I look forward to 3rd-party battery test comparisons between the new Surface and MBA. In my experience, Apple's battery life claims are more accurate, but maybe MS is on the up and up here.
When the third parties get their hands on the Qualcomm device they will find that the numbers have been massaged by Qualcomm one thing that is telling is the wattage used by the Elite SOC when compared to Apples SOC, the Elite is using nearly four times more watts 80 thanboth M3 and the new M4 which are beasts when it comes to speed/power and low wattage. 80 watts only looks good when compared to Intel or AMD laptops.
Some of these comparisons name sense, and some don’t. A major one that may not is the TOPS scores. Generally TOPS is measured as 16 but, 8 bit and 4 bit. Apple has. Even measuring and optimizing for 16 bit. Most others measure and optimize for 8 bit. While 16 bit is more accurate, LLMs work to 8 bit. So we see with the new M4, Apple says they optimized 8 /bit, and that’s where that new rating comes from. Apple is somewhat conservative. So I don’t know what Apple’s TOPS rating for the M3 are in 8 bit.
The elephant in the room though is Copilot. I haven’t used it myself but I know several people who have as part of Office365. They hate it, like truly despise it. Copilot is pushy and they find it getting in the way more than ever helping. Clippy on steroids. So the hardware is only part of the story. The real test will come with Apple’s AI efforts this fall. Let’s hope they aren’t as obnoxious as Microsoft’s apparently are.
I could be wrong, but on every demo of Copilot in MS Office that I have seen, you start it by clicking an icon in the toolbar, very different from Clippy. I don't see it as pushy if it works in that way.
The people I talked to said you can do it that way, but it will also pop up whenever it feels it could be of help. This is what they find most obnoxious, it’s not a little thing in the corner, but a Copilot window in the middle of the screen in front of what you’re trying to do.
Given the Surface has a fan, why compare it to the MacBook Airs rather than the MacBook Pros?
Surface Laptop 15" is $1299 for Snapdragon X Elite with 16 GB RAM and 256 GB SSD
MacBook Air 15" is $1299 for M3 with 8 GB RAM and 256 GB SSD
MacBook Pro 16" is $2499 for M3 Pro with 18 GB RAM and 512 GB SSD
-
Surface Laptop 13" is $999 for Snapdragon X Plus with 16 GB RAM and 256 GB SSD
Macbook Air 13" is $1099 for M3 with 8 GB RAM and 256 GB SSD
Macbook Pro 14" is $1599 for M3 with 8 GB RAM and 512 GB SSD
Microsoft/Windows products should be at least $400 cheaper than the equivalent MacBook and a lap top with more cores and a fan, shouldn't be compared to an Air with an 8 month old Processor in it, even doing that, they can't beat the aging M3 in single core and everything else will be worse, including editing !
Given the Surface has a fan, why compare it to the MacBook Airs rather than the MacBook Pros?
Surface Laptop 15" is $1299 for Snapdragon X Elite with 16 GB RAM and 256 GB SSD
MacBook Air 15" is $1299 for M3 with 8 GB RAM and 256 GB SSD
Microsoft doesn't mention unified memory in its marketing, so it could be DDR for the Surface which is approximately 50% as efficient as unified memory. If so, the RAM difference in this comparison is negligible. I also wonder what the SSD speed is for the Surface. The M3 MBA SSD has a read speed of 2880 and a write speed of 2108.
8 GB of RAM is still 8 GB of RAM.
The NPU may take 8 GB of RAM just for itself at times. 16 GB was chosen as the minimum for a reason.
I look forward to 3rd-party battery test comparisons between the new Surface and MBA. In my experience, Apple's battery life claims are more accurate, but maybe MS is on the up and up here.
When the third parties get their hands on the Qualcomm device they will find that the numbers have been massaged by Qualcomm one thing that is telling is the wattage used by the Elite SOC when compared to Apples SOC, the Elite is using nearly four times more watts 80 thanboth M3 and the new M4 which are beasts when it comes to speed/power and low wattage. 80 watts only looks good when compared to Intel or AMD laptops.
Which means the actual scores will be lower in a Laptop the public (third party) can buy which means the scores announced by Qualcomm are basically untrue. That will also mean the battery life will be also be different for the worse in some cases. Qualcomm quoting the higher scores from the higher wattage Hmm...?
