Who knows. Maybe patents for some solutions inside the chips that Apple holds? Maybe the fact that they had to redesign some parts of the chip in hurry due to the ongoing lawsuit, as the rumor goes. Plus I doubt Nuvia guys had anything to do with botched GPUI wish Gerard could tell us what was different at Qualcomm compared to Apple that prevented him from delivering something on par with M3. Is Apple using well researched productivity boosting tricks in their campus (maybe serving free brain enhancing concoctions to important engineers)?
I don’t there is much that prevents others from doing what Apple is doing, it is just that it takes time and hard work. People keep commenting as if all you need to do to design power and efficiency is to just decide to do it, like wondering what’s taking so long for some kid to put on his shoes.I wish Gerard could tell us what was different at Qualcomm compared to Apple that prevented him from delivering something on par with M3. Is Apple using well researched productivity boosting tricks in their campus (maybe serving free brain enhancing concoctions to important engineers)?
This means if Qualcomm made a laptop SoC using 2nd gen Oryon cores, the efficiency would be similar to Apple M3.1st generation Oryon CPU wasn't very good, efficiency wise.
View attachment 112184View attachment 112185
2nd generation Oryon CPU is much improved. ST Efficiency of Snapdragon 8 Elite is comparable to A17 Pro. One generation behind Apple.
View attachment 112181
I wish Gerard could tell us what was different at Qualcomm compared to Apple that prevented him from delivering something on par with M3. Is Apple using well researched productivity boosting tricks in their campus (maybe serving free brain enhancing concoctions to important engineers)?
True, true but wouldn't Gerard have assembled his dream team at Qualcomm? What went wrong?Individuals don't design SoCs with tens of billions of transistors like M3, teams do. Hypothetically, if the Qualcomm team as a whole isn't the equal of Apple's team as a whole, it wouldn't matter if Apple's lead wasn't as good as Qualcomm's lead.
Almost 40% faster than my S24+.samsung SM-S936U - Geekbench
Benchmark results for a samsung SM-S936U with a Qualcomm ARMv8 processor.browser.geekbench.com
Galaxy S25+ with Snapdragon 8 Elite visits Geekbench.
4.47 GHz peak clock speed.
True, true but wouldn't Gerard have assembled his dream team at Qualcomm? What went wrong?
It must be noted these numbers are shipments, not sales.Canalys told TechRadar Pro, “As this was the first full quarter of shipments for Snapdragon X Series PCs, we saw sequential growth of around 180% compared to Q2 2024.
| Shipments | QoQ | |
| Q2 | 400,000 | |
| Q3 | 720,000 | 180% |
The X Elite and X Plus chips, used for Windows on ARM (WOA), will reach about 2 million unit shipments in 2024
Or... Intel Foundry perhaps?Surprisingly, its main competitor (Qualcomm) is nowhere to be found on that list. This casts doubt about the future of Snapdragon X Elite and 8 Elite successors, giving further credence to rumours about the American chipmaker pivoting to Samsung Foundry once again.
So do I understand the video correctly, that his old SPEC runs are invalid, and the new ones were done using Ubuntu under WSL? At least when it comes to Lunar Lake? But then they are not labelled as such? Or were all runs redone under WSL?Important screenshots:
My understanding (based on the auto generated translated subtitles), is that prior Lunar Lake efficiency results are invalid. Results for other CPUs and for lunar lake max performance should still be fine. But Geekerwan tends to be light on the details, and I don't speak Mandarin.So do I understand the video correctly, that his old SPEC runs are invalid, and the new ones were done using Ubuntu under WSL? At least when it comes to Lunar Lake? But then they are not labelled as such? Or were all runs redone under WSL?
lol the lunar lake results didn’t change that much with new Linux updates.My understanding (based on the auto generated translated subtitles), is that prior Lunar Lake efficiency results are invalid. Results for other CPUs and for lunar lake max performance should still be fine. But Geekerwan tends to be light on the details, and I don't speak Mandarin.
