Ah, here's the thing:I never said that M cores are crappy, or there is anything wrong with them. The thing is I doubt you need 6 of them in the mobile SoC other than to boost benchmark scores. If all the background tasks can be handled by a cluster of 4 A520 cores, then 4 M cores will definitely be able to handle them sufficiently.
And what is the sustained MT throughput you need in mobile phone other than benchmarks?Ah, here's the thing:
Qualcomm is using the Oryon-M cores not only for efficiency but also for sustained MT throughput. That's why they are calling them "Performance Cores"
View attachment 110197
Essentially, the Oryon-M cores in Snapdragon 8 Elite take up the duties performed by both Cortex A7xx and Cortex A5xx in previous Snapdragon chips.
I would argue that putting 6 Oryon-M cores is actually a sensible engineering decision.All I am saying is that that could do well with 2+4 configuration matching what Apple is doing, but the decision to include 6 was motivated by the marketing as it would look bad if they had fewer than 8 cores next to android competition and they would not get the MT win in Geekbench vs Apple what looks good in the slides
Yes, this is the idea behind adding more weaker cores. Your analysis seems valid but is missing data, I mean we don't how power is being assigned to the individual cores. Are all 8 cores active or maybe it's just 2+2. And until somebody measures that we won't have definitive answer if one can be had Anyway I am afraid we went too far from the original topic of this thread. Maybe we should move further discussion back to the Qualcomm one?I would argue that putting 6 Oryon-M cores is actually a sensible engineering decision.
//
Geekerwan review have shown that Oryon-M is not as efficient as A18-E.
View attachment 110198
It's close, but there is a gap nonetheless.
What this means is that in terms of efficiency;
[4 Oryon-M] < [4 A18-E]
By putting 2 more Oryon-M cores, they made it more efficency than Apple.
[6 Oryon-M] > [A18-E]
This is due to the nature of Voltage-Frequency curves. As frequency increases, voltage increases exponentially. So power consumption also increases exponentially.
By putting more cores, you can run then at a lower frequency (and thus at a lower power consumption) to match the same performance level.
View attachment 110200
That is exactly what you are seeing in this GB6 MT curve.
Qualcomm's cores are individually less efficient than Apple's, but Qualcomm put more of them. Hence why in this MT power curve, 8 Elite is more efficient than A18 Pro at the lower part of the curve.
//
Well, yeah, they're something closer to Skymont or Zen4c than any actual real LITTLEs like what Apple ships.That is not guaranteed, because Qualcomm's "E-cores" are much more performant than Apple's.
Absolute performance wise yes.Well, yeah, they're something closer to Skymont or Zen4c than any actual real LITTLEs like what Apple ships.
But has higher consumption and that would hurt battery life.Absolute performance wise yes.
But IPC, core area and scale of the architecture are similar between Oryon-M and A18-E.
Oryon-M is faster because it clocks higher (3.53 GHz vs 2.2 GHz).
Sort of. For BL numbers in LITTLE your main factor is joules spent per workload.But has higher consumption and that would hurt battery life.
Individual SoC vendors vary the cache size, so it's possible Ex2500 is cache starved.JSON shows the actual clock speed was 3.2-3.3 GHz, so this is a pretty bad score for a Cortex X925 at that clock speed (Dimensity 9400 does 3000 at 3.6 GHz).
The bigger disaster, from what little I've heard and read, is that their trailing nodes aren't even getting fully matured to good yield levels before they move on to newer, more advanced nodes. When the stuff that should work like clockwork and yield at a high percentage is trailing competitors, you can't compete with them on price there and just hemorrhage money.This is a big dissaster...
"Samsung Foundry's second-gen 3 nm node yields are reportedly stuck at 20% "
Even SMIC has better yields than Samsung's! Something is clearly off on there. And Samsung should start training for better engineers.
If that happens Samsung Foundry is done in the high tier and will get a massive budget cut focusing on mid tier processors, at least the Exynos 1XXX line is not bad after all.Sketchy rumours claim that Samsung's next Exynos chip may be manufactured by TSMC:
Sketchy rumour alleges future Exynos chip could be manufactured by TSMC
Samsung LSI may soon ask TSMC to manufacture a future Exynos chip. This could be due to plummeting yields on Samsung's own 2nd-gen 3 nm node.www.notebookcheck.net
If Samsung Foundry cannot get SF3/SF2 to yield well, and if Samsung LSI wants to continue making flagship Exynos chips, then this seems to be the only option.
Otherwise,
- If Samsung doesn't make flagship chips anymore, then they will be at the mercy of Qualcomm and Mediatek for chip pricing.
- If Samsung insists on making Exynos 2500 on SF3 (or SF2?) despite the bad yields, then it will be an Exynos 990/2200 situation all over again.
Samsung Foundry will have to swallow their pride, and let Samsung LSI go to TSMC. A parallel to how Intel is fabbing their latest products on TSMC nodes, while their internal foundries catch up.
Plenty of their models were already using SD SoC's anyway.If that is real... RIP Exynos for good and RIP Samsung Semiconductors.... to not to say... the increase of the Samsung phones will be dramatic.
They don't, only flagship use Snapdragon in bulk, Samsung got deals with Mediatek and majority of their lowend/Midrange use Mediatek soc then Exynos. Even when they use Snapdragon they pick leftover soc like sd 888 or 7s gen 2 etcPlenty of their models were already using SD SoC's anyway.
Samsung are being very stupid about this regardless of the semiconductor business.
If they really want to be serious about Exynos they need to become flexible about their fab.
SMIC isn't trying anything quite as ambitious as GAAFets. In any case, Samsung needs to do more than train better engineers. They need to reach out to other non-edge foundries and work out IP sharing/collaboration agreements. Not sure if IBM still has research fabs but I think they retained a few before selling out to GF, and of course GF is an option as well.This is a big dissaster...
"Samsung Foundry's second-gen 3 nm node yields are reportedly stuck at 20% "
Even SMIC has better yields than Samsung's! Something is clearly off on there. And Samsung should start training for better engineers.
The roadblock for SMIC is EUV.SMIC isn't trying anything quite as ambitious as GAAFets. In any case, Samsung needs to do more than train better engineers. They need to reach out to other non-edge foundries and work out IP sharing/collaboration agreements. Not sure if IBM still has research fabs but I think they retained a few before selling out to GF, and of course GF is an option as well.
Maybe even Intel would be interested in a partnership.