"The current M3 Pro with 12 cores (as used in the MacBook Pro 16, for example) is only slightly faster—as is the 80-watt reference system from the Snapdragon X Elite. However, these results become even more impressive when you look at the processor's power consumption, which in the case of the iPad Pro 13 was around 15 watts during the benchmark (with a 65-watt power supply). This means that the new M4 SoC consumes even less power despite its higher performance, and the mere fact that such a fast chip can be passively cooled is unlikely to make the competition very happy."
What is the power consumption in the Qualcomm test? most web sites are quoting the high number from 80 watts which means the Apple M3, and the new M4 run at much lower wattage at their peak performance than the Elite 3-4 times lower, the actual peak for the Elite is actually close to 98 watts?
"As the most powerful Snapdragon X Elite chip, the X1E-84-100 is expected to consume up to 80 watts. This will most likely differ from one laptop design to another, with most designs expected to target the 45 W sweet spot." (Qualcomm however announcing the high score from 80 watts?)
Something like this comes up in the world of cars do you measure horsepower at the wheel or the crank? Porsche measures at the wheel which seems to be what Apple is doing.
So which one of them is actually better at processing ML/AI? Surely that should be the primary focus based on the title?
On paper, for the NPU performance of each SoC:
Snapdragon X Plus - 45 TOPS (INT8) Snapdragon X Elite - 45 TOPS (INT8) Apple M3 - 18 TOPS (FP16) Apple M4 - 38 TOPS (INT8)
Interesting. Based on TOPS, which equates to ‘Trillions of Operations a Second’ it would seem that the snapdragon elite does indeed shit all over the base M3 and even the M4 in this specific area.
The NPU was meant to be a big selling point for this particular chip. Looks like they actually delivered on that claim.
For those confused by this whole new world. A rating of 40 TOPS or higher is generally seen as the benchmark where AI (basic AI) can be processed on device as opposed to needing cloud assistance. Obviously there is a lot more to it, but the marketing seems to match the hype on this occasion.
The Elite isn't a direct comparison to base M3 though. It's more an M3 Max competitor with core counts.
And the Qualcomm doesn't fair so well in that comparison.
The Qualcomm is realistically an M2 and M2 Pro competitor. Too far behind to do anything worthwhile. They will improve, but at a lesser rate than Apple who is already too far ahead.
Qualcomm is going 150 miles per hour. Apple is going 200. As this continues, the gap between Apple's lead and Qualcomm will only increase exponentially.
It's best to read it, but here are a few highlights:
Scores shown for the MacBook Air 15" relative to that of the Surface Laptop 15" (Surface Laptop = 1.00x)
CPU Single Thread: 1.15x~1.17x CPU Multi Thread: 0.70x~0.85x NPU Performance: 0.52x GPU Performance: 1.29x~1.31x Battery Life: 0.84x
"The Surface division has always seen its
primary competition as the Mac, and the
new Surface Laptop and the new Surface
Pro tablet absolutely position the Windows
ecosystem stiffly against the MacBook
line. Microsoft knew that the Snapdragon X
Elite and its Arm-based architecture was a
pivotal part of its strategy against the Mac,
both with high performance Oryon cores
and an innovative NPU AI accelerator built
in. The M3 is still a great chip, but the new
Surface Laptop offers an amazing build
quality and physical design, innovative new
AI experiences that Mac hasn’t started to
compete with yet, and does it all for $200
less than an equivalent MacBook Air.
Microsoft and the Surface group have
created the most impressive new laptop to
hit the PC market in a long time. The new
Surface Laptop is managing to combine
a new processor architecture, an updated
emulation layer for Windows on Arm, the
new highest performance PC processor for
this class of laptop, an OS paradigm shift
to AI, and one of the best physical keyboard
and trackpad designs into a single package.
It manages to be both unique and innovative
but also familiar and comfortable for the
modern consumer."
For Battery runtime, You should also include the "Web browsing numbers" for another indicator of battery runtimes - for the M3 Air that is about 15 hours Wireless Web Browsing. For the Surface Laptop that is "Up to 13 hours of battery life" for the 13.8", and "Up to 15 hours of battery life" for the 15".
