Strix Halo - AMD

exoticspice1

Site Champ
Joined
Jul 19, 2022
Posts
368

Looks like reviews are out today. I'm curious to see how it compares to M4 Pro.
 

Looks like reviews are out today. I'm curious to see how it compares to M4 Pro.
My guess based on available information so far is that (compared with the 14/20 CPU/GPU core Pro) the Halo’s CPU will top out as/less powerful* and much more power hungry while its GPU will be faster in most, but not all tests and definitely more power hungry. The GPU is unlikely to reach 32-core Max except in native vs non-native/badly optimized games. Anything bandwidth intensive the 20-core Pro GPU might win given that most Strix Halo devices I’ve seen come with the same memory bus but lower speed RAM. If the price of 128GB is accurate, that’s a good deal even if it has a couple of caveats.

*in embarrassingly parallel multicore, the ST/lightly threaded scenarios it’ll be no contest, the M4 Pro CPU will be better.
 
Last edited:
My guess based on available information so far is that (compared with the 14/20 CPU/GPU core Pro) the Halo’s CPU will top out as/less powerful* and much more power hungry while its GPU will be faster in most, but not all tests and definitely more power hungry. The GPU is unlikely to reach 32-core Max except in native vs non-native/badly optimized games. Anything bandwidth intensive the 20-core Pro GPU might win given that most Strix Halo devices I’ve seen come with the same memory bus but lower speed RAM. If the price of 128GB is accurate, that’s a good deal even if it has a couple of caveats.

*in embarrassingly parallel multicore, the ST/lightly threaded scenarios it’ll be no contest, the M4 Pro CPU will be better.

Scanning the above quickly, NBC's analysis largely backs up my prediction, although I was thinking 4060 Laptop GPU equivalent which seems to be true for some of the tests, but in gaming, it can also compete with a 4070 laptop which is overall how NBC decided to describe it. It depends on how high the Nvidia GPU has been clocked (some OC's laptop 4050s even compete) and power levels. So definitely quite nice.

Unfortunately only one data point for my CB R24 graph.
 
Screenshot 2025-02-18 at 9.12.49 AM.png

Click to expand. Things are getting a bit scrunched up, haven't had time to prune the data points (unfortunately Numbers isn't great here, change the order or remove a data point and the entire graph's style has to be redone).

With respect to the Strix Halo in Asus ROG Flow (AI Max+ 395) I have to admit this a little worse on the CPU side than I thought it would be. I expected higher power/lower efficiency for the ST but not this bad and I thought there would be better performance than the Strix Point. Instead it is just same performance, much worse efficiency - not totally unexpected given the larger SOC design, but this is substantially worse. It almost uses as much power in ST as the base M4 does in multicore. Keep in mind, as per my conversation with @Artemis in the other thread, this is with the idle power removed and the idle power here is substantial - 11.2W and this all due to the SOC, no dGPU to blame and it is a less than 14" laptop/tablet hybrid. With only one data point I don't know if that is a feature of the Halo's design or Asus' implementation.

Similar story with multithreaded efficiency. Here performance is around where I thought it would be, but efficiency is way worse. I knew it would be lower, I didn't think the 14-core M4 Pro would be over 40% more efficient (side note: in this analysis I use the 14-core M4 Pro from the 14" Pro which had a slightly lower performance and slightly better efficiency than the 16" device reported in the Halo article). I don't know if Strix Halo is using TSMC's 4NX or 4NP. If the former, that could explain some of the efficiency losses for both ST and MT and it uses a similar structure to the desktop chip dies (2 CCDs and full AVX-512). Although ... I guess it's getting 41% more performance than the Strix Point for roughly the same power increase when the HX 370 nets a score of 1166, so maybe at lower power the Halo will be much more efficient and come closer to the 12-core M4 Pro? Running the numbers based on the Strix Point's performance curve that is certainly plausible. Of course the 12-core M4 Pro is considerably smaller with only 8+4 P+E-cores than the 16-core Halo with hyperthreading and we don’t know what the performance curve of the Halo actually is, it could be shallower/steeper than the Point’s. What we do know is that at this singular performance point, the Halo is using a fair bit more power than the 16-core M4 Max to just barely lose to the 14-core M4 Pro. So maybe 60-70W will be its sweet spot like 40-50W was for the Strix Point (wall power not TDP).

Price is higher than I thought it would be too (2500 Euros for 32GB model) - so now I'm not sure if the price I saw quoted for the 128GB model is accurate (I'd previously seen placeholder like $2600 for the 128GB model and $2100 for the 32GB model which I don't see how that would be possible unless Europe is getting screwed here).

