This is Part 4 of a series. Part 1 can be found here, Part 2 can be found here, and Part 3 can be found here.
Power:
Qualcomm talked about the power use of the X2 in a roundabout way, please welcome the new INPP metric. You know when there is a new metric to define how power is measured, something went wrong somewhere in development. INPP may be an exception but it is hard to tell given how strongly Qualcomm avoided giving out basic numbers like TDP and instead only wanted to talk INPP. What is INPP? Idle Normalized Platform Power, a metric that is more or less tangential to TDP.
INPP isn’t what Qualcomm wants it to be
We won’t go into the excuses about why Qualcomm declined to use industry standard metrics to talk about their chip but given the people we know who measure such things at the company, we don’t buy the excuses. That said, INPP is basically the max SYSTEM power draw minus the idle SYSTEM power draw, emphasis ours. It is actually a good measure but it has two crippling flaws.
First off it is system specific, your Asus notebook will have a different INPP than your HP notebook even if they have the same chip and components. Change your SSD or RAM and voila, different INPP. It also doesn’t measure performance so if that Asus laptop is 10% faster than the HP but has a 3% higher INPP, that means… it means it will be used as a club by unethical marketers for all the wrong reasons.
More importantly it tells you NOTHING about the thing we all wanted to know about, how much power does the CPU pull? Good luck teasing that from the INPP without having the ability to put taps on memory power lines. In short it doesn’t answer the question, and since that is all that Qualcomm wanted to talk about, you have to wonder why? Qualcomm is really good at self-inflicted wounds in the PC space and this is another example of why. To date Qualcomm has not disclosed the TDP of the X2, a sad commentary on an otherwise solid piece of silicon.
Cluster-Level Multi-Level Boost or two core boost
That brings us to one of the things we talked about 10 pages ago, the CPU frequency and boost. We said the Prime cores were running at 4.4/5.0GHz base/one core turbo clocks. Qualcomm has now updated that to say the base clocks are 4.45GHz and the 5.0GHz is for one core active only but the previously denied middle ground exists. Now they say that middle ground would be 4.8GHz for two cores, 4.47GHz for three active cores, and 4.45GHz if four or more cores are active. This is per cluster and remember there are two Prime clusters on the X2. Our biggest issue here, besides the silly ‘surprise’ reveal, is, why call out the 20MHz ‘gain’ when only 3C are active? If this was a P54C that would be impressive but good luck seeing the gains from that ‘boost’ in any real world workload. Snarky comments aside the 2C boost is welcome and a good thing.
Performance:
With that clock update we come to performance. Qualcomm put out a lot of numbers for the SDX2E as a chip and the units themselves. Most were quite impressive, as we keep saying the silicon is likely very good but we have no way of testing it. Before you point out that Qualcomm ran extensive ‘benchmarking’ sessions at their tech day, at Snapdragon Summit, and in past years for the X1, no they didn’t.
The sessions that Qualcomm called benchmarking were nothing like benchmarking. Once again they set up a lot of reference designs, something that is not a problem, and clearly marked whether or not they were thermally constrained or not, again a good thing. They put out a clear spreadsheet of expected scores for selected benchmarks, and had people there to answer questions if needed. Again these are all good things. The benchmarks themselves were selected Windows only commonly used benchmarks, other than the Windows only bit, quite fair.
Things went off the rails from there. Each station had a single test on it, and you could run that test by pressing start and waiting for the results. No changing of settings, no loading your own benchmarks, no nothing, only press the button and get a number which to be fair did match the given expected numbers. This wasn’t a benchmarking session, it was a demo session.
Qualcomm simply does not grasp the needs and expectations of the PC community and they keep making painful missteps because of it. Normally we would point this out to them at the event and explain the issues, suggest improvements, and wait for the next time to actually benchmark things. After four or five times of doing that without change, we gave up and had no more FSCKs to give on their ‘benchmarking’. It isn’t. Qualcomm doesn’t get it, please stop it with the footgun, it is painful to watch.
At this point we would like to point out how either of our two SDX1 systems fared when tested independently against the expected numbers but to date we can’t get either working. The first one got a desktop up after about 18 months of trying. If you open a browser and move the scroll wheel, the touchpad doesn’t work among other things, the whole system goes down. The second system had most of the hardware not recognized so after a few days of trying we gave up to write this tome. We may eventually get it to but we doubt it. There was no opportunity to actually benchmark the X2.
This is a big problem for Qualcomm. Why? Because as we said earlier, Nvidia is coming. If you are a developer and are thinking of deploying Qualcomm AI hardware like the AI100 or the upcoming generation of parts, what do you do? You buy a consumer system and kick the tires so get a feel for the software and paradigms before you commit to 6+ digits of servers.
To do this, you will buy an SDX1 system and quickly discover it doesn’t work. Not work right, but is not a functional product for any sort of serious development. If you buy an Nvidia Spark system, it will not work either for different reasons but those will be fixed. On the other hand Qualcomm has made it very clear they won’t fix their problems. Qualcomm just lost the AI developer community, or at least had no chance of making inroads into anyone serious.
