RAM is not part of the SoC chip, so I don’t quite follow.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
M3 Chip Generation - Discussion Megathread
- Thread starter scottrichardson
- Start date
- Sort by reaction score
RAM is not part of the SoC chip, so I don’t quite follow.
Sometimes it’s worth reading some of the thread rather than diving right in.
The main reason behind the gap is because Apple still use LPDDR5 on A17 Pro, I guess Apple will move to LPDDR5X next year with A18 Pro, then we should be expecting bigger jump on GPU
Good! I hope this will improve the state of the Android gaming.
I have a galaxy s23 ultra and iPhone 14 Pro Max and the Samsung usually has better gaming performance especially when it comes to sustained performance.
And you can see it in GFXbench as well.
Attachments
As I said. Read back on some of the thread. The Qualcomm gpus seem like they have much less precision and compute. Hence the huge disparity in geekbench scores. Anandtech explained the issue with visuals a couple of years ago.
The Apple A15 SoC Performance Review: Faster & More Efficient
He is talking about Genshin Impact, they usually have increased resolution and some effects have higher quality in iPhone, but Geekerwans tune up the settings so they get about the same resolution and therefore in that game the performance results are fair, not benchmarks like GFXBench and 3DMark they do the exact same quality in thoseAs I said. Read back on some of the thread. The Qualcomm gpus seem like they have much less precision and compute. Hence the huge disparity in geekbench scores. Anandtech explained the issue with visuals a couple of years ago.
The Apple A15 SoC Performance Review: Faster & More Efficient
www.anandtech.com
View attachment 2280467
The DRAM is put in the same package, in order to be close. (That's how to get the highest speeds.) So the package size dictates how big a DRAM chip(s) is in it. And the transistor size dictates how many bytes a DRAM can be.
I think it’s more of a cost issue than anything else. Package substrate could be made larger etc. What I am not sure is whether LPDDR supports adding more capacity to the same channel in the same way DDR does (it’s possible that per-channel capacity is the real hard constraint).
Btw, RAM on package is not to maximize speed (Apples RAM is not any faster than other LPDDR products) but again to save costs and reduce power consumption.
With all the unenlightening back and forth about whether A17 really is faster, no it isn't, no it is but only at the cost of power, well but that's only (realistically) in short bursts, blah blah; an obvious data source is being omitted!
What do the numbers look like for an A17 running either JetStream 2.1 or Speedometer 2.0?
Browser benchmarks have the great advantages that
(a) they are easy to run, by anyone
(b) they correspond to a workload we KNOW Apple really cares about for the phone (as opposed to both GB and SPEC which each in their own way don't really match how people use phones)
(c) they will stress code that's written and compiled by Apple to be a best fit for the A17. If there is functionality in the A17 that increases IPC, but requires special, or at least different code (either generated by the compiler, or otherwise exploited, eg in the JIT) we will see it there in a way we won't in GB6 (not yet recompiled for A17) or SPEC (who knows how that's been compiled by the sites that claim to be running it on iPhone?)
So how about it, people? If you own an iPhone 15 Pro/Max (or for that matter, something with an A16 or A15 AND you are running iOS latest, so the Safari is up to date!) go to
https://browserbench.org/JetStream/ and/or
and/or
and give us your numbers (trying to be as honest as possible; eg don't do this while you are also doing something else in the background!)
What do the numbers look like for an A17 running either JetStream 2.1 or Speedometer 2.0?
Browser benchmarks have the great advantages that
(a) they are easy to run, by anyone
(b) they correspond to a workload we KNOW Apple really cares about for the phone (as opposed to both GB and SPEC which each in their own way don't really match how people use phones)
(c) they will stress code that's written and compiled by Apple to be a best fit for the A17. If there is functionality in the A17 that increases IPC, but requires special, or at least different code (either generated by the compiler, or otherwise exploited, eg in the JIT) we will see it there in a way we won't in GB6 (not yet recompiled for A17) or SPEC (who knows how that's been compiled by the sites that claim to be running it on iPhone?)
So how about it, people? If you own an iPhone 15 Pro/Max (or for that matter, something with an A16 or A15 AND you are running iOS latest, so the Safari is up to date!) go to
https://browserbench.org/JetStream/ and/or
and/or
and give us your numbers (trying to be as honest as possible; eg don't do this while you are also doing something else in the background!)
From a digital circuit signal (in terms of data and address lines) point of view, there's no difference betwen LPDDR vs DDR. Apple could very well have dumped tons of LPDDR modules into the main board and have multiplexing circuitry for the data lines, controlled by the address lines. This is a very simplistic explanation. Might as well go DDR then.
Drawback is there will be tons of circuit lines which will be a nightmare to route from memory module to multiplexing components, and power consumption will likely increase as longer data lines means higher voltage to drive, and higher latency still, due to the need for data and address lines to settle down.
Great idea. Damn I should have run this test today in store as they had the 15 pro there.
My M2 Pro Mac Mini gets 456 +/- 22 on the first run and 474 +/- 7.8 on the second. Subsequent runs are around the second number.
Gaussian Blur is most certainly is going to be limited by RAM throughput, and since they use the same RAM I'd expect the scores to be the same.
So to put this in context the most recent numbers I can find are
- A13 w/iOS15 177
- A14 w/iOS15 207
- A15 w/iOS15 238
- A16 w/iOS16 ~360..370
(some claims that recent versions are as high as 400)
There is a problematic *CPU* benchmark because it includes such a large element of software-level improvements, but however you cut it, there seems to be a real A17 vs A16 boost by at least 60 points (~15%); pretty impressive on real code, and seems inline with the sorts of improvements we have seen historically. (Including how the whole "nothing changed" series of 15/16/17 seem to magically double the score while supposedly doing "nothing" for performance...)
Keep submitting more numbers (more iphones, more of the set of three, some of which are more GPU centric) so we can see what the real patterns are.
Big core area reduced by 20%, small core area reduced by 20%. GPU single-core area increased by 5%, and GPU overall area increased by 20%.
https://twitter.com/Tech_Reve/status/1706472603752575103/photo/1
Does this seem credible? Is an iPhone 15 2x faster than an iPhone 13? (Asking as a 13 mini user...)
Why not read what I said and THINK about it?
There seems to have been a substantial improvement in Safari SW between iOS15 and iOS16, so that
- an iPhone 17 TODAY is 2x as fast (for Speedometer 2) as an iPhone 15 when released
- an iPhone 17 TODAY is not 2x as fast (for Speedometer 2) as an iPhone 15 TODAY. But it is SUBSTANTIALLY faster, much more so than you would imagine from the idiot internet.
To get a full picture, we need more numbers.
My A12 iPhone gets 205, and my A15 gets 369. I *think* the big jump happened between A14 and A15 with the A15 hardware acceleration of JIT-style indirect pointers.
Even so, 369 to 460 is another impressive boost. We need an A16 number to see how those 90 points are distributed between the 15->16 and 16->17 transition, but from the limited data I can find on the internet, it seems 2/3 of the boost comes from 16->17.
Yep, my iPhone 14 Pro gets almost 400, but 380 on first run. Last run 399 +/- 4.8
Edit: on iOS 16.6.1.
Last edited:
My 15 Pro Max got 443 on the first run of Speedometer but averaged about 472 on multiple repeated runs after that. Results for all tests attached.
Attachments
It is just numbers from some benchmarks. Depending on what and how you use your phone you might not see any difference or you might see a big difference. I would suspect that with typical usage you would not notice any speed difference. It is mostly a “how many angels can dance on the head of a pin?” type of debate. Some find it interesting but its practicality might be rather limited.
Last edited: