Question Apple A15 announced

[jpiniero]

Seems like the CPU or GPU is not much faster than A14. Might be due to wanting lower power draw for better battery life.

The NPU did get a bump, 15.8 vs 11.8 tops.

Edit: The Pro does get an increase to 5 GPU cores from 4. Might be useful because of the 120 Hz VRR they added.

 

[Doug S]

LOL. This kills the fanboy. Why do you think Intel harps on constantly about "AC performance"? It is so they can pretend their 15W chips can dish out the performance even while it sucks down twice as much. It is also how this guy can actually write something like:

Well they DO dish out that performance, so long as you don't have a long running task that needs that performance level the entire time.

If you want to avoid that you choose for your testing a long running task where the initial turbo boost will be lost in the noise over a several hour long run. If you think running a noisy high speed fan is cheating, disable the fan on both for that test.

There are ways to get the outcome as "fair" as you want, but don't complain if others find that to be just as "pretend" as you consider tasks brief enough the Intel CPUs far exceed their 15W TDP.

It is well known that Geekbench is a very short duration and bursty benchmark, which will slant the results in hardware that can sustain a higher performance level during that run than it can sustain for longer periods. If you ran Geekbench in a loop over and over again then you get a different look at things, especially if you also count the runs until the battery runs out.

I would argue that Geekbench is a pretty decent benchmark for phones, in that at least it fits the usage model for most people. Few people peg the CPU for very long in their phone. PCs are a different matter and while some people have pretty phone-like usage pattern on their PC, others will peg it for long periods and benchmarks with longer duration like SPEC2017 are more appropriate.

 

[dmens]

Well they DO dish out that performance, so long as you don't have a long running task that needs that performance level the entire time.

If you want to avoid that you choose for your testing a long running task where the initial turbo boost will be lost in the noise over a several hour long run. If you think running a noisy high speed fan is cheating, disable the fan on both for that test.

There are ways to get the outcome as "fair" as you want, but don't complain if others find that to be just as "pretend" as you consider tasks brief enough the Intel CPUs far exceed their 15W TDP.

It is well known that Geekbench is a very short duration and bursty benchmark, which will slant the results in hardware that can sustain a higher performance level during that run than it can sustain for longer periods. If you ran Geekbench in a loop over and over again then you get a different look at things, especially if you also count the runs until the battery runs out.

I would argue that Geekbench is a pretty decent benchmark for phones, in that at least it fits the usage model for most people. Few people peg the CPU for very long in their phone. PCs are a different matter and while some people have pretty phone-like usage pattern on their PC, others will peg it for long periods and benchmarks with longer duration like SPEC2017 are more appropriate.

They dish out the performance while consuming grossly more energy than advertised. The reason Intel blabbers on about performance with AC power is because it hides the non-existent battery life when dishing out said performance. In addition, both AMD and Apple can dish out the same level of performance, with less power consumed.

The issue is not fairness, it is clarity of information. Work done over time and energy consumed are the only metrics that matter, not dumb things like absolute performance over short bursts which are easy gamed. I realize absolute performance at any cost is the kind of benchmark that gets fanboys wet but no CPU/GPU/xPU designer/architect cares the slightest about that.

 

[GC2:CS]

Teardowns report LPDDR4X.

Die size about 108 mm2.

Apple iPhone 13 Pro Teardown | TechInsights

We received the iPhone 13 and iPhone 13 Pro in our labs and identification and early costing analysis are available below.

www.techinsights.com

 

[Doug S]

Teardowns report LPDDR4X.

Die size about 108 mm2.

Apple iPhone 13 Pro Teardown | TechInsights

We received the iPhone 13 and iPhone 13 Pro in our labs and identification and early costing analysis are available below.

www.techinsights.com

Because like I've said it is basically an unchanged A14. An extra GPU core, more cache (I think I heard?) and a few tweaks for stuff like nested virtualization (which may have been present before and just not enabled or needed to be fixed before it could be enabled)

If you're using basically the same SoC you're not going to swap out for an LPDDR5 controller.

