Plausible A4 numbers

By: Maynard Handley (, October 2, 2020 10:19 pm
Room: Moderated Discussions

The cache sizes are probably untrustworthy (the A13 numbers are also wrong; GB still doesn't have a decent algorithm for testing this sizing).

The frequency is a little more plausible; GB5 seems to have a reasonable frequency tracking algorithm now (at least for Apple cores; still all over the place on highly turbo'd Intel cores).

So if one was hoping for another impressive leap forward, this is not it. Pretty much all the performance boost appears to be the result of frequency boost (thanks TSMC) and some performance boosting of the Crypto unit.

So what are we to make of this? Let me throw out some possibilities (not necessarily exclusive)

- the sky is falling. GW3 has left and no-one can do what he did.
- the sky is falling. Apple has picked up every realistic µArchitecture trick know to humanity and the fun is over.

- this was a deliberate re-steering of the platform. In phones, Apple reckons the A13's peak power/current draw mostly overshoots what phones need compared to battery life and overheating, so from now on the phone designs (perhaps like the watch designs already) will be more oriented to power. In this telling, the A14 is a very different design from the A13, using even more new tricks, smarts, and lots of logic -- but all deployed in a way that doesn't hit the same power spikes as before.
We'll see if this is true soon enough, once we see iPad and iPhone energy usage numbers for benchmarks.

- alternatively the reason the numbers are so similar is that for the most part this is the same chip, the equivalent of an Intel, a Broadwell rather than a Haswell. So mostly the exact same micro-architecture guts, maybe a few bug fixes or new tightly constrained changes (crypto?), supposedly some changes in the OS-level stuff (obviously the 4kB pages, also hypervisor stuff).
So is this a bad thing? Options include:
+ no it's reality. Apple needed to update a bunch of complicated stuff for Apple Silicon (the MMU/hypervisor stuff, maybe the usual round of even further security weirdness, maybe a bunch of invisible stuff surrounding caches, IOMMUs, general scalability to lots of IO and lots of cores) so basically they're doing in one year the sort of behind-the-scenes stuff that Intel or AMD have crawled their way to over ~20 years. Maybe it's not all fully working or fully performant, but it's good enough to get AS working up to the iMac level [8+8 cores?] and to debug for next year?

+ alternative 3. We always assumed that AS (and iPad Pro) would get essentially the same core as iPhone, even if the rest of the SoC, starting with the memory controller, differed. Perhaps they concluded that's no longer a good tradeoff; time to bifurcate the high-end cores. This doesn't necessarily have to mean total redesigns, more kinda a X1 vs A78 sort of deal, mostly the same design and core algorithms but with various key parameters (BP sizes, various queues and buffers, various buses,...) chosen as small for iPhone and big for AS?

+ alternative 4. There is a speed booster in there in the form of SVE/2, but we don't see it as providing any value because GB5 isn't compiled for it. Would be nice to believe, but I'm not optimistic. I think if it were there, at the least NEON would have expanded to 4 units, and that would have had a visible impact on the benchmark numbers. One can still hope that SVE/2 (and 4xNEON) was compiled into the AS version, not the iPhone version, but that seems like a more aggressive change. On the other hand, there is value to Apple starting AS with SVE/2 as baseline ISA, even if it's a somewhat crippled SVE (1x256 wide? 3x128 wide?)

