0 reactions
Apple M4 - Geekerwan Review with Microarchitecture analysis.
Here’s the review by Geekerwan on the M4 released on billbili https://m.bilibili.com/video/BV1NJ4m1w7zk?buvid=XUE18A0EBAE9E520C1B43F0C3258B0B0982D7&from_spmid=default-value&is_story_h5=false&mid=7SUA%2Bv5CrE02siPSjBVEP38FTQ%2FSZMtL1rElX6M3iMo%3D&p=1&plat_id=122&share_from=ugc&share_medium=android_i&share_plat=android&share_session_id=5328aabc-5769-4bcb-92fb-977d1e818509&share_source=COPY&share_tag=s_i&spmid=united.player-video-detail.0.0×tamp=1716324595&unique_k=7mRwC5f&up_id=25876945&share_times=1
For those in regions where billbili is inaccessible like myself, here’s a thread from twitter showcasing important screenshots.
https://x.com/faridofanani96/status/1793022618662064551?s=46
There was a misconception at launch that Apple’s M4 was merely a repackaged M3 with SME with several unsubstantiated claims made from throttled geekbench scores.
Apple’s M4 funnily sees the largest micro architectural jump over its predecessor since the A14 generation.
Here’s the M4 vs M3 architecture diagram https://x.com/faridofanani96/status/1793022632306024525?s=46.
* The M4 P core grows from an already big 9 wide decode to a 10 wide decode.
* Integer Physical Register File has grown by 21% while Floating Point Physical Register File has shrunk.
* The dispatch buffer for the M4 has seen a significant boost for both Int and FP units ranging from 50-100% wider structures. (Seems to resolve a major issue for M3 since M3 increased no of ALU units but IPC increases were minimal (3%) since they couldn’t be kept fed)
* Integer and Load store schedulers have also seen increases by around 11-15%.
* Seems to be some changes to the individual capabilities of the execution units as well but I do not have a clear picture on what they mean.
* Load Store Queue and STQ entries have seen increases by around 14%.
* The ROB has grown by around around 12% while PRRT has increased by around 14%
* Memory/Cache latency has reduced from 96ms to 88ms. https://x.com/faridofanani96/status/1793022632306024525?s=46
All these changes result in the largest gen on gen IPC gain for Apple silicon in 4 years. https://x.com/faridofanani96/status/1793022632306024525?s=46
In SPECint 2017, M4 increases performance by around 19%.
in SPECfp 2017, M4 increases performance by around 25%.
Clock for clock, M4 increases IPC by 8% for SPECint and 9% for SPECfp. https://x.com/lafaiel/status/1793122963807707292?s=46
But N3E does not seem to improve power characteristics much at all. In SPEC, M4 on average increases power by about 57% to achieve this.
Neverthless battery life doesn’t seem to be impacted as the M4 iPad Pro last longer by around 20 minutes.
#technology #hardware
| 17,000 mlx | |
| 17 | |
| 0 |