I personally change the power settings on the Apple Silicon MacBooks I use to "Low Power Mode" - "Only on Battery" to get the most runtime while on battery, the tasks I'm doing don't suffer from this mode - if I'm doing anything intensive (compressing large sets of files, working with large graphics or exporting many files, compiling code), I plug it in to get back to fully performance (or change the setting back to Never). It adds over 2 hours runtime to each charge, possibly as much as 4 more when combined with reducing screen brightness. - Low Power Mode reduces the maximum clock speed of all cores to about half, and may prioritize efficiency cores more. In general you get about a 40-60% reduction in maximum performance.
There are more differences in the CPUs as well that explains some of the performance difference. The M3 8 CPU has 4 Efficiency Cores, and 4 Performance Cores. Has a TDP of 20 Watts The P cores use about 20-25 times as much power as the E Cores and provide somewhere around 3x the performance - this does vary though CPU frequency changes frequently in the newer chips as load changes - an E core doing the same task considering it takes something like 3x longer to complete it has to run would still result in about 1/6th to 1/8th the amount of power used. (https://eclecticlight.co/2023/11/27/evaluating-m3-pro-cpu-cores-1-general-performance/ and some other related pages). The efficiency cores do take longer to do the same task, but use substantially less power to do so. Most background tasks in MacOS use the Efficiency Cores.
The Snapdragon X series does not have efficiency Cores. So the X Plus has 10 Performance Cores - TDP of 23 Watts and the X Elite has 12 Performance Cores has a TDP of something like 45 Watts TDP (I've seen at least one article that said up to 80, but that seems too high, ASUS specifically mentions 45 Watts for their model, https://wccftech.com/snapdragon-x-elite-package-power-can-reach-almost-100w/). If it can go higher, that would start to impact battery runtimes.
The Elite TDP puts it more in the range of the M3 Pro - or over it
The M4 with 4 Efficiency and 4 Performance - the single core performance is higher than the X Series, and the Multi-core exceeds the X Plus, and comes within 5-8% of the X Elite. Its TDP is apparently still 20 Watts.
Tried to do something simple, then kept stumbling on more and more info that made things less simple.
I looked at this, which is very pro Qualcomm, and I don't think it really shows much of an advantage for the Snapdragon X. On aggregate, the Snapdragon X appears no better than the M3 MBA or the Core Ultra 7 155H whatever the model name is. Each system has their advantages and there isn't a clear winner across the board in terms of performance. Signal is calling that a win for Snapdragon, but that is a questionable position imo. It has to be better than x86 for almost everything to get OEMs, Enterprise and consumers to switch.
Meteor Lake is coming, or has already shipped (?), and I expect that will score well on a plethora of battery benchmarks, as well as maintain Intel's other advantages. AMD's next gen is also coming and I expect that to be competitive and outright beat Snapdragon X and Intel in several types of workflows.
I still can't believe the M3 MBA is the fastest at multi-platform Office 365 benchmark, per the Signal results. It's basically the single core performance driving that, but it is MS Office. Or perhaps I should say that office automation apps are like web browsers and they require a lot of single core performance. Working with MS Word is basically an exercise in torture, at least for the sort of documentation I would use it for, if I actually used it. It's really not designed for scientific documentation, so, I don't use it for that. It's really really unpleasant. The only Office app that is good is Excel, as long as you don't load a lot of data into it. Ok, I take it back, Excel sucks just because of that.
On the Procyon AI inference benchmark, I would like to know what the sub-benches consist of, how they are compiling, the prevalence of INT8, INT16 and INT32 benches, etc. Like Signal says themselves, this needs to be sorted out. Moreover, I'm unsure how sensitive trained models are to the NPUs. You need a lot of SIMD throughput to train a model. Clusters of GPUs are needed if you want to train a natural language interface, but to run the trained model, the computational resources are much less. Same for all the other types of inference applications. Training takes a lot of resources. Running the trained model much less so.
How much less is really going to be dependent on the model architecture and what it is designed to do. This stuff is just going to be the wild west for a while.
Nice to see more competition in this area. With TSMC and ARM-based tech, the Apple M-series had a huge advantage to x86 and Intel. Now we see Microsoft being able to offer something similar and add innovation with touch based inputs and CoPilot designed for Windows.
And while things are moving forward in PC-land, we have Apple blocking touch from macOS, penny-pinching on memory, and there is still no LLM in macOS. WWDC better be good, the transition to M4 should move fast, and the basic AI features must get here in Q3.
I'm curious how the X Plus and the X Elite "Thin and Light" are both listed with 23W according to that Anandtech article. the X Elite has 2 more cores than the Plus - so either it must be running at lower clock, or it is actually using a little more power (each core is likely using somewhere around 1W if they are anything like the Apple Performance Cores)
I'm curious how the X Plus and the X Elite "Thin and Light" are both listed with 23W according to that Anandtech article. the X Elite has 2 more cores than the Plus - so either it must be running at lower clock, or it is actually using a little more power (each core is likely using somewhere around 1W if they are anything like the Apple Performance Cores)
These are the latest Snapdragon X SKUs:
These are all one SoC chip, binned into 4 SKUs. I would guess the X1E-80, X1E-78 and X1P-64 are the "23W" SKUs. "23W" in quotes because it seems Qualcomm has just dropped that distinction, or stopped using TDP as a SKU distinction all together. Will have to wait for measurements.
And "thin and light" isn't much of a category these days either? Like, these Surface Laptops are as thick as the MBP14 or MBP16. Maybe the categories these days are consumer laptops, business laptops, and gaming/workstation laptops? The MBA is basically in a unique category.
In real life applications unified memory is going to make all the difference. Constantly accessing off-chip memory to copy it to and from various cores wastes cycles and power. The pure core peformance may be slightly better on the Snapdragon but Apple’s elegant design ultimately delivers considerably better performance where it counts.
AI should change it's [sic] moniker to BS, because that's all it is!
Oh, I don't know. AI can certainly churn out awful text – its article writers could really use an editor – but as a general source of Apple-related news, AI remains useful for...
Snapdragon Scores inside in an Actual OEM Computer better at what? High M1 low M2 with 2-4 more cores running at 3-6 times the wattage of a iPad Pro M4 (15 watts).
It’s a minefield for anyone buying a PC (so many utter pieces of crap out there) but great to see some decent competition for the MBA. Apple’s shortcoming as always is that stinginess with soldered in GPU’s, hard drives and RAM. It’s always annoyed me that these otherwise brilliant machines from the world’s second richest company are hobbled by their tightarseness.
Comments
The 23 W version is what's in most devices.
From Anandtech (Oct. 2023)
"Thin and Light" version has a 23W TDP:
It won't be Qualcomm they have no control over the AI part directly however it will be a problem for them going forward because Recall Microsoft is on the case.
https://www.notebookcheck.net/Apple-M4-SoC-analysis-AMD-Intel-and-Qualcomm-currently-don-t-stand-a-chance.839332.0.html
The NPU may take 8 GB of RAM just for itself at times. 16 GB was chosen as the minimum for a reason.
Which means the actual scores will be lower in a Laptop the public (third party) can buy which means the scores announced by Qualcomm are basically untrue. That will also mean the battery life will be also be different for the worse in some cases. Qualcomm quoting the higher scores from the higher wattage Hmm...?
https://www.notebookcheck.net/Apple-MacBook-Pro-16-2023-M3-Pro-review-Efficiency-before-performance.772025.0.html Wattage used for Apple in a M3 is laptop 26-27 watts.
The following Quotes from 2 links.
"The current M3 Pro with 12 cores (as used in the MacBook Pro 16, for example) is only slightly faster—as is the 80-watt reference system from the Snapdragon X Elite. However, these results become even more impressive when you look at the processor's power consumption, which in the case of the iPad Pro 13 was around 15 watts during the benchmark (with a 65-watt power supply). This means that the new M4 SoC consumes even less power despite its higher performance, and the mere fact that such a fast chip can be passively cooled is unlikely to make the competition very happy."
What is the power consumption in the Qualcomm test? most web sites are quoting the high number from 80 watts which means the Apple M3, and the new M4 run at much lower wattage at their peak performance than the Elite 3-4 times lower, the actual peak for the Elite is actually close to 98 watts?