Basically the numbers seem to validate the disappointing performance expectations Apple set at the iPad event, but don't tell us why! We'll get one more data point when we see the A14 energy consumption values, but won't see the big picture till the macs arrive.
 Next Post in Thread >
TopicPosted ByDate
Plausible A4 numbersMaynard Handley2020/10/02 10:19 PM
  Plausible A4 numbersMaynard Handley2020/10/02 10:22 PM
    Plausible A4 numbersGabriele Svelto2020/10/03 05:46 AM
  Plausible A4 numbersDoug S2020/10/02 11:53 PM
    Plausible A4 numbersMaynard Handley2020/10/03 09:24 AM
      Plausible A4 numbersDoug S2020/10/03 12:54 PM
  Plausible A4 numbersAndrei F2020/10/03 01:35 AM
    Plausible A4 numbersRober Schwärzt2020/10/03 06:53 AM
      Plausible A4 numbersAnon A. Anonymous2020/10/03 07:05 AM
    Plausible A4 numbersanon2020/10/03 07:01 AM
      Plausible A4 numbersJonas Palm2020/10/03 08:26 AM
        Gerard Williams IIIanon2020/10/03 09:26 AM
          Gerard Williams IIIMaynard Handley2020/10/03 09:30 AM
            Gerard Williams IIIanon2020/10/03 11:31 AM
              Gerard Williams IIIMaynard Handley2020/10/03 11:38 AM
                Gerard Williams IIIanon2020/10/03 01:56 PM
                  Gerard Williams IIIMaynard Handley2020/10/03 02:00 PM
          Gerard Williams IIIMichael S2020/10/03 10:33 AM
            Gerard Williams IIIanon2020/10/03 11:28 AM
            Gerard Williams IIIme2020/10/03 03:09 PM
              Gerard Williams IIIMaynard Handley2020/10/03 06:36 PM
                Gerard Williams IIISimon Farnsworth2020/10/04 02:46 PM
                  Gerard Williams IIIMark Roulo2020/10/04 06:06 PM
                  Gerard Williams IIIDoug S2020/10/04 06:35 PM
                    ICL v A14anon2020/10/04 07:43 PM
                      Based on? (NT)Rayla2020/10/04 07:45 PM
                        Dataanon2020/10/04 08:15 PM
                          Datavvid2020/10/05 01:02 AM
                            Dataanon2020/10/05 07:28 AM
                      ICL v A14Robert2020/10/05 04:25 AM
                        ICL v A14anon2020/10/05 07:34 AM
                          ICL v A14Doug S2020/10/05 10:50 AM
                            ICL v A14Anon2020/10/05 12:31 PM
                Gerard Williams IIIAaron Spink2020/10/05 09:36 AM
                  Gerard Williams IIIMaynard Handley2020/10/05 11:38 AM
                    Gerard Williams IIISimon Farnsworth2020/10/06 03:22 AM
                      Gerard Williams IIIMaynard Handley2020/10/06 09:06 AM
                    Gerard Williams IIIDummond D. Slow2020/10/06 03:55 AM
                      Gerard Williams IIIBrett2020/10/06 04:18 AM
                        Gerard Williams IIIMichael S2020/10/06 07:29 AM
                          Gerard Williams IIIBrett2020/10/07 04:45 AM
                            Gerard Williams IIIMichael S2020/10/07 05:58 AM
                            Gerard Williams IIIAdrian2020/10/07 06:09 AM
                              Gerard Williams IIIBrett2020/10/08 02:38 AM
                            Intel supply constraints and 10nmJames2020/10/08 05:56 AM
                              Intel supply constraints and 10nmGroo2020/10/08 02:43 PM
                      Gerard Williams IIIMaynard Handley2020/10/06 09:14 AM
                        Gerard Williams IIIDummond D. Slow2020/10/06 02:38 PM
                          Gerard Williams IIIMaynard Handley2020/10/06 04:33 PM
                        Gerard Williams IIIAaron Spink2020/10/06 06:08 PM
            Gerard Williams IIIMark Roulo2020/10/03 04:02 PM
              Gerard Williams IIIMichael S2020/10/03 04:24 PM
                Gerard Williams IIIMark Roulo2020/10/04 10:05 AM
    Plausible A4 numbersRonald Maas2020/10/03 08:06 AM
  Geekbench scores?Rober Schwärzt2020/10/04 02:49 AM
    Geekbench scores?Adrian2020/10/04 04:37 AM
      Geekbench scores?Robert2020/10/04 05:20 AM
        Geekbench scores?Doug S2020/10/04 07:39 AM
Reply to this Topic
Body: No Text
How do you spell avocado?