The fate of Samsung's custom CPU development efforts has been making the rounds of the rumour mill for almost a month, and now we finally have confirmation from Samsung that the company has stopped further development work on its custom Arm architecture CPU cores. This public confirmation comes via Samsung’s HR department, which last week filled an obligatory notice letter with the Texas Workforce Commission, warning about upcoming layoffs of Samsung’s Austin R&D Center CPU team and the impending termination of their custom CPU work.

The CPU project, said currently to be around 290 team members large, started off sometime in 2012 and has produced the custom ARMv8 CPU microarchitectures from the Exynos M1 in the Exynos 8890 up to the latest Exynos M5 in the upcoming Exynos 990.

Over the years, Samsung’s custom CPU microarchitectures had a tough time in differentiating themselves from Arm’s own Cortex designs, never being fully competitive in any one metric. The Exynos-M3 Meerkat cores employed in the Exynos 9810 (Galaxy S9), for example, ended up being more of a handicap to the SoC due to its poor energy efficiency. Even the CPU project itself had a rocky start, as originally the custom microarchitecture was meant to power Samsung’s custom Arm server SoCs before the design efforts were redirected towards mobile use.

In a response to Android Authority, Samsung confirmed the choice was based on business and competitive merits. A few years ago, Samsung had told us that custom CPU development was significantly more expensive than licensing Arm’s CPU IP. Indeed, it’s a very large investment to make in the face of having the up-hill battle of not only to designing a core matching Arm’s IP, but actually beating them.

Beyond the custom CPU’s competitiveness, the cancellation likely is tied to both Samsung’s and Arm’s future CPU roadmaps and timing. Following Deimos (Cortex-A77) and Hercules (Cortex-A78?), Arm is developing a new high-performance CPU on the new ARMv9 architecture, and we expect a major new v9 little core to also accompany the Matterhorn design. It’s likely that Samsung would have had to significantly ramp up R&D to be able to intercept Arm's ambitious design, if even possible at all given the area, performance, and efficiency gaps.

In practice, the end result is bittersweet. On one hand, the switch back to Cortex-A CPUs in future Exynos flagship SoCs should definitely benefit SLSI’s offerings, hopefully helping the division finally achieve SoC design wins beyond Samsung’s own Mobile division – or dare I hope, even fully winning a Samsung Galaxy design instead of only being a second-source alongside Qualcomm.

On the other hand, it means there’s one less custom CPU development team in the industry which is unfortunate. The Exynos 990 with the M5 cores will be the last we’ll see of Samsung’s custom CPU cores in the near future, as we won't be seeing the in-development M6. M6 was an SMT microarchitecture, which frankly quite perplexed me as a mobile targeted CPU – I definitely would have wanted to see how that would have played out, just from an academic standpoint.

The SARC and ACL activities in Austin and San Jose will continue as Samsung’s SoC, AI, and custom GPU teams are still active, the latter project which seems to be continuing alongside the new AMD GPU collaboration and IP licensing for future Exynos SoCs.

Related Content:

Source: KXAN

POST A COMMENT

42 Comments

View All Comments

  • Andrei Frumusanu - Monday, November 4, 2019 - link

    It doesn't make sense in mobile designs as idle execution units would be preferably clock gated instead of having wasteful overhead logic to enable SMT. Arm has said that they have no plans of ever adopting SMT in any energy efficient oriented design as it's simply counterproductive.

    Interesting take on an in-order wide SMT core, but that would likely just be something useful for a small core as ST performance would be too bad as a main core. Even the Cortex-A65 will remain datacenter centric with no plans for mobile.
    Reply
  • Kevin G - Monday, November 4, 2019 - link

    You are correct that idle units would be clock and/or power gated. However, leveraging SMT has to be weighed against waking up the front end of the core that would otherwise be idle. Basically it is waking up unused execution units vs. waking up an idle core. SMT would be beneficial if the front end can be shared efficiently (large micro-op caches etc.) with the big power delta coming from the execution units. How many free execution units and how well a front end can be shared is of course dependent upon implementation so any (dis)advantage to SMT is not clear cut. Most designs have chosen the big.little approach but I would not say that that is the definitive answer without seeing the raw power consumption numbers from implementations. Reply
  • brucethemoose - Monday, November 4, 2019 - link

    I wonder how much physical die space SMT would add to a little ARM core. Low priority/background Android tasks seem to be well threaded, hence every new design seems to add even more little cores. Reply
  • yeeeeman - Monday, November 4, 2019 - link

    I am currently using a Galaxy S7 device with the first custom core, the exynos m1. It is a pretty good chip still, fast for everyday stuff, average to good in terms of battery life. It can still run emulators at full speed like PPSSPP or dolphin (not all games). Thermals are quite good and to be honest, even though the newer phones do feel a bit snappier, I will keep using this phone as much as possible since the size/weight is perfect for me. Reply
  • MrCommunistGen - Monday, November 4, 2019 - link

    I think this is a bit of a shame. For the rest of the mobile ARM environment to catch up to Apple's designs in performance someone probably needed to push for a bigger, wider core.

    Although I don't think that any of the Samsung designed M cores ever lived up to expectations, I feel that at least a good chunk of that had to do with the implementation and not the cores themselves.

    The high level impression I remember from Andrei's Deep Dives was that they always targeted too high of a final clock speed - forcing voltages past the point of absurdity, and they had BSP and scheduler issues.
    Reply
  • brucethemoose - Monday, November 4, 2019 - link

    Yeah, Exynos cores seemed to be held back by weird, relatively arbitrary decisions, like excessively high clock targets or those bizarre schedulers.

    Not that those are trivial issues to fix... But I too feel like the cores weren't living up to their potential in actual implementations.
    Reply
  • anonomouse - Monday, November 4, 2019 - link

    Bigger and wider cores are fine, but only if they can get the performance out of it to justify that. When they're at ~3x larger in area and double in power against Arm, but can get only at most ~5% more performance except for some cherry picked microbenchmarks and broken memory component tests, then the tradeoffs don't make sense. Reply
  • master381 - Monday, November 4, 2019 - link

    Would appreciate expanding acronyms on company divisions/terms (SLSI, SARC, ACL). Thanks. Reply
  • Death666Angel - Monday, November 4, 2019 - link

    SLSI is Samsung LSI, their chip division.
    SARC is the Samsung Austin R&D Center.
    ACL is the San Jose Advanced Computing Lab.
    Reply
  • TomWomack - Monday, November 4, 2019 - link

    I wonder who will pick up the team? There's a big ARM site in Austin already, but I'm not sure where ARM would put another three hundred CPU designers Reply

Log in

Don't have an account? Sign up now