GPU looks nice. Most of the rendering/macOS native gaming benchmarks put it in-between the 40-core and 20-core M4 GPUs (no 32-core model tested, but my sense is it should be better than or at worst equivalent to the Halo's GPU) except in a couple of tests Apple is known to do well in where the 20-core M4 GPU actually beats it. It uses slightly more power in the CP 2077 test than the M4 Max while getting better performance/efficiency than the M4 Pro/Max, but obviously that game is not macOS native. Also, no really memory intensive benchmarking done here (this was the smaller 32GB RAM capacity but still). Like another commenter on the analysis article already said, I would've liked to see some AI benchmarks since that is a major purpose the Halo is being advertised for - also wouldn't have minded some Blender render benchmarks although RDNA 3.5’s ray tracing has generally been subpar. There are other reviews out there I just haven't had time to look at them, so maybe some of them have those.


 
Last edited:
View attachment 33916

Click to expand. Things are getting a bit scrunched up, haven't had time to prune the data points (unfortunately Numbers isn't great here, change the order or remove a data point and the entire graph's style has to be redone).

With respect to the Strix Halo in Asus ROG Flow (AI Max+ 395) I have to admit this a little worse on the CPU side than I thought it would be. I expected higher power/lower efficiency for the ST but not this bad and I thought there would be better performance than the Strix Point. Instead it is just same performance, much worse efficiency - not totally unexpected given the larger SOC design, but this is substantially worse. It almost uses as much power in ST as the base M4 does in multicore. Keep in mind, as per my conversation with @Artemis in the other thread, this is with the idle power removed and the idle power here is substantial - 11.2W and this all due to the SOC, no dGPU to blame and it is a less than 14" laptop/tablet hybrid. With only one data point I don't know if that is a feature of the Halo's design or Asus' implementation.

Similar story with multithreaded efficiency. Here performance is around where I thought it would be, but efficiency is way worse. I knew it would be lower, I didn't think the 14-core M4 Pro would be over 40% more efficient (side note: in this analysis I use the 14-core M4 Pro from the 14" Pro which had a slightly lower performance and slightly better efficiency than the 16" device reported in the Halo article). I don't know if Strix Halo is using TSMC's 4NX or 4NP. If the former, that could explain some of the efficiency losses for both ST and MT and it uses a similar structure to the desktop chip dies (2 CCDs and full AVX-512). Although ... I guess it's getting 41% more performance than the Strix Point for roughly the same power increase when the HX 370 nets a score of 1166, so maybe at lower power the Halo will be much more efficient and come closer to the 12-core M4 Pro? Running the numbers based on the Strix Point's performance curve that is certainly plausible. Of course the 12-core M4 Pro is considerably smaller with only 8+4 P+E-cores than the 16-core Halo with hyperthreading and we don’t know what the performance curve of the Halo actually is, it could be shallower/steeper than the Point’s. What we do know is that at this singular performance point, the Halo is using a fair bit more power than the 16-core M4 Max to just barely lose to the 14-core M4 Pro. So maybe 60-70W will be its sweet spot like 40-50W was for the Strix Point (wall power not TDP).

Price is higher than I thought it would be too (2500 Euros for 32GB model) - so now I'm not sure if the price I saw quoted for the 128GB model is accurate (I'd previously seen placeholder like $2600 for the 128GB model and $2100 for the 32GB model which I don't see how that would be possible unless Europe is getting screwed here).

GPU looks nice. Most of the rendering/macOS native gaming benchmarks put it in-between the 40-core and 20-core M4 GPUs (no 32-core model tested, but my sense is it should be better than or at worst equivalent to the Halo's GPU) except in a couple of tests Apple is known to do well in where the 20-core M4 GPU actually beats it. It uses slightly more power in the CP 2077 test than the M4 Max while getting better performance/efficiency than the M4 Pro/Max, but obviously that game is not macOS native. Also, no really memory intensive benchmarking done here (this was the smaller 32GB RAM capacity but still). Like another commenter on the analysis article already said, I would've liked to see some AI benchmarks since that is a major purpose the Halo is being advertised for - also wouldn't have minded some Blender render benchmarks although RDNA 3.5’s ray tracing has generally been subpar. There are other reviews out there I just haven't had time to look at them, so maybe some of them have those.



On the other site Xiao_Xi posted a pretty good video review and though I'm not sure how the reviewer measures power, it seems to confirm to me that the Strix Halo would be substantially more efficient at lower power levels, not a surprise perhaps and we still need to get exact power measurements at the lower power/performance points as NBC provides. But I strongly suspect that the "70W" TDP setting is fairly far along the curve.

The reviewer also did AI workloads and confirmed that for programs optimized Apple Silicon, the memory bandwidth of the MBP should provide better performance. The reviewer didn't do a GPU Blender benchmark unfortunately.

I will also try to do a silicon analysis based on the techpowerup article soon.
 
Price is higher than I thought it would be too (2500 Euros for 32GB model) - so now I'm not sure if the price I saw quoted for the 128GB model is accurate (I'd previously seen placeholder like $2600 for the 128GB model and $2100 for the 32GB model which I don't see how that would be possible unless Europe is getting screwed here).

Arstechnica quoting $2300 for what appears to be the same model:


Is the detachable keyboard included in the price? I dunno, someone is either wrong or maybe just Europe is getting the short end of the stick?
 
Back
Top