Editorial: Samsung's new Exynos 990 isn't fast enough to maintain its custom M core

Samsung's scant mention of M

Since 2010, Samsung's Exynos mobile processors have sought to deliver ARM chips that could compete with Qualcomm's Snapdragon in the market for Android phones and tablets and Chromebook netbooks. Samsung once sought to position its Exynos SoCs as rivaling the performance and sophistication of Apple's Ax-series of mobile Application Processors.

Samsung has now delivered five generations of its custom ARM core since the M1 "Mongoose" first appeared in 2015. But its latest M5 core used in the new Exynos 990 is apparently its last. The custom "M" ARM cores the company has been designing at its Samsung Austin R&D Center (SARC) in Texas have simply not been competitive in either their performance or efficiency, the two most important benchmarks of mobile CPU capability.

Earlier this month, Samsung reportedly laid off hundreds of SARC chip designers. At its Samsung Tech Day media event this week, it didn't even bother to name its latest core, instead simply referring to the new Exynos 990 as having two "powerful custom cores" in addition to the off-the-shelf A76 and A55 cores it licenses from ARM.

Pay no attention to the CPU behind the curtain

Samsung promoted its new chip as delivering a 20% increase in performance over its previous generation, but as Android Police noted, "seeing as how the Exynos 9820 SoC got trounced by the Snapdragon 855 earlier this year, we wonder if a 20% improvement will be enough for Samsung to catch up to Qualcomm's performance levels."

The relative performance of Samsung's custom "M" cores to the Kyro cores used in Qualcomm's latest Snapdragon 855 Plus is easy to see because Samsung ships virtually identical versions of its Galaxy S flagships with both chips. It uses its own Exynos chip internationally— and domestically in Korea— while being required to use Snapdragon chips in its premium models sold in the U.S., Japan, and China.

Samsung's M-based Exynos CPUs don't compare favorably with Qualcomm's chips, but look particularly weak next to Apple's A13 Bionic, which is beating Qualcomm's latest CPU tech by an insane 77 percent margin in superior CPU core performance. Apple noted a similar 20% jump in CPU and GPU performance while achieving even greater improvements in power efficiency on its A13 Bionic.

Samsung focused its attention on other features of its SoC, including a refreshed ARM Mali GPU, a new NPU for AI tasks such as facial recognition, and an improved ISP with support for five cameras with up to 108 megapixel sensors. The company also promoted its new integrated 5G modem. All together, these details suggest the appearance of fancy new products, perhaps next spring.

There are, however, some problems.

First, Samsung's premium phone sales have collapsed. Also, the Chromebooks, premium Android tablets, and other third party buyers that Samsungs once sold its Exynos chips to have also dried up and blown away, a problem Exynos shares with Qualcomm's high-end Snapdragons. On top of that, even Samsung will only be able to ship its new Exynos chip in phones outside of the U.S., China, and Japan, limiting its potential impact to Korea and other international markets where a Qualcomm modem isn't required.

This adds up to an increasingly dire future outlook for Samsung's Exynos line of chips. With high-end sales already in trouble, it increasingly makes less sense for the company to be dumping money into its custom "M" core architecture that isn't paying off but is increasingly falling behind.

An apparent end of M

Samsung first delivered its custom "Mongoose" M1 core design in 2015, which it used in some models of the Galaxy S7 and Note 7. It was followed by M2, M3, M4 and the latest M5 generations across the last five years. But these Exynos chips were only usable in a small slice of Samsung's premium-priced Galaxy S and Note models sold outside of the largest markets for smartphones — where Qualcomm's CDMA patents made it impractical to sell an Exynos with Samsung's integrated modem.

Samsung also knew that CPU speed wasn't a major driver of its smartphone sales. For buyers who wanted to see fast benchmarks, Samsung could simply cheat to provide impressive numbers, which was much cheaper than developing entirely new chip architectures.

The company has also preferred to show off its other technical achievements, including its lead in OLED panels, Qi charging, and elsewhere. Adding a licensed NPC that promises to accelerate artificial intelligence is easy to do, as demonstrated by Huawei. No need to custom design a CPU core. And focusing attention on 5G also gains media attention, even if 5G networks are not available and won't materially benefit most users for years.

Samsung's System LSI chip fab was also working to stay in the lead in chip fab process, which could help even a sub-optimal core design run fast simply by scaling it down and ramping up its clock speed. Samsung also faced undifferentiated commodity competition from Chinese phone makers who were using basic ARM designs of their own. Sales of middle-tier Androids have been destroying demand for high-end Galaxy S options.

So for Samsung, spending vast amounts of money on custom core design it could only really use on a fraction of its newest, most expensive phones, while it was also dealing with post-peak premium Galaxy S sales in a maturing market, was sort of like changing the tires on a car that was currently on fire. Five years of limited M-core development have only resulted in a third-place runner-up that it could sell in select markets anyways.

Custom ARM core design isn't a cake walk

In contrast to Samsung's odd avoidance of even mentioning "M5," Qualcomm name-drops its own "Kyro" custom core as often as possible, and Apple loves to come up with fast sounding marketing names for each major new generation of its custom CPU core designs: this year it was Lightning and Thunder, the A12 Bionic introduced Tempest and Vortex, A11 Bionic used Monsoon and Mistral; A10 Fusion blew out Hurricane and Zephyr and A9 spun up Twister.

But Apple's storming deluge of ARM-twisting custom-core development hasn't just been clever marketing. Industry observers have been in awe of the amount of work the company has devoted to building its custom Ax ARM chips across the last decade, noting that Apple's fastest chips are now bumping up into the raw performance territory of Intel's x86 offerings, while crushing x86 in power efficiency.

The fact that Apple is catching up to Intel should highlight how incredibly complex and expensive it is to play in the big leagues of state-of-the-art CPU design. Intel has ruled the roost in PC CPU performance since the 1990s, even in the face of the very determined consortium of IBM, Motorola and Apple to defeat it with Power PC in the 1990s, and despite intense competitive efforts more recently by AMD.

There are several universities actively working on launching satellites, but none are developing their own state-of-the-art CPU processor designs. Advanced custom silicon processor architecture design is beyond even space research.

Why has Samsung fallen behind in ARM chip design, despite years of world-leading hit sales volume of smartphones? The next article will take a closer look at how Apple was able to sneak past larger more entrenched silicon experts at Intel, Qualcomm, and Samsung while the media doubted its progress and bet that rivals would easily catch up.