The most painful part of all of this is that Qualcomm has the tools to fix things. The first demo of the X Elite in the fall of 2023 was done on a fully working internal Linux distro because Windows was far from ready. Both states are still true but Qualcomm won’t admit to having working software much less release it. Serious developers know this and get the message loud and clear, avoid Qualcomm PCs. Nvidia faced the same issues so they put out the Spark with Linux, Qualcomm made excuses and still does.
X86:
The last bit is the x86 emulation in the Snapdragon X Elite line. When the X1 was launched Qualcomm made a big deal about it, this time it was more of an afterthought. Compatibility numbers were completely absent but things were claimed to me much better than the X1 launch. They even pointed out that the compatibility on X1 has gotten much better since release, a very true statement. That said it isn’t good much less good enough. Much of the ‘compatibility’ updates come from sticking points being natively recoded for ARM, something that is not x86 compatibility.
Moving back to the good hardware, crap software, no software enablement theme, we come to the heart of the issue. The core Nuvia team which Qualcomm bought was the core Apple team. They were responsible for the M1 and M2 among other CPUs, both of which have very solid x86 emulation. You don’t hear about ARM Apple users complaining about x86 Windows emulation do you? You don’t hear about Apple bragging about x86 compatibility figures that start with an 8 do you? Nope, it just works.
For Qualcomm we have, err, a less optimal situation. This means that the team either collectively got closed head wounds or forgot all they know about CPU design. SemiAccurate can state that neither is true, they are fine, heads and internal knowledge bases. In fact the emulation hardware and firmware in the X1 is a generation ahead of that in the Apple products so it is better right? Nope, it is crap. Why? Microsoft incompetence is a big part of the problem and Qualcomm’s lack of software enablement means there will be no real progress. This was a golden opportunity squandered by all involved.
Overall:
There are a lot more details to the new Snapdragon X2 Elite SoC, things like a new sensor hub, VPU, user ID features, and a new on-die network. It all adds up to a chip that SemiAccurate likes a lot, the silicon appears really good but, well, you can’t tell that from the released products. Unfortunately Qualcomm has gone out of their way to make sure the product that you can buy is unacceptable. Toss in Microsoft’s software incomptence and you have something that should be avoided. The X1 should be avoided at all costs, it simply doesn’t work right and after 18 months of trying, is unlikely to ever be made right. The X2 may fix a few of these critical flaws but given the starting point, it is unlikely to ever be a usable product. And this is directly Qualcomm’s fault.
Why do we say this? A chip is a pretty piece of silicon without software support and the X Elite line utterly lacks any sort of software enablement. The X1 CPU was delayed by about 9 months while Microsoft got it’s act together. Yes that is a joke, they didn’t get their act together but they released WART anyway, and it was a debacle. Remember the release, pulling, and whimpering re-release of Recall? And to date the ‘AI’ software that was going to change everything is still under the category of, “Maybe someday we will have something that doesn’t suck”.
Microsoft and Qualcomm pushed the AI message hard for different reasons. Microsoft just wanted to offload datacenter costs, nothing more. Qualcomm is acting like a Microsoft sock puppet without understanding why or what the consequences are. They followed the Microsoft messaging and sales were abysmal, AI just didn’t sell even with a free large screen TV tossed in. When they switched to their own messaging, basically battery life, sales picked up to slightly above abysmal, a big gain actually even if it wasn’t real. How do we know about this messaging change? Qualcomm said it outright to a room full of press at the Snapdragon Summit a bit over a month ago. We applaud their honesty and hope these are the green shoots of a growing spine.
But back to the X1, it simply doesn’t work right. It isn’t a PC, it is a good CPU with a barely working software wrapper in a closed garden. Fixes that you expected to come in short order are no nearer 18 months later than they were at launch. This isn’t by chance. Extensive chats with over a dozen people at Qualcomm made it very clear that while they get it, there is no chance they will actually do anything about it much less fix the problems. Excuses ranged from the outright laughable to technically impossible with 3rd parties occasionally blamed. It was painful.
So with that track record in mind, we can honestly say the Snapdragon X Elite is not a functional product despite the silicon. With the X2 impending, will it be better? The silicon is unquestionably first class but it is still being put into a product that isn’t up to the standards of crap, and has massive unfixable security problems. This isn’t an overstatement, the one fix we can point to is that the X2 uses ACPI which will lessen a big pain point of the X1 but that is about it. The lack of software enablement, the antagonistic responses from Qualcomm to developers just asking questions, and missing documentation is still there. Don’t touch Snapdragon X2 Elite products until it is very clear Qualcomm has changed their ways for good. Maybe the X3 will be in functional products upon release.S|A
This is Part 4 of a series. Part 1 can be found here, Part 2 can be found here, and Part 3 can be found here.
