Investment Thesis
One of the (plenty) major semiconductor news items this year has been Apple's (AAPL) announcement of transitioning its Macs to Apple Silicon over the next two years. Since this process would start at the end of 2020, it has been expected that those first Macs would be powered by Apple's upcoming A14 SoC. By many, it had been expected to be a leap ahead of the Intel (INTC) silicon Apple would replace (not the least due to some concerns about Intel stagnating), most notably because it would be built on TSMC's (TSM) 5nm (aka 5N) process which is seen as leading edge, while Intel is on 10nm still.
As part of its September iPad and Apple Watch announcements, Apple has unveiled this A14 chip, and it doesn't look like it will fulfill the high expectations. While Apple touted solid numbers, a closer look revealed this comparison was not against the previous generation, but against the A12.
While admittedly the expectations on Apple each year are quite high given its execution, the announcement followed after two previous semiconductor launches in September: Intel's Tiger Lake mobile CPUs and Nvidia's (NVDA) Ampere GPUs. Both products significantly raised the bar for performance, despite being based on what is now previous-gen process technology.
In that light, Apple Silicon is not on track for making a great introduction as it will likely be trailing on performance. This is not what many would have expected.
As one caveat, Apple often develops higher performing X versions of its SoCs, such as A13X. If that's also the case here, then that (still unannounced) A14X chip could find its way into Apple's upcoming Macs instead. Even then, some of the following analysis would likely still remain true.
A14X
Apple presented the A14 as follows:
It features close to a 40% increase in transistor count. But due to moving to TSMC's 5N process, silicon area has likely decreased somewhat.
The CPU retains its big. Little design with two big and four smaller cores. Apple touts a 40% increase in performance against the A12. But given that the A13 already improved by 20%, the gen-on-gen improvement should be around 16%. The big cores incorporate the AMX engine that was introduced in the A13, but seemingly an improved variant (from 6x to 10x performance increase). This is an in-CPU engine for number crunching such as machine learning.
This is in addition to the dedicated neural engine, which was doubled from 8 to 16 cores, according to Apple. This results in 11 TOPS.
The GPU is also stated as being improved by 30%, but compared to the A13, this is less than 10%. This implies it might be the same GPU, but just clocked a bit higher.
Lastly, the ISP (for images) has been improved.
Altogether, a solid update (+16% CPU, +8% GPU), but nothing groundbreaking. Just like Intel's Tiger Lake, one of the key points seems to be its AI improvements. Indeed, AnandTech also noted that the improvement was rather "meagre".
Apple Silicon vs. Tiger Lake
Apple's A14 chip packs a ton of transistors. This has allowed it to pack many features in one SoC, from its CPU to GPU to a powerful neural engine.
This is made possible, without blowing up the silicon area because of TSMC's latest 5N process, which improves logic density by 1.84x. Although the gain is lower for some other types of silicon such as SRAM memory and I/O.
Nevertheless, as I discussed in a recent article, I argued that there is more about process technology than just transistor density: Intel Vs. TSMC: Process Technology Leadership Is More Than Transistor Density. As I already stated in that article: we have yet to see an Apple CPU go past 3GHz in frequency.
Meanwhile, Intel has followed a different approach (in part out of necessity, of course, given its longstanding process delays) and launched Tiger Lake on a less dense process, but excessively focused on performance (and power) instead. Intel called this intra-node improvement 10nm SuperFin.
Indeed, peak frequencies were improved by around 20% or more, from 3.9-4.1GHz to 4.8GHz, and may reach 5GHz in future chips. On the GPU side, the result of Intel's new Xe architecture and SuperFin even yields up to a 2x increase in performance at the same power.
The performance this all has resulted in can be seen below:
Compared to Tiger Lake, the A13 and A14 score 1338 and 1583 vs. 1620 in single core, and 3519 and 4198 vs. 5640 in multi-core. In this benchmark, the increase vs. A13 is a bit higher than 16% but falls well short of Tiger Lake.
To be sure, Apple's performance is great (to say the least) for tablet form factors, as this is a comparison against a laptop chip. This means a ~5W chip approaches the performance of a 15-28W chip. However, when it comes to the raw performance figures, the reality is that Apple's approach simply falls flat in these less power constrained environments against almost 5GHz chips. Contrary to the mobile space, Apple does achieve leadership figures here.
This seems to confirm the thesis in the process leadership article mentioned above: Intel seems to be effectively mitigating its process delays by meaningfully improving its existing leading edge processes.
Still, given the power difference, I am sure A14 has a (much) higher power efficiency than Tiger Lake. But that is a function of their divergent design approaches. Intel aims for 5GHz frequencies or more to cater to products with 7W to 45W thermal cooling capacity and even more for desktop CPUs. Those frequencies are the bar Intel has (to) set itself as far as peak performance is concerned, given its legacy as PC vendor. Apple, on the other hand, by being less aggressive on frequency can make much different design trade-offs to balance performance and power efficiency. Given the lower power envelopes it is targeting for its iPhones and iPads, this obviously makes sense.
Discussion
Apple's approach is more than valid for phone and tablet form factors, but for laptops, it will not result in the performance crown as Apple has there. Apple may achieve the battery life crown, and an A14X, if it exists, may surpass the 4-core Tiger Lake in multi-core performance, but the difference between ~3GHz and close to 5GHz is just too high to be competitive on purely a core-against-core basis.