I wonder whether running the benchmarks through a hypervisor is affecting the results.
lol the lunar lake results didn’t change that much with new Linux updates.
He speculated that it would in the first video despite the fact that Intel is the single most aggressive in working to upstream Linux firmwares/drivers for their products and Qualcomm is obviously behind Lunar Lake in Linux mainline.
A bunch of usual DIY neckbeards online reflexively asssumed that since Geekerwan only mentioned possible Lunar Lake issues with Linux, the claim “Snapdragon results are wrong especially the low end” was somehow invalid and Geekerwan wouldn’t miss that. The reality is this line of reasoning was just plainly motivated against Qualcomm after Andrei F. (Among others, the first literally didn’t match what we saw about QC’s power floor and I believe their perf/W curves) claimed it was bad data.
And ironically/sure enough: the Lunar Lake curve didn’t change too much. What did was Qualcomm’s and it predictably shifted two entire watts to the left and is now on par with Intel’s Lion Cove in Lunar Lake for SpecInt.
Kind of pathetic Intel’s expensive one-off Skunk Works Black Project rip-off-Apple is that meh on these things despite all the money and area but yeah, the actual ST integer ppw curve between QC and Intel are similar.
Before:
View attachment 112305
After:
View attachment 112306
The lunar lake linux updates haven't come, that's why geekerwan ran the benchmark in a linux virtual machine under a Windows host. He presents this as a way to fix the issue, but I'm skeptical.lol the lunar lake results didn’t change that much with new Linux updates.
He speculated that it would in the first video despite the fact that Intel is the single most aggressive in working to upstream Linux firmwares/drivers for their products and Qualcomm is obviously behind Lunar Lake in Linux mainline.
A bunch of usual DIY neckbeards online reflexively asssumed that since Geekerwan only mentioned possible Lunar Lake issues with Linux, the claim “Snapdragon results are wrong especially the low end” was somehow invalid and Geekerwan wouldn’t miss that. The reality is this line of reasoning was just plainly motivated against Qualcomm after Andrei F. (Among others, the first literally didn’t match what we saw about QC’s power floor and I believe their perf/W curves) claimed it was bad data.
And ironically/sure enough: the Lunar Lake curve didn’t change too much. What did was Qualcomm’s and it predictably shifted two entire watts to the left and is now on par with Intel’s Lion Cove in Lunar Lake for SpecInt.
Kind of pathetic Intel’s expensive one-off Skunk Works Black Project rip-off-Apple is that meh on these things despite all the money and area but yeah, the actual ST integer ppw curve between QC and Intel are similar.
Before:
View attachment 112305
After:
View attachment 112306
So do I understand the video correctly, that his old SPEC runs are invalid, and the new ones were done using Ubuntu under WSL? At least when it comes to Lunar Lake? But then they are not labelled as such? Or were all runs redone under WSL?
I found the context for this on r/hardware. Andrei F commented on Geekerwan's Lunar Lake review saying that the X Elite Spec2017 power curve is lower than it should be. With that context, it sounds like Geekerwan would have retested all CPUs.My understanding (based on the auto generated translated subtitles), is that prior Lunar Lake efficiency results are invalid. Results for other CPUs and for lunar lake max performance should still be fine. But Geekerwan tends to be light on the details, and I don't speak Mandarin.
I wonder whether running the benchmarks through a hypervisor is affecting the results.
To my understanding it should, but we lack the details to be honest. There are 2 flavours of WSL, one of which is not enforcing the core affinity based on what I have read some time ago ( https://github.com/Microsoft/WSL/issues/3827 next to last comment) not sure about WSL1. It might be there are some knobs in the OS to force the behaviour but I have not looked deep enough and reviewers don't seem to care to detail their setups to such degreeI wonder whether running the benchmarks through a hypervisor is affecting the results.