Hands-On With the Xiaomi Mi5 - High-End at a Mid-Range Price

106 Comments

View All Comments

• jjj - Thursday, February 25, 2016 - link

  What i said was efficiency not power, efficiency factors in perf so, if you pay attention, you can safely assume that i have factored in perf.
  Your claim that a Kryo core is more powerful than A72 is based on what? Because you have nothing at all to support that claim and all existing evidence supports my claim,that in perf and power A72 is a lot better. If you take the Kirin 950 in Geekbench ( just because we actually have numbers in this benchmark) in integer and FP single core , it delivers 16% lower perf on average (lower diff in integer ,more in FP) with A72 at 2.3GHz while Kryo needs 50% more power at 2.15GHz- by looking at power for the 2 SoCs with 2 cores - just 2 cores because core 3 and 4 are lower clocks for Kryo. Kirin 950 is an early implementation , not the latest core revision, an older interconnect,likely not an ideal memory controller on an immature process. We will see better implementation and maybe even a slightly better Kirin soon.So i firmly stand by my claim that A72, based on these power numbers for Kryo , is a lot more efficient.
  Additionally, the CPU related benchmarks on this page, place the Kirin ahead of Kryo, even with A72 at just 2.3GHz plus slower RAM and NAND - in all fairness Kryo might be crippled in this benchmark by Xiaomi's ROM.
  As for your insults , after you learn how to read, maybe you can also learn how to have a conversation.

• jjj - Thursday, February 25, 2016 - link

  Ofc this math includes the all else power and if all else power would be equal (risky to assume that), it would get much worse for Kryo.

• lilmoe - Thursday, February 25, 2016 - link

  For me, I'll reserve any judgement until a deep dive is done for all the major new cores; Kryo, M1, and A72.
  
  Adreno always benches better than Mali, but Mali always seems to perform better in real life (IE: more efficient).

• jjj - Thursday, February 25, 2016 - link

  In GPU it is a bit more complicated. The math includes the CPU so less efficient CPUs will harm the GPU perf too. Another very important matter is die size and that always gets left aside. IF Qualcomm was pushing clocks to save on die size, they could change that, go bigger , lower clocks and gain efficiency. Sadly i have no clue how big the GPU was in SD810. In any case, the GPU should be fine for 1440p even if it throttles a lot. At 1080p some could see it as overkill.
  As for M1, we'll see how hard they pushed it, would be nice if it stays cooler than others.
  But Kryo vs A72 is surprising. Another way to look at these numbers is core 3+4 for both. Kryo at 1405mW clocked at just 1.6GHz but here there is an extra complication because Kryo seems to have a 512KB L2 cache for the 2 lower clock cores (vs 1MB L2 for the other 2) and that's why core 3 adds so much more than core 4. Both excluding that L2 or including it seem unfair, but it is less than the other 2 cores, less total than Kirin and some cache would be needed so we leave it in. For Kirin there is this graph http://images.anandtech.com/doci/9878/power-big_57... . There you can kinda guess core 3+4 power at 2GHz, it's not very precise but seems at about 1050mW. Went with 2GHz because in Geekbench Integer+FP average Kryo at 1.6Ghz would be at best a couple of % faster. So A72 delivers similar perf in Geekbench in this early implementation at much lower power when doing the math this other way that better isolates the core and clocks are not pushed too far. Even here the memory speed is different so perf and thermals will vary a bit and how memory heavy an app is would matter. The budget version of this device does have the same memory speed as the Kirin so would be more interesting to compare that one with the Mate 8.
  Ofc all this is based on the limited available data, but that's all we got so far. Ignoring the little cores is not fair either because they are more efficient and they do offload some work, that is a plus in perf+power and a minus in cost.

• lilmoe - Friday, February 26, 2016 - link

  Yes, CPU power draw plays a big role in GPU efficiency, not question there.
  
  Both Kryo and M1 are at their early stages. Their "maturity" is comprable to A57 and Swift/Cyclone rather than A72 and Twister at this point (again, "maturity" not "performance"). I was a little surprised at first when Qualcomm added "little cores" to their design. But it's understandable; they either need more work on the dynamic scaling (lower voltage at low/medium clocks), or it's a decision dictated by how Android scheduling works based on evidence and recent articles. Either way, they still have lots of space for improvement, and we might see a quad core, all big, configuration with better voltage tuning in next iterations. But I wont hold my breath; they might need 10nm or less for an all big/wide configuration to be feasible on Android, and no less than Quad core, with only 1 or 2 cores configured for turbo.
  
