I am surprised that I have to explain such trivial details. Isn't it obvious that the CPU will be running at lower clocks to achieve lower performance? S8E is what, around 15% faster in multi-core, right? So to match the lower MC performance of A18 Pro you would need to lower the clocks, by approximately 15%, right? The rest of course is a guess. We don't really know the core activation pattern or the clocks these CPUs sustain in MC workloads. My guess judging by previous data on Oryon is somewhere in the ballpark of 3.3-3.5 Ghz for the prime cores.
Well I actually was expecting proof not explanations for your assumption. You claimed 8 Elite runs at much lower clocks than the A18 at the same wattage, how exactly do you know at what clocks it runs in various scenarios? Nobody published any such info.
Qualcomm GPUs have been inspected in a lot of detail, and their compute performance is well know. Please read in-depth analysis by well-informed tech sources like ChipsAndCheese, which discuss the limitations of Qualcomm GPU designs. I am looking forward to see the compute performance of the new GPU, Qualcomm has a lot to improve here, and I hope they did some work in this area.
Fascinating, they obviously didn't test the 8 Elite GPU which actually does use a new GPU architecture, and not just a slightly a new architecture.
And yet your opinion was about 8 Elite's GPU, strange. Not to mention I didn't see a lot of compute tests from them, or tests for 8Gen2, 8Gen3. And still you think have a general portrait of all of Qualcomms GPU's. Curious.
Qualcomm claims to have the fastest NPU on the market, but in actual industry-standard benchmarks the NPU performance is not very good. It is one thing to advertise XXXX TOPS, it is quite another one to use it.
Well maybe the benchmark isn't good. Or you say Qcom makes false claims? Microsoft seemed quite confident in X Elite's NPU performance for example so yeah in actual practice that benchmark is irrelevant.
An example from related domain: Qualcomm's GPUs feature some very impressive peak FLOPS numbers since they pack a lot of ALUs behind one scheduler. But many real-world workloads cannot utilize such wide architecture efficiency. There is a reason why GPU industry has converged at 32-wide wavefronts.
So you say, I see no indication of this.
That is fair. I don't know much about these things and they are not my primary interest. It is very much possible that Qualcomm is an industry leader in these components.
I mean, it's the Modem, so really? "very much possible"?
Apple has the best E-core design currently on the market. They have 1/3 of performance of the P-core but consume 10-20 less power. If I remember correctly, M4 is around 0.5 watts at peak, and 0.25-0.3 watts in multicore operation. Of course, that is according to Apple's internal power-reporting tools, which might be inaccurate. We lack good tooling to properly estimate power, unfortunately.
It's actually more like
0.5-1 watt at peak per core for A-series chips from what I saw. Big Cores also run at lower clocks in multicore workloads and even lower if all cores are maxed out. This is not a computer chip that's not power constrained, so that 15% better perf is an exaggeration. And there's no reason for apple to not add 2 extra cores if the power penalty would be in the margin of error.
I just don't find it useful. Qualcomm certainly does, since they build the thing, and their customers probably find it useful too. I don't. I also don't understand why one would buy Alpine hiking shoes to do a light walk in the part, but there are people who do that as well.
Of course, we will wait for apple to do it, then no doubt it will be useful.
What I said is based on the work of people lime Dougall Johnson, who reverse-engineered details of Apple's microarchitecture, the documentation published by Apple and Qualcomm, and clang patches that describe CPU-specific optimizations. You are welcome to look up these courses and do your own comparison.
Yeah that doesn't say much at all.
Also I hope you understand Nuvia first designed a server CPU core. When Qcom bought them, they had to adapt the core architecture to mobile devices, so it had to be optimized for laptops and now in the 8 Elite they use the 2nd gen Oryon core which was specifically optimized for smartphones, as Qcom confirmed on stage. Did Mr. Dougall Johnson analyze the second gen Oryon Core in the 8 Elite?
That is a weird thing to say. The IPC is different, the dispatch width is different, the timing of certain instructions is different. It is obviously an evolution of Firestorm rather than a complete redesign (like Lunar Lake was), but M4 for example is quite a significant departure from M2.
So not really "new microarchitectures". Got it.
That is not what I mean by "unified cache". Apple's GPU designs forego shader register files and specialized storage, merging everything into a single unified L1. What's more, that storage can spill to higher levels of the memory hierarchy . There is no other GPU on the market that can dynamically allocate shader registers or spill them to L2.
Different versions or variations of a similar thing. Yeah others did it before them.
You talked like it was something that became a standard and a necessity in new GPUs in order to be competitive.
Maxwell features unified virtual memory, not unified physical memory. If I remember correctly Intel has been building UMA systems since at least 2012.
Mybe, my idea was about unified memory in general.
www.cnx-software.com
