it is more cost-effective for Intel than Arrow Lake. But Lunar Lake's 8-core CPU clearly can't perform MT tasks very well. It's not designed to, but regardless Arrow Lake with its larger core count can easily blow past it in both performance and efficiency and can go down to the power levels that Lunar Lake operates at (though not the lowest possible levels for Lunar Lake).
So how does Arrow Lake compare to its competitors? Well the Apple M4 is so far ahead of just about everyone in single threaded performance/efficiency that it’s not really worth comparing. There were two Intel Arrow Lake devices with slightly different efficiencies and I chose the better of the two. This had an identical efficiency to the Strix Point chip albeit at a higher power draw/performance which is a small win (slightly better node plays a part here). The Qualcomm chips get similar performance but at much lower power.
In multithreaded efficiency, there are two Arrow Lake data points (~46W, MSI, and ~34W, Zenbook). This 16-core/thread Arrow Lake is very similar in performance and efficiency to the 12-core/24-thread HX 370 Strix Point chip at the higher power draw in the MSI device. This sort of tracks with expectations even though the Arrow Lake is on a slightly better node it has 14-16 threads - 2 are on a low power island (and don’t participate in multicore workloads making it effectively a 14-core), 6 are P-cores, and the remaining 8 are E-cores, while if I remember correctly the uplift from hyperthreading in CB R24 is roughly 25% making the HX 370 roughly a 15-thread equivalent (though all of those are effectively P-cores in the context of multithreading given that "c"-cores should behave similarly enough to P-cores here). But something clearly is wrong with the Arrow Lake Zenbook at the lower power draw getting only 740 points, it should be doing better. The Arrow Lake CPU has exact same efficiency here as at the higher power draw, which shouldn’t happen. At the lower power draw, with the same number of cores, its efficiency should increase as clock speeds go down. But it doesn’t. Not sure why. The Zenbook oddly had the better single threaded efficiency between the two, so it isn't like the Zenbook is just worse all around. The AMD HX 370 and Qualcomm X1E-78 show expected MT performance curves as power increases. The Arrow Lake H-series can also go to higher power levels, but it remains to be seen what performance gains that yields.*
Higher power/performance Arrow Lake HX-CPUs with 24-cores/thread are expected as are larger 16-core/32-thread Strix Halo and Fire range AMD CPUs. Based on results so far, even when these chips match the performance of the Apple M-series, it'll inevitably be at higher power draws even with the larger core/thread counts. These larger core/thread counts would in theory allow for greater down clocking and efficiencies to achieve the same performance, but relative to the M4, Apple's microarchitecture advantages means that still doesn't bring them to parity even in a test that is very multithreaded like CB R24. The closest these x86 processors get in MT efficiency to an Apple processor occurs at ~37W the HX 370 just misses the base M4's performance/efficiency, but the base M4 achieves this with significantly fewer threads available and only 4 P-cores. For any task that is single-threaded or lightly threaded, Apple's advantage grows. And when Apple adds P-cores for chips that are designed for greater multithreaded throughput, the efficiency of M4 Pro and Max becomes even tougher for AMD/Intel to match.
*EDIT: The MSI TDP chart in the Arrow Lake analysis article gives us some clue as to the Zenbook’s behavior. While NBC didn't measure wall power at each TDP, we can see that as power decreases performance for the Intel Arrow Lake H chip drops off significantly faster than for the AMD chip assuming they are drawing similar power at the same TDP, which they are definitely aren't given the wall power results. Even so, given the Zenbook results, this is not what I was expecting from the performance/power curve, so hopefully they add more wall power measurements for Arrow Lake H at different performance settings. It’s possible Artie Lake H just doesn’t do well at low power for some reason.
Last edited:
