Nuvia: don’t hold your breath

dada_dave said:
🙃 max frequency: 12MHz. I’d say something went wrong! I certainly hope so.
Click to expand...
Yes I’m sure there is a reasonable explanation. There is no way it’s that a bad. I also doubt the 1600 GB score mentioned in the article earlier. These machines are definitely not finished.
 
Jordanmarkstone said:
Interesting comment from Andrei about power usage/thermals.


Leaves us more in the dark 😄
Click to expand...
Thanks for pointing this out. I had always assumed people were referring to the thermal envelope of the laptop, rather than the chips itself

I notice he doesn’t challenge the article itself which quotes the 1600 GB score. Perhaps he wouldn’t be allowed to? Also credit to notebookcheck for not saying it’s a reasonable score.

OT also saw he takes a (reasonable) shot at Intel here!
 
Jimmyjames said:
Thanks for pointing this out. I had always assumed people were referring to the thermal envelope of the laptop, rather than the chips itself

I notice he doesn’t challenge the article itself which quotes the 1600 GB score. Perhaps he wouldn’t be allowed to? Also credit to notebookcheck for not saying it’s a reasonable score.
Click to expand...
Agreed: with preproduction numbers that leak you can get some wild stuff. I wouldn’t take numbers too seriously until right before launch. I remember some of the valid preproduction Alder Lake numbers were trash but were running in low power states with like only 4GB of particularly slow DDR4 memory or something ridiculous. Now why was that config being tested? Who knows, could be any number of reasons, but OEMs like to fiddle around with stuff and some of the numbers are just fantasy, high and low. So no real point in challenging them and as you say he almost certainly can’t.

Jimmyjames said:
OT also saw he takes a (reasonable) shot at Intel here!
Click to expand...


Eugh … I thought we were past doing shit like this. Evidently not.
 
dada_dave said:
Eugh … I thought we were past doing shit like this. Evidently not.
Click to expand...

There is money involved, of course we are not past it. Intel cheating on SPEC is a well known thing. And if they are not cheating, they are cherry-picking to extreme.
 
leman said:
There is money involved, of course we are not past it. Intel cheating on SPEC is a well known thing. And if they are not cheating, they are cherry-picking to extreme.
Click to expand...
Cherry-picking I can accept - well to a degree, I remember when the M1 first came out Intel’s marketing was pretty desperate and I’ll admit it irked me that some tech journalists couldn’t see through the obvious hatchet job. But again what irked me was more about how some journalists couldn’t tell the difference and less about Intel trying to present themselves in the best light, desperate as it was. As much as our community likes to discuss the meaning and merits of individual numbers and treat them as meaningful, benchmarking is essentially the advertising business and trying to accentuate your advantages and downplay your competitors in advertising is normal.

But cheating on benchmarks especially when it is so easily disproven is counterproductive. I mean if all it takes is renaming or wrapping the function, it’s going to be discovered and quickly what you’ve done. And then it functions against you. Extreme cherry-picking can too, Apple has taken more than a few deserved lumps for that, but cheating is another level precisely because it tries to keep the reality distortion past the control of your advertising department.

Then again maybe tech journalism is so dead that nobody cares anymore. I dunno, I hope not.
 
iMac base model M3, (4p/4e @4.06GHz): 3044sc, 11669mc
Qualcomm Oryon Elite X (12p @4.01Ghz): 2574sc, 12562mc

So the 8 core iMac is about 7% slower in multicore than the 12 core beast Nuvia chip.
 
Again
1. These leaked benchmarks have clock speeds that match the lower "23W Device TDP" model. We have no idea what "Device TDP" means, but we do know this is the SKU for fanless MacBook Air-like machines. Maybe these leaked results are for a fanless machine? Who knows!
2. We still don't know power usage or boost behaviour (the most important factor in judging the success of this product).
3. Geekbench 6 for aarch64 Windows typically performs MUCH worse than Linux and macOS. I know this from personal experience (look at my Geekbench profile, I mess around with a lot of hardware for fun). 2547 points under Windows could be between 10 and 15% higher in Linux (2801 or 2929 points, with 2929 being just 4-6% behind the average base M3 scores).

I reject the premature narrative that this chip has failed based on flaky context-less benchmarks being deliberately interpreted in the least flattering way 🙂
 
Aaronage said:
Again
1. These leaked benchmarks have clock speeds that match the lower "23W Device TDP" model. We have no idea what "Device TDP" means, but we do know this is the SKU for fanless MacBook Air-like machines. Maybe these leaked results are for a fanless machine? Who knows!
2. We still don't know power usage or boost behaviour (the most important factor in judging the success of this product).
3. Geekbench 6 for aarch64 Windows typically performs MUCH worse than Linux and macOS. I know this from personal experience (look at my Geekbench profile, I mess around with a lot of hardware for fun). 2547 points under Windows could be between 10 and 15% higher in Linux (2801 or 2929 points, with 2929 being just 4-6% behind the average base M3 scores).

I reject the premature narrative that this chip has failed based on flaky context-less benchmarks being deliberately interpreted in the least flattering way 🙂
Click to expand...
I don’t think anyone is taking the numbers too seriously yet. We all recognize that preproduction numbers can be wonky and that we’re missing context. Yeah okay there’s a little ribbing at Nuvia’s expense in comparison to Apple benchmarks, but of course we know that these are early numbers. As an extreme example from above, while Qualcomm’s Adreno GPUs tend to score very badly on OpenCL GB, I very much doubt they’ll get a score of “10” in the finished product, which I didn’t even think a modern GPU could score that low and still finish the benchmark! So, hey!, at least I learned something new! 🙃
 
Aaronage said:
Again
1. These leaked benchmarks have clock speeds that match the lower "23W Device TDP" model. We have no idea what "Device TDP" means, but we do know this is the SKU for fanless MacBook Air-like machines. Maybe these leaked results are for a fanless machine? Who knows!
2. We still don't know power usage or boost behaviour (the most important factor in judging the success of this product).
3. Geekbench 6 for aarch64 Windows typically performs MUCH worse than Linux and macOS. I know this from personal experience (look at my Geekbench profile, I mess around with a lot of hardware for fun). 2547 points under Windows could be between 10 and 15% higher in Linux (2801 or 2929 points, with 2929 being just 4-6% behind the average base M3 scores).

I reject the premature narrative that this chip has failed based on flaky context-less benchmarks being deliberately interpreted in the least flattering way 🙂
Click to expand...
I’m not stating it’s a failure. I’m stating Qualcomm came out guns blazing about Apple Silicon beating performance, and even M3 beating performance, and that is unproven. If they wanted to be judged on a final product, they should have waited until they released one.

If they are free to use suspect pre-release products for marketing, then why wouldn’t we be able to criticise those numbers? Your response seems like an overreaction. The press has been nothing but fawning over these as yet unreleased and untested chips.

I reject the narrative that they, having started this nonsense, should be free from its consequences. I also don’t see why a consumer should care that Windows scores lower on Geekbench. That is the main product they are releasing.
 
Last edited:
Jimmyjames said:
I also don’t see why a consumer should care that Windows scores lower on Geekbench. That is the main product they are releasing.
Click to expand...

On that last part, if that’s true it’s because the issue may be within GB rather than reflective of the actual performance of the chips in a Windows device.
 
dada_dave said:
On that last part, if that’s true it’s because the issue may be within GB rather than reflective of the actual performance of the chips in a Windows device.
Click to expand...
Does anyone actually know if that's true? Is it GB or is it the OS? Have we actually seen tests that isolate the problem to GB vs the problem being windows or linux?
 
amonduin said:
Does anyone actually know if that's true? Is it GB or is it the OS? Have we actually seen tests that isolate the problem to GB vs the problem being windows or linux?
Click to expand...
@Aaronage posted results earlier but those were through VMWare so it could be a combination of Windows through a VM having performance issues which Linux through a VM does not share. That said, he alluded to other results that pointed in the same direction. I haven’t had time to check.

While GB has had issues with Windows before, I think thread scheduling for multicore tests, my impression was those problems were fixed awhile ago and wouldn't be relevant to the single core results anyway. @Aaronage’s post was the first I’d heard of it having a problem specifically with Windows on ARM. Therefore while it’s not out of the realm of possibility that the problem is in GB, yeah it could also be Windows on ARM has performance issues. I do seem to recall that in general SPEC ran faster on Linux than Windows, but I don’t think it was substantial. I could be wrong.

In summary: 🤷‍♂️

🙃
 
What were the same app compasisons on windows v intel-Mac like? Is windows about the same performance as Mac?
 
dada_dave said:
@Aaronage posted results earlier but those were through VMWare so it could be a combination of Windows through a VM having performance issues which Linux through a VM does not share. That said, he alluded to other results that pointed in the same direction. I haven’t had time to check.

While GB has had issues with Windows before, I think thread scheduling for multicore tests, my impression was those problems were fixed awhile ago and wouldn't be relevant to the single core results anyway. @Aaronage’s post was the first I’d heard of it having a problem specifically with Windows on ARM. Therefore while it’s not out of the realm of possibility that the problem is in GB, yeah it could also be Windows on ARM has performance issues. I do seem to recall that in general SPEC ran faster on Linux than Windows, but I don’t think it was substantial. I could be wrong.

In summary: 🤷‍♂️

🙃
Click to expand...
This is my Surface Pro X (8cx/SQ2) native Windows vs WSL https://browser.geekbench.com/v6/cpu/compare/5186796?baseline=5053384
About 12% faster ST in WSL
 
Back
Top