So again, much of the differences between Tiger Lake and A14 are a result of different target markets, which shows in favor of Apple for some characteristics of the chip, while it favors Intel in others. Most notably, Intel loses quite some power efficiency by having to target 5GHz frequencies for its chips.
Still, one point of argument often seen is Intel's loss of process technology. But despite that being the case on paper (at least in density), that simply does not show in any tangible way on the product side: while an 11B transistor chip on Intel's 10nm would be a fairly large chip for a consumer product, it would still be possible to manufacture this in principle, and the SuperFin enhancement clearly holds up in performance.
Concerning power, any battery life advantage of A14 might not be as great either when considering that this is more a function of standby CPU power consumption, and Tiger Lake has also substantially improved power management, so it will have both high peak performance and long battery life. (Snapdragon's efforts at entering the PC market have already shown that just a battery life advantage may not be enough, especially if performance does not hold up to the competition.)
Nvidia Ampere
Nvidia's Ampere was the first silicon announcement in September, although serving a much different segment.
Against its predecessor, Ampere benefited from moving from a 16nm process (Apple A14 instead competed against 7N A13). Nvidia touted it as one of the largest generational GPU improvements.
Gaming Ampere is based on Samsung's 8nm, which is more of a descendent of Samsung's 10nm technology than a competitor to TSMC's 7nm. Nevertheless, despite this, clearly, inferior process technology, Ampere was received well by investors and gamers.
In other words, whereas Intel got heavily penalized on the stock market for its process delays, Nvidia mostly got away with using Samsung's 8nm process. On the other hand, while Apple clearly mentioned the 5nm process, it did not discuss any benefits of this such as improved energy efficiency.
Risks
The data (benchmarks) in this article is quite limited, just one benchmark. It serves to give a general indication of what might be expected once Apple Silicon and Tiger Lake.
Secondly, Apple said its Macs would be powered by Apple Silicon. This may not necessarily be the same as its iPhone or iPad silicon, to which I compared Tiger Lake here.
Apple might have build a completely separate chip for its Macs (with different characteristics hence). Nevertheless, some have expressed doubt about this possibility given the lower volume of Macs, making it hard to justify developing completely separate chips.
As another possibility, there may also be a faster A14X chip in the works, something that Apple would often design for iPad Pros. Those aren't fundamentally different chips. They would simply feature more of the same to increase performance: more CPUs cores, more GPU cores, etc.
More cores will improve performance, of course, but, for some workloads, that scale less linearly with core count, such as CPU workloads, the argument in this article about the performance disadvantage of a peak frequency of around 3GHz vs. one of up to 5GHz remains valid.
Lastly, financially, Apple is not a silicon vendor. Apple Silicon is just a means to accomplish its overall product goals. For Apple, the move to Apple Silicon is likely for a large part out of its goal to unify its ecosystem to one architecture.
Some have argued that Apple Silicon might drive an increase in Apple Mac sales (perhaps due to pertained performance advantages), but this is not something I would have agreed with anyway. So, the financial impact of less competitive performance (in the worst case) might not be tangible. Also, the transition will take two years and Apple will likely still release some Intel-based Mac.
Takeaway
Some likely would have expected (vast) leadership performance from Apple Silicon. However, the CPU performance improvements Apple touted (when normalizing to the previous generation), and confirmed by one benchmark, were mostly evolutionary. What surprised me most was actually that Apple didn't make any statements at all about power efficiency, despite moving to the widely anticipated 5N process, which on Wall Street has been seen as a large threat to Intel.
The meager Apple A14 Silicon improvements are even more pronounced when comparing it against the two other major silicon announcements earlier in September, which are based on what now has to be considered trailing edge technology.
Tiger Lake retains Intel's 10nm process but has improved this node substantially ("SuperFin"), akin to the improvements one normally sees from an inter-node jump such as 7N to 5N. The result is that Tiger Lake improves performance gen-on-gen by more than Apple's A14 does. Initial estimates also indicate that A14 will indeed trail behind Tiger Lake or be about on-par at best.
Nevertheless, another way of looking it is that the A14 for now resides in fanless devices and comes quite close to matching Tiger Lake in laptops. In principle, there might be some room available to improve performance for laptops.
But, at present, while any A14-based Macbooks may feature great battery life, they are likely not going to win any awards for (breakthrough) performance. So that part of the Apple Silicon advantage thesis seems to fall flat, at least this generation. While a faster A14X chip with more cores may change the picture somewhat, at the end of the day, a chip that was designed to operate in a 2-3GHz regime, while being very efficient in power, will simply struggle in many workloads against a chip designed for (close to) 5GHz frequencies.
So, overall, while much has been discussed about the move to Apple Silicon, the initial impressions of the A14 generation do not show any substantial improvements that may give Apple a large, meaningful advantage (in the sense of becoming a definite selling point). This is despite featuring what on paper looks like a process node advantage, concerns about which caused Intel stock to drop by ~20% in July. (While this article was mostly about the A14, Apple 5nm Silicon so far does not seem to seem to validate the Intel stock drop hence.)
As Intel's Tiger Lake showed, Intel too isn't standing still and competition remains an ongoing race each generation.
Disclosure: I am/we are long INTC. I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.
The Link LonkOctober 09, 2020 at 12:49AM
https://ift.tt/3lqnvjU
SuperFinished: Apple Silicon Not Holding Up - Seeking Alpha
https://ift.tt/2YXg8Ic
Intel
No comments:
Post a Comment