Hacker News Reader

Comment by jeffbee | Hacker News Reader

If Samsung was in the same league as Apple Silicon we'd see some evidence of it by now. Before Apple released the M1 they were selling similar parts in iPads. We knew these were fast before they stuck them in Macs. But Samsung is the opposite: their flagship mobile processors are terrible. Apple beats them by 50% on single-core performance. If Samsung was about to jump into the game, I would expect some signs of better parts.

View on HN

This is hearsay, but from what I understood from talking about this long ago with someone deeper in the business, the issues are more organizational than technical, with Apple getting huge benefits of both vertical integration and avoiding possible anticompetitive probes.

This results in Samsung's SoC team having to design each Exynos like it was a la carte offering, and mobile team having to "shop" like normal client, without being able to do SoC and device design in lockstep. Which means mobile team can't specify every last bit of feature they want, and SoC team can't splurge on cache and the like because then the prices might go beyond what would make the SoC theoretically profitable on "open market".

Lol they use off the shelf ARM cores again, and the major issues since the custom cores of years past have dissipated. Sure, Qualcomm adds a slight bit of low-level code and has a different power policy on the 888 vs the 2100, but it’s not actually that different and both are on the Samsung 5NM (by nomenclature anyways) process, which is what actually draws criticism in terms of “Samsung” in the context of SOC’s as of 2021. The reason is simple: the density of the process is solid (EUV and all come in handy) but the leakage is rather disproportionately poor for given points toward the upper range of the voltage-frequency curve e.g. 2.7-3GHz, where mobile “big” or “huge” cores tend to run such as the Cortex X1, Cortex A78 which are both found in the former two SOC’s from Samsung and Qualcomm.

In other words: Samsung and Qualcomm are both behind Cupertino in CPU single-threaded performance but it’s worth noting why this is especially bad: ARM still optimizes for performance per area/efficiency with reference cores & furthermore the implementations of these cores by Samsung and Qualcomm do not utilize remotely as much cache that ARM recommends for the Cortex X1 nor L3 (4MB, but they could go to 16!) which is typical, and the 888 actually reduces the maximum frequency of the X1 (2.84GHz - as ppposed to topping out at 3.09/3.1 like ARM spec or as with their last SOC, the 865) presumably because they deemed the power trade-off not worthwhile for their mainstream flagship chip. Throw in Samsung fab inferiority (to TSMC) and the engine lights here just stack up.

Microsoft is reportedly using an X1 & A710 on a TSMC 5NM process, which is odd given they ought to use an X2 but whatever. Qualcomm and Samsung’s X1’s score 1000-1100 on Geekbench 5’a Single thread test, which is indeed an excellent ledger of general performance. With TSMC 5NM & more cache, this chip is going to have phenomenal performance per watt and accounting for the Windows Geekbench penalty probably have X1’s hitting 1200 in single thread tests, which is about 500 off an M1, but also at less power - ARM quotes the X1 on TSMC 5NM (which Microsoft will be using instead of Samsung 5NM) as a 3.2-3.6 watt core at peak. The A710’s? We’re talking about a 1W profile (and 800-100 GB5 performance if the A78’s are anything to go by, and an absurdly high performance to area ratio).

And really, this is just about providing a decently performant ARM option that retains the ever-obvious performance per watt advantage. The Qualcomm 8CX is just not powerful enough, because it uses old Cortex A76 cores, as does Microsoft’s dogshit SQ1/2. Battery is life is reportedly solid though.

This is one big step in actually hitting “good enough” for ARM on Windows. It’s about competing with X86 and keeping the Windows ecosystem competitive with Apple’s offerings, not necessarily beating them overnight.

Samsung CPU performance meets AMD GPU performance per watt? I couldn't imagine a worse combination heh. I don't think this will compete with the M1 very well. At least I'm glad the industry is moving towards putting cheap high-performance/efficiency SoCs in all laptops finally.

There was a generation here or there where AMD(and Nvidia) had poor performance per watt. And even then it was due to pushing clock speeds and voltages on the high end parts to compete on performance. Hardly the norm or even the case today. RDNA2 seems to be very efficient. Cooling aside the RDNA mobile GPU in my Macbook gets a lot done with 20w-30w of power.

Indeed and the reported Exynos GPU utilizing AMD IP is likely a bit hot on phones but the sustained leaks for 3D games look promising, if they are accurate in that they were recorded for sustained activity then it sets a sort of upper bound on the power draw given it’s in a phone.

Peak:

https://twitter.com/fronttron/status/1430192491450363913?s=2...

Sustained:

https://twitter.com/fronttron/status/1434492507283279884?s=2...

Reply on news.ycombinator.com