The real upgrade comes with Jade-C, and it will include the cores we'll see in A16...

 

[uzzi38]

Because like I've said it is basically an unchanged A14. An extra GPU core, more cache (I think I heard?) and a few tweaks for stuff like nested virtualization (which may have been present before and just not enabled or needed to be fixed before it could be enabled)

If you're using basically the same SoC you're not going to swap out for an LPDDR5 controller.

The real upgrade comes with Jade-C, and it will include the cores we'll see in A16...

...

A14 is 88mm^2. It's a significantly larger die.

 

[roger_k]

Andrei being cryptic and click-baity

https://twitter.com/x/status/1443362561089490944

 

[nxre]

Andrei being cryptic and click-baity

https://twitter.com/x/status/1443362561089490944

very large jump in CPU? that's interesting, I though IPC was unchanged and only frequency increased. Maybe we're missing something

 

[sallymander]

very large jump in CPU? that's interesting, I though IPC was unchanged and only frequency increased. Maybe we're missing something

I'm guessing it's an improvement in efficiency - the A15 is running at slightly higher clockspeeds than the A14 (in iPhones at least) while getting much better battery life.

They did increase the size of the batteries in the new iPhones too but the iPhone 13 Mini gets more battery life than the regular iPhone 12.

 

[Mopetar]

I'm guessing it's an improvement in efficiency - the A15 is running at slightly higher clockspeeds than the A14 (in iPhones at least) while getting much better battery life.

They did increase the size of the batteries in the new iPhones too but the iPhone 13 Mini gets more battery life than the regular iPhone 12.

Are we sure all of that's just down to the SoC though? There's other hardware that uses battery in a phone and there's also firmware and software improvements that can make a difference as well.

Normally no one cares since they're buying a phone, not an A-series SoC, but it's important when it comes to discussions about just the CPU.

I'm not saying Apple didn't make any improvements, but it seems like the core design is largely the same, as are the clocks. Performance improvements would have to come from dedicated hardware improvements or Apple being able to better utilize parts of the SoC other than the CPU cores for that performance. Outside of the TSMC node maturing and allowing for better bins, there isn't any other way I can think of to square what we've seen with Andrei's tweet.

 

[Doug S]

...

A14 is 88mm^2. It's a significantly larger die.

A15 doubled the size of the system cache to 32MB and added a GPU core. The system cache appears to be around 9 mm^2 in A14 die photos, so doubling that makes an extra 9 mm^2. The four GPU cores appear to be about 12 mm^2, so a fifth adds 3 mm^2.

That's already ~12 of the 20 mm^2 it grew, without accounting for similar growth in the NPU, IPU, etc. Supposedly the NPU remains 16 cores but is 15.8 TOPs vs 11 TOPs, so they may have improved (and therefore likely enlarged) its NPU cores as they are unlikely to gain 43% by clock rate alone.

There have also been improvements in the IPU, video encode/decode blocks etc. which would consume additional area, and miscellaneous improvements in the other functionality that's too small to identify or mark on a die photo. Plenty of room for growth outside the recycled CPU cores and (presumably) GPU cores.

 

[eek2121]

Thank you, this is most interesting.

But now I am completely confused. Why are scores for the 15W and the 35W Cesanne this close? Why does 65W Vermeer perform significantly better? I feel like there is something essential missing...

P.S. Look at the difference in scores for the 15W and the 28W Tiger Lake — whopping 30%. That makes much more sense IMO.



Regardless of our personal preferences and taking sides, I just want to get to the truth: how energy efficient are various modern architectures and their configurations when doing computational work and how does that energy efficiency relate to their peak and the sustained performance. I mean, we are not talking ethics here, this kind of question should have an objective, empirical answer, right?

I think the story behind Apple CPUs is fairly straightforward. Tiger Lake is also fairly clear. But AMD story is just muddy. Something doesn't check out.

Of course something is off. I have a laptop with a 5900hx and it uses no more than 45W…even for bursty performance. To say the 5800u behaves differently from every other AMD chip is a bold claim.

