Apple chips, no way

Semiconductor Industry Observations

2024-05-15 15:09Posted on the official account of Anhui Semiconductor Industry Observation

Half a year after the release of Apple's M3 chip, the successor M4 suddenly broke into everyone's sight, although it is a processor used on the iPad Pro, but the performance improvement is not small at all.

According to Apple, M4 compared to M2, the CPU is 50% higher, the overall GPU rendering is increased by 4 times, and it can also provide the same performance while reducing power consumption, in short, it is another Apple-style performance leap.

But Apple has intentionally or unintentionally ignored the M3, which may be Apple's shortest-lived processor. The last Apple processor with such a short lifespan may have been traced back to the A5X, which debuted on the iPad 3 in March 2012, and the iPad 4 with the A6X was announced in October of that year, with a short lifespan of only 7 months.

M3 is now facing the same dilemma as the original A5X, powerful performance has always been one of the biggest motivations for consumers to change their phones, but just half a year later, Apple has launched a stronger and better product, and the latest processor has become the next new processor in the blink of an eye, and this iteration speed is inevitably a little too fast.

An even more fatal question has been thrown: What are the chances of Apple's self-developed chips?

01 M3 that "died without a problem".

At the MacBook conference in the second half of last year, Apple announced the M3 series chip using 3nm process, which is also the first M-series chip using 3nm.

Apple says the M3 chip is powered by 25 billion transistors – 5 billion more than the M2 – and features a next-generation 10-core GPU that delivers up to 65% faster graphics performance than the M1, an 8-core CPU with four performance cores and four efficiency cores that can deliver up to 35% faster CPU performance than the M1, and supports up to 24GB of unified memory.

The M3 Pro chip is equipped with 37 billion transistors and an 18-core GPU, which is up to 40% faster than the M1 Pro. Support for unified memory has been increased to a maximum of 36GB. The 12-core CPU design consists of 6 performance cores and 6 efficiency cores, delivering up to 30% more single-threaded performance compared to the M1 Pro.

The number of transistors in the M3 Max chip has increased to 92 billion, the 40-core GPU is up to 50% faster than the M1 Max, and it also supports up to 128GB of unified memory, and the 16-core CPU with 12 performance cores and 4 efficiency cores is up to 80% faster than the M1 Max.

From Apple's description, we can find that Apple is trying its best to avoid the direct performance comparison of M3 and M2 chips, which is an incredible thing, after all, the earliest M1 chip was released in 2020, and the 2023 chip is compared with the M1 series in the case of upgrading the process and core, which can only be said to be a little invincible.

So why did Apple launch the M3 chip and the corresponding MacBook when it knew that the performance was not improving much?

The answer is very simple, that is, the sales performance of Mac products with the previous generation M2 chip is too dismal.

According to The Elec, Mac sales have "plummeted" due to a severe downturn in the PC market, and Apple completely suspended the production of M2 series processors in January 2023.

Unlike the A-series chips, which are packaged in fan-out wafer-level packaging (FO-WLP), which TSMC calls InFO (Integrated Fan Out), the final processing of M2 chips uses a general-purpose flip chip packaging process. Once TSMC completes the pre-processing process, the wafer processing products will be sent to Amkor's South Korean facility for packaging work. In January and February 2023, TSMC did not deliver this part of the product, and the production line was idle for two months, and people familiar with the matter said that although wafers have arrived one after another since March, the number is only half of that of previous years.

Apple CEO Tim Cook also admitted during the Q1 2023 earnings call that Apple faces a "challenging" situation in the PC market. "The industry is shrinking," Cook said, "Apple's share is low, but we have a competitive advantage in Apple chips, so strategically, Apple is in a good position in the market, but it's going to be a little bit difficult in the short term." ”

In order to save the lack of competitiveness of Mac products equipped with M2, Apple's choice is to launch the M3 chip with 3nm process and the corresponding MacBook at the end of 2023.

But when Apple made up its mind to launch a Mac with a 3nm process in 2023, it ran into a new problem - TSMC's N3B process.

TSMC has had a long nightmare with the N3 node. The original N3, the N3B, had 25 EUV layers, almost twice as many as the N5, which also made it difficult and expensive to produce, and most customers were reluctant to pay a high fee given the lackluster improvements in performance, power, and density.