  Samsung, among other, "privileged", ARM partners has pretty early access to ARM designs. If the A72 was better than their M1 (same goes for the CCI), I'm sure they'd have went with that. Unless other design (and cost saving) decisions are at play, I'd voucher for the M1 over the A72. Heck, the M1 might even be a slightly modified A72.
  
  There's also the difference in process nodes. I'm happy we now have a better apples to apples comparison between SD820 and Exynos 8890 since they're both on Samsung 14nm LPP. But until we see an A72 on Samsung's process vs TSMC, it'll still be hard to tell, even with more data.
  
  I honestly don't believe these resent GPUs are overkill for 1080p, even for 720p. I envy how Apple can get away with a 720p screen on a ~5" device till this day, run by such a beefy GPU. I believe it has more to do with cost savings than optimal configuration (since their brand image allows it, and they'd sell either way). It's really sad that general perception dictates that Android OEMs need to have MUCH higher specs than Apple's to be considered as "good" alternatives till this day. I wish Samsung would enable the users to scale down the resolution (just like they can in the Game Tuner app) on the OS level. That 1440p pentile screen will sill look good scaled down to 720p since each pixel will have 5+ sub-pixels. Heck, color reproduction might even improve that way.
  
  Optimizing for Android on the hardware level is proving to be extremely difficult. A royal pain in the behind. Google seems to be more interested on optimizing their software for more cores at this point. There's still WAY too much overhead in their rendering pipeline.

• jjj - Friday, February 26, 2016 - link

  To add something i've omitted, the slower Kryo are supposed to be optimized for power so that might have helped Kryo in this core 3+4 math.
  Don't see the point in Qualcomm going back to 4 instead of 2+2 since the single threaded perf gain this config allows, helps marketing. There were some rumors about MSM8996 Pro or SD823 but from China and China rumors are 90% fake.
  In the end "overheating" in itself is not the problem, the lack of disclosure is. If the specs would disclose sufficient details so the consumer can be well enough informed, all would be ok. "Turbo" or"boost"or w/e marketing term one wants to use, is a good thing, as long as the consumer goes in knowing what he gets. Synthetic benchmarks don't help either in showing the real world behavior of a thermally managed SoC, things need to change there.
  The M1 in Geekbench seems to match Kryo in memory , a very high score. . Why Samsung is using it, hard to say, they could have aimed higher, missed targets and got stuck with it, or maybe it is even better than A72 and not pushed too far. For M1 we really don't have much info yet, or at least i haven't seen much. On the interconnect side, Kirin uses CCI-400, that's why i've mentioned it as a possible area where things can get better. On the process, Qualcomm chose to go with Samsung not TSMC, sure the deal might have been about more than just the process since Samsung needs foundry clients and Qualcomm needs the S7 win. But doubt Qualcomm would have went with Samsung if the process was far worse than what TSMC has.
  On the GPU side, as far as i can tell in actual games i do deem the perf as enough for current games but i don't see the need for 60FPS, some might. Sure mobile games could be much better and , no matter what, GPUs won't really be enough at any point soon.
  The convo about quality, res, FPS and what matters most is tricky. I do agree that users should be able to adjust res. A smart device could adjust all 3 on the fly in trying too offer the best experience possible - that's not easy but could be done at some point. ARM announced recently the Mali DP650 that supports variable refresh rates , can't wait to see that in phones.
  More cores can be a plus for gaming, if developers would try to use the small cores as much as possible and gain some GPU TDP. We might see even more than 10 cores and forcing games on more efficient cores could be a big plus. Consoles using 8 rather small cores could make console ports interesting.
  
  Just spotted a fresh article, Digitimes is saying that Samsung is using SD820 only in the US (others are saying US and China) and that only the SD820 model uses the heatpipe.
  Remains to be seen how true this is, Digitimes is not always reliable.

• jjj - Friday, February 26, 2016 - link

  The single core Kryo number is odd, Compare 1 and 2 cores with Exynos and ... maybe it's the way power was measured or maybe Xiaomi forgot a faucet running and it could end up being a bit better overall. Adjusting the core 1 number wouldn't change the core 3+4 math though.

• aegisofrime - Monday, February 29, 2016 - link

  I found it odd as well. One possibility that I thought of is what if the Huawei had a full charged battery while the Xiaomi was on low charge. The Xiaomi would then be drawing full current since the measurement methodology was measuring input charge power. Conversely the Huawei would then be trickle charging thus pulling lesser power. I do consider this highly unlikely since its Andrei doing the testing here.

• Dovahkiin - Tuesday, March 1, 2016 - link

  I wouldn't be surprised, that given it's likely a development firmware, it likely has flags toggled that keep portions of the SOC at full power for debug purposes.

• Dovahkiin - Tuesday, March 1, 2016 - link

  POTENTIALLY, QC 3.0 may be a factor too.