Unsure what the story is with the charts. I suspected SoC power was being included initially. x86 chips have more complicated memory controllers and have to support multiple PCIE4 lanes. However, given the chart shows a burst and then a drop (something i haven’t seen on all 6 of my Ryzen 1, 2, and 3 chips), it makes me wonder if something is misconfigured.

 

[roger_k]

Of course something is off. I have a laptop with a 5900hx and it uses no more than 45W…even for bursty performance. To say the 5800u behaves differently from every other AMD chip is a bold claim.

Who is saying that 5800U behaves differently? And 45W is obviously the turbo limit for 35W Cezanne chips, this should be obvious from the graphs that have been linked. I suppose that's what you have?

Anyway, since you own the hardware, could you do us a favor and run a couple of experiments? If you limit the maximal power consumptions (not sure how to do it with AMD CPUs, there must be some sort of utility) to 15W and run some benchmarks on your 5900hx, what kind of results do you get?

 

[itsmydamnation]

Who is saying that 5800U behaves differently? And 45W is obviously the turbo limit for 35W Cezanne chips, this should be obvious from the graphs that have been linked. I suppose that's what you have?

Anyway, since you own the hardware, could you do us a favor and run a couple of experiments? If you limit the maximal power consumptions (not sure how to do it with AMD CPUs, there must be some sort of utility) to 15W and run some benchmarks on your 5900hx, what kind of results do you get?

I have a 5800-H ( 8 cores - 16 threads)
just done some CB23 runs
few caveats, I currently have 322 processes running including ~ 4 web server frame works, ~ 16 connections open to mysql , teams , webex, etc, so my idle is hardly idle ( my idle is about 6-8% CPU utilisation @ 3.5ghz)
also remember lots of these laptops have junk memory , mine is 16GB 3200 22:22:22:52 , for example my cheap desktop memory is 3600 16:19:19:39


13watt CPU
multi core clock: 2.1ghz
per core wattage: 1.7
multi score: 6418

pinned to a single core
single core clock:4.1ghz
single core wattage: 7 watt
single core score: 1202


15watt CPU
multi core clock: 2.27ghz
per core wattage: 1.9
multi score: 6856

multi core NO SMT* clock: 2.6ghz
per core wattage: 1.9
multi score: 6573


pinned to a single core
single core clock:4.15ghz
single core wattage: 8 watt
single core score: 1253



25watt CPU
multi core clock: 2.7ghz
per core wattage: 3
multi score: 7494

multi core NO SMT* clock: 3.05ghz
per core wattage: 3
multi score: 7083

pinned to a single core
single core clock:4.15ghz
single core wattage: 8 watt
single core score: 1253 ( temp limited)

unpinned single core
single core clock:4.15ghz
single core wattage: 8 watt
single core score: 1283

35watt CPU
multi core clock: 3.1ghz
per core wattage: 4.3
multi score: 9225


will run higher wattage tests and update



* via affinity so still 16 processes running

ZOMG amd SMT improves PPW ..... intel P cores in shambles

 

[roger_k]

I have a 5800-H ( 8 cores - 16 threads)
just done some CB23 runs
few caveats, I currently have 322 processes running including ~ 4 web server frame works, ~ 16 connections open to mysql , teams , webex, etc, so my idle is hardly idle ( my idle is about 6-8% CPU utilisation @ 3.5ghz)
also remember lots of these laptops have junk memory , mine is 16GB 3200 22:22:22:52 , for example my cheap desktop memory is 3600 16:19:19:39


13watt CPU
multi core clock: 2.1ghz
per core wattage: 1.7
multi score: 6418

pinned to a single core
single core clock:4.1ghz
single core wattage: 7 watt
single core score: 1202


15watt CPU
multi core clock: 2.27ghz
per core wattage: 1.9
multi score: 6856

multi core NO SMT* clock: 2.6ghz
per core wattage: 1.9
multi score: 6573


pinned to a single core
single core clock:4.15ghz
single core wattage: 8 watt
single core score: 1253



25watt CPU
multi core clock: 2.7ghz
per core wattage: 3
multi score: 7494

multi core NO SMT* clock: 3.05ghz
per core wattage: 3
multi score: 7083

pinned to a single core
single core clock:4.15ghz
single core wattage: 8 watt
single core score: 1253 ( temp limited)

unpinned single core
single core clock:4.15ghz
single core wattage: 8 watt
single core score: 1283

35watt CPU
multi core clock: 3.1ghz
per core wattage: 4.3
multi score: 9225


will run higher wattage tests and update



* via affinity so still 16 processes running

ZOMG amd SMT improves PPW ..... intel P cores in shambles

Thank you! All of these are significantly lower than usually reported scores, so your results either confirm that AMD CPUs indeed run way above their TDP levels for the duration of the tests , or that the background processes on your laptop make the results unreliable. At any rate, performance scaling makes sense to me.

What results do you get if you use the default AMD power limit config?

P.S. Anandtech reports 15W 4800U CB23 multicore as 9300, which is close to your 35W limit. Assuming that your background processes eat up 10% or so of possible score this is actually consistent with the notion that a 15W Ryzen config runs closer to 30W when running these benchmarks…

 

[BorisTheBlade82]

Okay, I did some measurements with an AMD R7 4700U (Renoir, 8C/8T) within a cTDP range of between 10w and 30w:

It shows the duration of a CB23MT run, the total consumption in Joules or Wattsconds, the measured avg. Power (almost equal to the cTDP) and the So called PES. The latter basically being a combination of Performance and Efficiency weighted 1:1.

What do I see wrt to these numbers:

• The SKU is the most power efficient when running with 12w.
• But the sweet spot wrt performance-efficiency seems to be around 15 to 16w.
• With less than 12w the overhead of the uncore starts to limit the power efficiency.

This leads me to believe, that not only did AMD design the Renoir APU for a sweet spot of around 15w. That in itself is not possible without taking into consideration the core design.
They rather designed the Zen2 core to be the most performance efficient with around 1,5w. This makes sense for the mobile space as well as for the server market. There they employ 64c with a TDP starting from 225w. So after the IOD (and a huge consumption of the Interconnect via package) there are no more than around 2w per core left.

BTW
If you like more number crunching wrt power and performance efficency or if you might try it yourself have a look at https://github.com/BrsVgl/PerformanceEfficiencySuite
There are a lot of interesting samples including Cézanne, Vermeer, ICL, TGL-U, TGL-8C, RKL, Tremont et al.

 

[itsmydamnation]

Thank you! All of these are significantly lower than usually reported scores, so your results either confirm that AMD CPUs indeed run way above their TDP levels for the duration of the tests , or that the background processes on your laptop make the results unreliable. At any rate, performance scaling makes sense to me.

What results do you get if you use the default AMD power limit config?

P.S. Anandtech reports 15W 4800U CB23 multicore as 9300, which is close to your 35W limit. Assuming that your background processes eat up 10% or so of possible score this is actually consistent with the notion that a 15W Ryzen config runs closer to 30W when running these benchmarks…

few things remember:
I have a H not a HS , HX or U so i would have one on the worse side for binning for low wattage,
second is the trash ram, its pretty bad.
3rd i cant control the CPU i can only set the bounding parameters , so voltage behaviour could be very different on a U for example.
if I can be bothered i will test on my 4700U, i might have some previous data on this forum.

That said for my work ( i write a lot of perl + sql) my Zen3 is way faster then my Zen2 like probably a factor of 2x at the same number of threads ( 8). so that all scalar int and FP ( alot of FP actually)

 

[nxre]

Per anandtech:
A15 Performance cores are roughly ~12% faster but consume much less energy, leading to ~17% improvement in energy efficiency.
A15 Efficiency cores are roughly ~28% faster at same energy consumption.

Seems to contradict the previous reports that energy consumption was up from the A14.

 

[insertcarehere]

Per anandtech:
A15 Performance cores are roughly ~12% faster but consume much less energy, leading to ~17% improvement in energy efficiency.
A15 Efficiency cores are roughly ~28% faster at same energy consumption.

Seems to contradict the previous reports that energy consumption was up from the A14.

Seems like the increased L2 cache is contributing to increasing performance & efficiency, bodes pretty well for M1X moving forward.

 

[Annas231]

So bottom line...there is any laptop , that draws at maximum 25W at outlet that is better than M1 macbook , in terms of performance cpu+gpu ?
Thank you

 

[DrMrLordX]

So bottom line...there is any laptop , that draws at maximum 25W at outlet that is better than M1 macbook , in terms of performance cpu+gpu ?
Thank you

Kind of the wrong thread for that. Otherwise: faster in what software? You may find it a little difficult to bench M1 vs Wintel laptops. It's easiest to bench M1 against older Intel-based Macs.

 

[BorisTheBlade82]

So bottom line...there is any laptop , that draws at maximum 25W at outlet that is better than M1 macbook , in terms of performance cpu+gpu ?
Thank you

Let me answer this in my position as a newbie: Basically no!
Hands down, in an Apples to Apples comparison (pun intended) there is simply no competitor to the M1 wrt Performance and power Efficiency weighted 1:1 - at least under ST. In MT Renoir and Cezanne are strong competition because they simply have more big Cores and SMT as well. But I suspect that M1X will wipe the floor with them - and not because of a process advantage.
Regarding the iGPU there is no competition no matter how you look at it.
*Preparing to get roasted*

 

[Leeea]

So bottom line...there is any laptop , that draws at maximum 25W at outlet that is better than M1 macbook , in terms of performance cpu+gpu ?
Thank you

Hell no, but:

m1 macbook pulls around 50 watts from the wall under max load* with full battery.
If it has a battery to charge, that apparently can increase to 87 watts from the wall.

So basically, your asking for the impossible, and yea, not even mac books perform to your expectations.

*Preparing to get roasted*

You are correct, but only because Apples own products do not live up to the expectations set.

Even the m1 mac mini, with no screen, no battery, no etc pulls 39 watts from the wall under full load**.



*https://www.google.com/search?q=m1+macbook+50+watts
**https://www.google.com/search?q=m1+mac+mini+39+watts

 

[moinmoin]

*Preparing to get roasted*

You won't get roasted, but...

and not because of a process advantage.

Not solely you mean? Because process advantage is of course a contributing advantage, one of many advantages Apple has assembled over time.

 

[Doug S]

Hell no, but:

m1 macbook pulls around 50 watts from the wall under max load* with full battery.
If it has a battery to charge, that apparently can increase to 87 watts from the wall.

So basically, your asking for the impossible, and yea, not even mac books perform to your expectations.


You are correct, but only because Apples own products do not live up to the expectations set.

Even the m1 mac mini, with no screen, no battery, no etc pulls 39 watts from the wall under full load**.



*https://www.google.com/search?q=m1+macbook+50+watts
**https://www.google.com/search?q=m1+mac+mini+39+watts

I wonder what Apple bases that 39 watt max on? The highest figure they could measure from real world code, the highest synthetic load they could come up with (i.e. power virus for all CPU cores, GPU cores, NPU cores, maximum memory traffic, network traffic, etc.) or that was simply the sum of the max draw of all components.

Unless we know what that 39W figure represents, it is in no way comparable to any figures we have for AMD and Intel systems. Given the jump to 50W for the Macbook you can bet that measurement includes "display at max brightness".

 

[Annas231]

Hell no, but:

m1 macbook pulls around 50 watts from the wall under max load* with full battery.
If it has a battery to charge, that apparently can increase to 87 watts from the wall.

So basically, your asking for the impossible, and yea, not even mac books perform to your expectations.


You are correct, but only because Apples own products do not live up to the expectations set.

Even the m1 mac mini, with no screen, no battery, no etc pulls 39 watts from the wall under full load**.



*https://www.google.com/search?q=m1+macbook+50+watts
**https://www.google.com/search?q=m1+mac+mini+39+watts

thats applies the same for every other laptops ...
So, what other laptop can outperform the m1 macbook at 50W outlet ?
But the others already answer me, but thank you anyway