"As the most powerful Snapdragon X Elite chip, the X1E-84-100 is expected to consume up to 80 watts. This will most likely differ from one laptop design to another, with most designs expected to target the 45 W sweet spot." (Qualcomm however announcing the high score from 80 watts?)That maybe that was what the guy at semiaccurate was talking about? https://www.semiaccurate.com/2024/04/24/qualcomm-is-cheating-on-their-snapdragon-x-elite-pro-benchmarks/
Something like this comes up in the world of cars do you measure horsepower at the wheel or the crank? Porsche measures at the wheel which seems to be what Apple is doing.
Never realized that Microsoft just completely and blatantly flat out COPIED the old MacBook Air design.
What a rip!
Talk about disingenuous. Apple should bring back the old Stee Jobs "Redmond, start your photocopiers!" banners to WWDC.
The Elite isn't a direct comparison to base M3 though. It's more an M3 Max competitor with core counts.
And the Qualcomm doesn't fair so well in that comparison.
in some ways, the M3 base beats it.
https://beebom.com/early-snapdragon-x-elite-benchmarks-cant-beat-apple-m3/
Heck, even the iPad beats the elite now.
The Qualcomm is realistically an M2 and M2 Pro competitor. Too far behind to do anything worthwhile. They will improve, but at a lesser rate than Apple who is already too far ahead.
Qualcomm is going 150 miles per hour. Apple is going 200. As this continues, the gap between Apple's lead and Qualcomm will only increase exponentially.
https://signal65.com/wp-content/uploads/2024/05/NewSurfaceLaptop2024_Signal65LabInsights.pdf
It's best to read it, but here are a few highlights:
Scores shown for the MacBook Air 15" relative to that of the Surface Laptop 15" (Surface Laptop = 1.00x)
CPU Single Thread: 1.15x~1.17x
CPU Multi Thread: 0.70x~0.85x
NPU Performance: 0.52x
GPU Performance: 1.29x~1.31x
Battery Life: 0.84x
"The Surface division has always seen its primary competition as the Mac, and the new Surface Laptop and the new Surface Pro tablet absolutely position the Windows ecosystem stiffly against the MacBook line. Microsoft knew that the Snapdragon X Elite and its Arm-based architecture was a pivotal part of its strategy against the Mac, both with high performance Oryon cores and an innovative NPU AI accelerator built in. The M3 is still a great chip, but the new Surface Laptop offers an amazing build quality and physical design, innovative new AI experiences that Mac hasn’t started to compete with yet, and does it all for $200 less than an equivalent MacBook Air.
Microsoft and the Surface group have created the most impressive new laptop to hit the PC market in a long time. The new Surface Laptop is managing to combine a new processor architecture, an updated emulation layer for Windows on Arm, the new highest performance PC processor for this class of laptop, an OS paradigm shift to AI, and one of the best physical keyboard and trackpad designs into a single package. It manages to be both unique and innovative but also familiar and comfortable for the modern consumer."
https://nanoreview.net/en/cpu-compare/qualcomm-snapdragon-x-elite-vs-apple-m4
even the iPad base m4 beats the top of the line Qualcomm.
I personally change the power settings on the Apple Silicon MacBooks I use to "Low Power Mode" - "Only on Battery" to get the most runtime while on battery, the tasks I'm doing don't suffer from this mode - if I'm doing anything intensive (compressing large sets of files, working with large graphics or exporting many files, compiling code), I plug it in to get back to fully performance (or change the setting back to Never). It adds over 2 hours runtime to each charge, possibly as much as 4 more when combined with reducing screen brightness.
- Low Power Mode reduces the maximum clock speed of all cores to about half, and may prioritize efficiency cores more. In general you get about a 40-60% reduction in maximum performance.
There are more differences in the CPUs as well that explains some of the performance difference.
The M3 8 CPU has 4 Efficiency Cores, and 4 Performance Cores. Has a TDP of 20 Watts
The P cores use about 20-25 times as much power as the E Cores and provide somewhere around 3x the performance - this does vary though CPU frequency changes frequently in the newer chips as load changes - an E core doing the same task considering it takes something like 3x longer to complete it has to run would still result in about 1/6th to 1/8th the amount of power used. (https://eclecticlight.co/2023/11/27/evaluating-m3-pro-cpu-cores-1-general-performance/ and some other related pages). The efficiency cores do take longer to do the same task, but use substantially less power to do so. Most background tasks in MacOS use the Efficiency Cores.
