Qualcomm has compared the CPU to the new M3 chip, similarly claiming that its chip outperforms the leading ARM-compatible competitor for single-threaded CPU performance. Qualcomm also says the Oyron chip is 21% faster than the M3 in multi-core performance. That’s quite the bold claim.
Late last year, I wrote about Microsoft’s plans for its upcoming Surface Pro 10 and Surface Laptop 6, which are expected to feature some major upgrades in the form of refreshed designs, new chips, and next-gen NPU silicon that will power some of the upcoming advanced AI features Microsoft is building into the next version of Windows.
I know that the June wave will include the Surface Pro 10 and Surface Laptop 6, as that’s when the first Qualcomm X Series PCs are expected to start shipping across the board. These PCs depend on the Windows Germanium platform release, which won’t RTM until April, making it too late for any PCs launching that month.
According to my sources, Microsoft intends to ship a commercial focused version of Surface Pro and Surface Laptop that feature the current in-market designs, but with newer Intel 14th-gen chips and a scoped set of features designed for commercial customers.
https://www.windowscentral.com/hard...face-pc-announcements-may-take-place-in-marchAs I understand it, consumers will be able to purchase these commercial focused devices if they wish, but the recommendation is to wait for the true Surface Pro 10 and Surface Laptop 6 coming in June with next-gen Arm chips and updated designs.
https://www.theverge.com/2024/4/8/2...k-air-surface-arm-qualcomm-snapdragon-x-eliteMicrosoft is so confident in these new Qualcomm chips that it’s planning a number of demos that will show how these processors will be faster than an M3 MacBook Air for CPU tasks, AI acceleration, and even app emulation. Microsoft claims, in internal documents seen by The Verge, that these new Windows AI PCs will have “faster app emulation than Rosetta 2” — the application compatibility layer that Apple uses on its Apple Silicon Macs to translate apps compiled for 64-bit Intel processors to Apple’s own processors.
Core Count | Total Cache | Max Clocks | Dual-Core Boost | GPU | NPU | |
---|---|---|---|---|---|---|
Snapdragon X Elite X1E-84-100 | 12 | 42MB | 3.8GHz | 4.2GHz | 4.6TFLOPs | 45 TOPS |
Snapdragon X Elite X1E-80-100 | 12 | 42MB | 3.4GHz | 4.0GHz | 3.8TFLOPs | 45 TOPS |
Snapdragon X Elite X1E-78-100 | 12 | 42MB | 3.4GHz | - | 3.8TFLOPs | 45 TOPS |
Snapdragon X Plus X1P-64-100 | 10 | 42MB | 3.4GHz | - | 3.8TFLOPs | 45 TOPS |
Snapdragon X Elite | Snapdragon X Plus | |
---|---|---|
Geekbench 6.2 | 2850-2900 single-core 15,100-15,400 multi-core | 2,400-2,425 single-core 12,800-13,100 multi-core |
Cinebench 2024 | 126-128 single-core 1,140-1,200 multi-core | 107-109 single-core 825-845 multi-core |
Chrome (Beta) - Speedomter 2.1 | 460-500 | 410-430 |
Chrome (Beta) - Jetstream 2.1 | 330-340 | 280-290 |
UL Procyon AI | 1,750-1,800 | 1,750-1,800 |
UL Procyon Office | 6,500-6,900 | 5,700-5,900 |
PCMark 10 | 13,500-14,100 | 12,500-12,800 |
Blender | 430-470 | 340-360 |
3DMark | 41.9-44FPS | 37.1-38.5FPS |
Qualcomm Is Cheating On Their Snapdragon X Elite/Pro Benchmarks
So what are they cheating on? The short version is that the numbers that they are showing to the press and are not achievable with the settings they claim. Qualcomm is showing a different set of numbers to OEMs and these also are not achievable with the settings they claim. This information comes from two tier 1 OEMs and other sources. (Note to Qualcomm: No it wasn’t him, really, we knew long before last week) SemiAccurate is 100% confident in saying that some of the numbers Qualcomm was showing off can not be reproduced with the settings they claim.
Step forward a little in time. After OEMs got initial samples and made something close to the final designs, SemiAccurate got reports of poor performance. By poor we mean far sub-50% of the numbers Qualcomm was telling them in the technical docs and presentations. Trying to help we told some Qualcomm engineers about the findings and asked if there were any known issues with the silicon that would cause this. They repeated the claim that the silicon was clean, something we still believe to be true, but the state of WART was horrific. We also believe this to be true. They asked us to pass back to the OEMs that likely culprits were cooling and having the right benchmark build that was ARM native. Both are fair comments and we passed that back. Then silence for a while.
Later, with more Snapdragon X Elite samples in the wild and many more revisions of WART, we got similar reports from OEMs and another Tier 1. Both reported numbers that were nowhere close to what Qualcomm promised. How not close? Above 50% this time but one used the term ‘Celeron’ to describe performance. The claims of better than Apple’s Rosetta 2 x86 emulation are clearly not real on what is probably the release hardware and software. Actually the silicon emulation may be better but everything else is unquestionably not.
A deep source at Qualcomm told us that the benchmarks were cheats, told us how they were cooked, and told us that Qualcomm was well aware of it. This same technique meant the numbers looked far better than they could be on a non-trivial set of tests. Ironically some other benchmarks could have looked much better than those presented if Qualcomm adequately disclosed their testing details.
We realize it is anecdotal but Qualcomm won’t let anyone do independent testing on these parts. Windows on ARM is still said to be in a painfully awful state, or perfect by Microsoft standards. We are weeks away from the official announcement on May 20th at Build and the laptops are going on sale in June. If you take build time, shipping, and all the rest into account, what software and hardware we have now is what you are going to see in the stores, or really really really close to it. There is no excuse for not letting independent sites test the hardware other than the obvious bits this story is about. Qualcomm wants to control the message and has a host of social media influencers primed and ready to go. Curious that. Or not.
Because the claims can’t be independently verified, the influencers will parrot back how ‘good’ the product is based on something we know is not true. Between the initial publicity blitz and the time things go on sale, the cheating values will become ‘truth’ and few if any will question them months down the road. Cynical hardly begins to describe this mechanism. The most interesting part is whether OEMs will parrot back numbers they know they can’t replicate or release their own. Time will tell.
https://semiaccurate.com/2024/04/24/qualcomm-is-cheating-on-their-snapdragon-x-elite-pro-benchmarks/What more is there to say? Qualcomm has been making promises about disclosure and testing for about six months now. They have broken every one of those and at least to SemiAccurate, repeatedly lied. Because of this the digging we have been doing since before the Hawaii disclosure, some of which lead to this older article, has kept us focused on how these benchmark claims were cheating. While we don’t use this term lightly, we are absolutely sure that Qualcomm is doing exactly that with their numbers. Worse yet they know the reasons why the numbers are tainted but continue to present them to OEMs and the press. The Qualcomm Snapdragon X Elite/Pro SoCs still seem to be very good silicon, everything above that is the polar opposite.
As for power consumption, Qualcomm provides them in a variety of scenarios; for simplicity, we will list them for a CPU-only load. There’s a lot of variance in silicon devices, so Qualcomm lists the values as they should be able to be achieved by 95% and 50% of the manufactured chips, respectively. In practice, the better parts will simply get binned as the higher-end SKUs, though, so the distinction doesn’t have any significant implications in reality. However, it’s still important to keep in mind.
For context, Apple’s M3 Pro (built on TSMC’s 3 nm process node) consumes around 42W under CPU loads, and the older M2 Pro (TSMC’s 4 nm process, similar – if not identical – to the one used for the Snapdragon X1 series) uses around 55W. Intel’s Core Ultra 7 155h, meanwhile, uses around 80W.
X1 Elite (X1E84100) X1 Elite (X1E80100) X1 Plus (X1P64100) Total package power (95% parts) 98.50W 52.92W 42.52W Total package power (50% parts) 82.33W 43.40W 35.01W