The problems with N3 eventually caused TSMC to miss the 2-year upgrade cycle of the main process nodes, and the N3B was finally put into production in the fourth quarter of 2022, while the N3E had to wait until mid to late 2023 to be put into production, which not only led to Apple's forced to postpone the chip plan - its original plan to use the N3 process on the iPhone in 2022, many customers also retreated from the N3B, such as Zen 5, Intel GPU and Broadcom custom ASIC, or continue to use N5, Either move to N3E as a subsequent improvement process.

In fact, except for Apple's M3 series and A17 Pro, almost no manufacturers are willing to use TSMC's N3B process, which also lays the groundwork for a sweetheart deal between Apple and TSMC.

In August 2023, foreign technology media The Information broke the news that Apple placed a huge order for 3nm chips with TSMC, but required TSMC to bear the responsibility for unqualified chips, and the initial yield rate of TSMC 3nm was about 70%, and Apple could save billions of dollars after reaching such an agreement.

On the one hand, TSMC's 3nm is unique in the world, on the other hand, Apple is unwilling to pay more for unqualified wafers, not so much Apple's tough requirements, but TSMC's half-pushing, so that Apple has formed a de facto monopoly on N3B.

This situation can't help but remind us of TSMC's N10 and N20, two equally short-lived processes.

At that time, after TSMC's 16nm/20nm process experienced a peak in the fourth quarter of 2016, some customers moved to 10nm in 2017, including Huawei HiSilicon's Kirin 970, Apple's A11 and A10X, of which A11 is TSMC's main customer for 10nm.

But in addition to these two, it is rare to hear about TSMC customers who use 10nm, which once reached 11% of TSMC's revenue in TSMC's financial report, but after 2017, this proportion has declined rapidly, and the 7nm process has become the choice of more manufacturers.

As for TSMC's 20nm, it is a notorious process, although many processors have used this process, such as Qualcomm Snapdragon 810, Apple A8, Nvidia Tegra X1, MediaTek Helio X20, etc., but without exception encountered serious heating and runaway power consumption, allowing everyone to experience the horror of "Taiwan leakage".

20nm is poor, customers either quickly migrate to the next 16nm node, or reluctantly use the 28nm process, 2015 not only made mobile phone and processor manufacturers sit on pins and needles, but also TSMC experienced the darkest moment.

Now look at TSMC's N3B, there is a high probability that it will follow in the footsteps of N10 and N20. According to WikiChip, the N3 node appeared to be a one-time node, and TSMC's engineers ran into some hurdles in the way forward, so they decided to change halfway, and at a high level, N3B had different PPAs, as well as the company's claims to be "very different" design rules aimed at increasing yield, with the result that subsequent N3E would not provide any path to direct migration from N3B, and eventually N3B became a dead knot for designers, This is also the reason why TSMC wants most customers to use N3E.

The question is, why did TSMC deliberately launch an N3B? The simple explanation is that TSMC is only trying to fulfill its promise to early adopters of technology like Apple, and perhaps the N3B process will fade out of everyone's attention after the end of this year.

Of course, not only the N3B, but also the Apple M3 series and the A17 Pro that use the N3B, Apple may have foreseen the end of these processors long ago after deciding to adopt this short-lived process.

02 The lackluster M4

Back in 2020, when Apple's first self-developed Mac processor debuted, a whirlwind blew in the technology and semiconductor circles, and the once invincible Intel was actually kicked out by Apple.

This is one of the most glorious moments in Apple's history, with some citing the switch from the Motorola 68000 platform to the Power PC platform, then to the Intel x86 platform, and now choosing the self-developed ARM platform, in an attempt to prove the beginning of a new era.

In the first quarter of 2021, due to the impact of the pandemic and other factors, the overall PC industry shipments increased by 55% year-on-year, and Apple's Mac sales ushered in a staggering 111.5% growth. In the three years from 2020 to 2022, the Windows PC market grew by 6%, while the Mac market grew by 60%, and the Mac market share exceeded 10% in 2022, reaching 10.8%, and in the U.S. market, Apple's home base, the Mac share was even higher, reaching 17% in the fourth quarter of 2022.

But the good news is all left in the past three years, from the beginning of 2023, Apple's Mac has begun to fall endlessly: in Q1 2023, Mac global shipments will be 4.819 million units, with a market share of 8.7%, a year-on-year decrease of 34.2%; In Q2, Mac shipped 5.293 million units globally, with a market share of 8.9%, down 0.3% year-on-year; In Q3, Mac shipped 6.266 million units worldwide, with a market share of 9.7%, down 24.2% year-on-year.

In Apple's Q3 2023 earnings call, Apple's chief financial officer Luca Maestri blamed the decline in Mac sales on too much competition in the market and the release of pent-up sales in 2022, resulting in an overproportionate figure.