The Snapdragon X series does not have efficiency Cores.
So the X Plus has 10 Performance Cores - TDP of 23 Watts
and the X Elite has 12 Performance Cores has a TDP of something like 45 Watts TDP (I've seen at least one article that said up to 80, but that seems too high, ASUS specifically mentions 45 Watts for their model, https://wccftech.com/snapdragon-x-elite-package-power-can-reach-almost-100w/). If it can go higher, that would start to impact battery runtimes.
The Elite TDP puts it more in the range of the M3 Pro - or over it
The M4 with 4 Efficiency and 4 Performance - the single core performance is higher than the X Series, and the Multi-core exceeds the X Plus, and comes within 5-8% of the X Elite. Its TDP is apparently still 20 Watts.
Tried to do something simple, then kept stumbling on more and more info that made things less simple.
Meteor Lake is coming, or has already shipped (?), and I expect that will score well on a plethora of battery benchmarks, as well as maintain Intel's other advantages. AMD's next gen is also coming and I expect that to be competitive and outright beat Snapdragon X and Intel in several types of workflows.
I still can't believe the M3 MBA is the fastest at multi-platform Office 365 benchmark, per the Signal results. It's basically the single core performance driving that, but it is MS Office. Or perhaps I should say that office automation apps are like web browsers and they require a lot of single core performance. Working with MS Word is basically an exercise in torture, at least for the sort of documentation I would use it for, if I actually used it. It's really not designed for scientific documentation, so, I don't use it for that. It's really really unpleasant. The only Office app that is good is Excel, as long as you don't load a lot of data into it. Ok, I take it back, Excel sucks just because of that.
On the Procyon AI inference benchmark, I would like to know what the sub-benches consist of, how they are compiling, the prevalence of INT8, INT16 and INT32 benches, etc. Like Signal says themselves, this needs to be sorted out. Moreover, I'm unsure how sensitive trained models are to the NPUs. You need a lot of SIMD throughput to train a model. Clusters of GPUs are needed if you want to train a natural language interface, but to run the trained model, the computational resources are much less. Same for all the other types of inference applications. Training takes a lot of resources. Running the trained model much less so.
How much less is really going to be dependent on the model architecture and what it is designed to do. This stuff is just going to be the wild west for a while.
Now we see Microsoft being able to offer something similar and add innovation with touch based inputs and CoPilot designed for Windows.
And while things are moving forward in PC-land, we have Apple blocking touch from macOS, penny-pinching on memory, and there is still no LLM in macOS.
WWDC better be good, the transition to M4 should move fast, and the basic AI features must get here in Q3.
Disruption makes it all interesting again.
I'm curious how the X Plus and the X Elite "Thin and Light" are both listed with 23W according to that Anandtech article. the X Elite has 2 more cores than the Plus - so either it must be running at lower clock, or it is actually using a little more power (each core is likely using somewhere around 1W if they are anything like the Apple Performance Cores)
These are all one SoC chip, binned into 4 SKUs. I would guess the X1E-80, X1E-78 and X1P-64 are the "23W" SKUs. "23W" in quotes because it seems Qualcomm has just dropped that distinction, or stopped using TDP as a SKU distinction all together. Will have to wait for measurements.
And "thin and light" isn't much of a category these days either? Like, these Surface Laptops are as thick as the MBP14 or MBP16. Maybe the categories these days are consumer laptops, business laptops, and gaming/workstation laptops? The MBA is basically in a unique category.
Wait. Which "AI" are you talking about?
https://browser.geekbench.com/search?utf8=%E2%9C%93&q=+Samsung+X+Elite+Qualcomm+Oryon
https://browser.geekbench.com/search?utf8=%E2%9C%93&q=inspiron+snapdragon+Elite+Qualcomm+Oryon
https://browser.geekbench.com/search?utf8=%E2%9C%93&q=+lenovo+X+Elite+Qualcomm+Oryon
https://browser.geekbench.com/search?utf8=%E2%9C%93&q=ASUS+X+Elite+Qualcomm+Oryon
https://browser.geekbench.com/mac-benchmarks Single Core
https://browser.geekbench.com/mac-benchmarks Multi Core