Apple's CEO told CNBC that he expects the Mac's market performance to improve during the Christmas shopping season, and he believes that the Mac business will improve significantly in the fourth quarter of 2023 due to Apple's M3 series chips and new Macs equipped with this chip.

According to Apple's financial report for the first quarter of 2024, Mac's revenue in the quarter was $7.735 billion, compared with $10.852 billion in the same period last year, far lower than the $9.63 billion expected by analysts.

M3 did not drive Mac sales, to the surprise of analysts and Apple executives, but also let the M3 MacBook stand on the cusp, you know, Apple spent a lot of money on M3, foreign analyst Jay Goldberg said that Apple spent $1 billion for the tape-out of the M3 series of chips, and the astronomical tape-out cost was much lower than expected.

Interestingly, well-known analyst Ming-Chi Kuo said in October 2023 that Apple may launch a newly designed MacBook Pro in 2025 if the M3 chip still fails to boost MacBook shipments, and may also consider (not yet confirmed) a lower-priced version of the MacBook to further increase shipments, targeting shipments of more than 8 million to 10 million units per year.

In other words, Apple is abandoning the route of the Arm version of the Mac with chip upgrades as the main selling point, and instead considers making a fuss about the design and price, is it a bit familiar? This is exactly what the Intel version of the MacBook did before, and the large size can't be sold? Then make a 12-inch MacBook, missing the upgrade highlights? Then add the bells and whistles of hardware features like TouchBar and butterfly keyboard, in short, there is no selling point and you have to force a selling point.

At this point and then, how proud Apple Mac was when the M1 was launched, and how frustrated it is now, Apple thought it had created a storm for the popularization of Arm notebooks, only to find itself just a pig on the cusp of the wind after three years, and the Mac has returned to its original market position.

The success of M1, on the one hand, is the surge in demand brought about by home office, on the other hand, of course, it also has the credit of TSMC, although the TSMC N3B process in the previous article has pulled a big span, but at this point in time in 2020, TSMC N5 is several steps ahead of Intel and Samsung's respective processes, which also contributes to M1's extremely excellent power consumption control.

Although Apple is still emphasizing the improvement of M2 and M3 compared with M1, the best selling and most recognized products on the market are still M1.

With the large-scale mass production of TSMC's N3E process, the M3 series is destined to become Apple's outcast, but Apple does not seem to give up, and has prepared the N3E M4 at a fairly fast speed, using it to replace the expensive M3, and then do what he thinks after reducing the cost.

According to Bloomberg's Mark Gurman, Apple aims to release new Macs from the end of this year to the beginning of next year, with the M4 chip on the new iMac, the low-end 14-inch MacBook Pro, the high-end 14-inch and 16-inch MacBook Pro, and the Mac mini.

According to Gurman, Apple plans to launch more M4 Macs throughout 2025, including the 13-inch and 15-inch MacBook Air, which will be updated in the spring of 2025, the Mac Studio that will be updated around mid-2025, and the Mac Pro, which will be updated later in 2025.

It said that the M4 chip family includes the entry-level Donan, the more powerful Brava and the high-end processor, codenamed Hidra, and Apple plans to highlight the AI processing power of these three processors and how they will integrate with the next version of macOS to be released at WWDC this summer.

Donan 芯片将用于入门级 MacBook Pro、新款 MacBook Air 和低端版 Mac mini,而 Brava 芯片将用于高端 MacBook Pro 和高价版 Mac mini。 至于 Mac Studio,苹果正在测试采用尚未发布的 M3 系列芯片和 M4 Brava 处理器变体的版本,Hidra 芯片则会搭载于苹果最高端的台式机 Mac Pro之上。

So will the M4 outperform the M3? Not necessarily.

First of all, N3E is not a qualitative leap compared to N3B, and it can even be considered that the two chips are actually Apple's two 3nm answer sheets, and when we put M3 and M4 together, we will find that even the improvement on paper is a bit lackluster.

First of all, the CPU part, the full version of M4 includes 4 performance cores and 6 efficiency cores, compared with M3 2 more efficiency cores, and the beggar version of the iPad has moved the knife on the core, for 3 performance cores + 6 efficiency cores configuration, and the GPU part, M4 is almost the same as M3, the same is 10 cores, and it also supports M3 has dynamic caching, ray tracing hardware acceleration and mesh coloring and other functions.

The NPU part is the biggest improvement of M4, M4's 16-core neural network computing power is as high as 38 TOPS, compared to M3's 18TOPS, which is a full 20 TOPS higher, and it also exceeds Apple's A17 Pro's NPU's 35TOPS computing power, according to Apple, this will help a lot for future AI applications.

Looking at the actual running score, in the latest GeekBench 6 running score test, the iPad Pro equipped with the full version of M4 has a single-core score of 3767 points, and a multi-core running score of 14677 points, compared with the previous generation M3 (3053), the single-core performance has increased by 25%, and the multi-core performance is roughly equivalent to the latest 11-core M3 Pro (15261), while the GPU's Metal test score is 53792 points, an increase of 13.4% compared with M3 (47420). It doesn't surpass the much larger M3 Pro. It should be noted that the processor frequency of the iPad Pro in the benchmark score is pulled to 4.4GHz, and the actual frequency of the M3 Max, which has the highest frequency, is only about 4.1GHz.

In the Geekbench ML 0.6.0 test, which represents the NPU, the M4 scored 9234, about 22.9% higher than the 7511 score of the iPad Pro with M2, and only 10.4% higher than the M4 (8365) of the 14-inch MacBook Pro.

At this time, you may think that something is wrong, isn't it 20TOPS more computing power, why is the running score only increased by 10.4%? In fact, Apple is playing a word game here, M3's 18TOPS is obtained at FP16 accuracy, while M4's 38TOPS is obtained at INT8 accuracy, the latter is actually discounted, that is, 19TOPS, so that the actual TOPS improvement is only about 5%.

Compared with the original Intel x86 processor, M1 is undoubtedly a cross-era processor, but the subsequent M2, M3 and M4 seem to be only improving on the basis of the M1 architecture, and there is no huge improvement, M4's frequency on the iPad is even pulled to 4.4GHz, you must know that the frequency of M3 running on MacBook Pro is only 4.05GHz.

Why did Apple choose to let the iPad Pro debut the M4 chip? The main thing is to drive iPad sales. In February this year, iPad revenue became a rare bright spot in the first fiscal quarter, Apple's net revenue from iPad in the quarter was $9.396 billion, an increase of 29.7% from $7.248 billion in the same period in 2023, exceeding Wall Street analysts' expectations of $7.76 billion.

However, in terms of the performance improvement of the M4, whether this intention can be realized may be a big question mark.

03 The "chasm" that Apple can't cross

For M3 and M4 chips, the current situation is the same as the dilemma faced by everyone in the chip industry. For example, process limitations, physical law limitations, and even mask limitations. And the current situation of these chips can't help but remind us of Apple's self-developed baseband curse.

From the earliest revelations that may be unveiled in 2023, to the earliest debut in 2028 revealed at the end of last year, in the blink of an eye, Apple has spent almost 10 years pondering the so-called self-developed baseband, and after billions or even tens of billions of dollars of investment, it is still unknown whether it can replace the current Qualcomm baseband.

There is also the so-called self-developed wireless chip, when the news came out in 2021, the stock prices of companies such as Broadcom and Skyworks Solutions fell in response, but for more than 3 years, the project has progressed slowly, and at the end of last year, there was news that Apple's wireless chip team seems to have been suspended, and the self-developed WiFi chip is unlikely to be carried on the iPhone in 2025.

At the end of last year, Bloomberg broke the news that Apple also intends to develop its own image sensor, because it is the core of mixed reality and autonomous driving, and Apple's self-developed microLED screen and non-invasive blood glucose detection system are also advancing, as a result, this year, two of the three plans have been blocked: the microLED team has been disbanded and reorganized, and the U.S. Food and Drug Administration (FDA) has also clearly opposed non-invasive blood glucose-related devices.

With the fading of the popularity of Vison Pro, as well as the collapse of Apple's car, the future of Apple's self-developed image sensor team may also be clouded.

The success of the A series of chips has made Apple full of confidence in the chip, and the success of the M series chip has further expanded Apple's confidence.

But contrary to expectations, Apple has been able to call it successful so far, and it still only has the two series of chips A and M, and even the latter is still trapped by the continuous decline in Mac sales, and now it is up to the iPad to be the last straw.

The cost of tape-out of 3nm chips alone is as high as one billion US dollars, if there is no high-sales corresponding product to dilute this part of the cost, then what is the point of reducing costs for self-development in the first place? If the sales of the M4 iPad Pro are still unsatisfactory, then which category will Apple look for to save the day?

Perhaps at this time, it can be said that there is nothing that can be done if it is as strong as an apple.

View original image 319K

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

•