Hacker News Reader

Hacker News

Comment by MR4D | Hacker News Reader

Parent Full thread

MR4D

>Eh? This is a flimsy claim.

Anandtech would like to disagree with your word choice.

https://www.anandtech.com/show/14892/the-apple-iphone-11-pro...

View on HN

Did you read your own link?

> In virtually all of the SPECint2006 tests, Apple has gone and increased the peak power draw of the A13 SoC; and so in many cases we’re almost 1W above the A12. Here at peak performance it seems the power increase was greater than the performance increase, and that’s why in almost all workloads the A13 ends up as less efficient than the A12.

> The total power use is quite alarming here, as we’re exceeding 5W for many workloads. In 470.lbm the chip went even higher, averaging 6.27W. If I had not been actively cooling the phone and purposefully attempting it not to throttle, it would be impossible for the chip to maintain this performance for prolonged periods.

In other words, to get those good specint numbers, power was sacrificed to do it. 5W per-core power draw is right in line with a typical x86 laptop CPU, too. 4800HS sits at 35w, or 4.3w per core.

You do realize that's 5W total, not 5W per core? Apple is currently around 4-5x the performance per watt of Intel's most efficient parts.

It's a single-threaded test. That's 5W on a single core.

Here's the A12's power/frequency curve: https://images.anandtech.com/doci/14892/a12-fvcurve.png

That's the curve per core (that's how those charts work). 3.85W per core @ 2.5ghz is what a big core A12 is spec'd at.

That's fair. However, according to the same chart, it draws 1W at 2000 MHz, which should give 80% the single core performance. This is how Apple is getting 4x the perf/watt of Intel and AMD competitors in Geekbench and similar workloads. Apple is able to acheive very high multithreaded performance within a 5-6W TDP by running all the cores around 80% of peak performance.

Those curves also apply to Intel & AMD. As in, you can drop frequency on AMD to also achieve significant improves in perf/watt. That's not a unique aspect of the A12. That curve is more "this is how TSMC's 7nm transistors behave" type of thing.

> 4x the perf/watt of Intel and AMD competitors in Geekbench

Geekbench only measures perf, not perf/watt. It does not try to achieve maximum perf/watt, nor has Apple tuned to the A12/A13 to achieve maximum perf/watt in Geekbench either. Geekbench's single thread numbers where it "competes with Intel & AMD" are also these ~5W per core power figures.

I'm not sure where you're getting this random 4x better number from anyway?

Researching the very most efficient parts from Intel and AMD today, 3x the perf/watt would be more accurate. Operating at a 5-6W TDP, the 2018 Apple A12X gets a multicore Geekbench score of 4730. Operating at a TDP of 15W, the i7-1065G7 (Ice Lake) gets multi-core Geekbench score of 4865. This is on Intel's 10 nm process that's comparable to TSMC 7 nm. Near equal performance for 3x the power.

I'd expect a 2020 A14X or whatever it's called to comfortably beat what they could achieve in 2018, so getting 4-5x the perf/watt of Intel and AMD's best is what I'd expect when operating at similar points in the frequency/power curve. The A12X was around 4-5x the perf/watt of what Intel and AMD had out in late 2018.

https://browser.geekbench.com/v5/cpu/2639065 https://browser.geekbench.com/v5/cpu/2638528

Unfortunately I could only find this old chart [0] showing how power draw scales with frequency on Intel. However for the sake of demonstration it should be more than enough.

The chip needs around 25W at 2.5GHz and 200W at 4.7GHz. 8x more power for 1.88 times the performance. In other words Intel chips running at 2.5GHz are 4.25 times more efficient than Intel chips running at 4.7Ghz. No magic. Once Apple has chips that go this far they will suffer from the same problems.

Here is a slightly newer chart [1] that demonstrates a 57% increase in power consumption for a 500Mhz frequency gain (12% performance gain).

[0] https://www.extremetech.com/wp-content/uploads/2018/09/Clock...

[1] https://www.extremetech.com/wp-content/uploads/2016/02/power...

> Those curves also apply to Intel & AMD. As in, you can drop frequency on AMD to also achieve significant improves in perf/watt. That's not a unique aspect of the A12. That curve is more "this is how TSMC's 7nm transistors behave" type of thing. Geekbench only measures perf, not perf/watt. It does not try to achieve maximum perf/watt, nor has Apple tuned to the A12/A13 to achieve maximum perf/watt in Geekbench either. Geekbench's single thread numbers where it "competes with Intel & AMD" are also these ~5W per core power figures.

While similar curves also apply to Intel and AMD, their mobile parts are drawing 10-20 watts per core to acheive the very top results. When you're using all the cores together under a 6W TDP (as apple is doing), Apple is able to achieve a much higher Geekbench result than Intel or AMD parts set to a comparable TDP, or even 3x the TDP. Compare multi-core Geekbench scores of Apple's parts running at a 6W TDP to Intel or AMD's most efficient parts running at a 15W TDP, and you'll see that Apple outperforms them while drawing 1/3 the power. Similar curves apply, but Apple can achieve far morea at 1W per core than any x86 competitors.

If it was that easy then everyone would do it. I call this the curse of the single number. There is this complicated machinery with lots of parts with different shapes, some are bigger some are smaller. However, the customer is not aware of the complexity and only sees a single number like 5W and maybe another number that showcases the performance score of the chip. Surely, since that is the only information we have about power consumption and performance it must be true in all situations. The reality is that those two numbers were measured during different situations and combining them into a meaningful calculation might actually not be possible.

For example. Geekbench measures peak performance of all cores at the same time and the power draw may go above 5W.

The 5W TDP may refer to normal day to day use where one or two cores are active at the same time for the duration of the user interaction (play a game for 5 min or something) and once the user stops using the phone it will quickly go back to a lower TDP.

You're comparing peak single with all-core TDP/cores. Ryzen 2 cores peak at about 10 Watts per core...they probably have at least 30% frequency increase headroom...plus whatever bump they get when they move to 5 nm. I expect at least a 50% per core performance increase when they refresh their laptop/desktop lines, and probably twice the cores. And they'll also be saving a few hundred dollars per laptop...

Look at the A12's power/frequency chart: https://images.anandtech.com/doci/14892/a12-fvcurve.png

Increase frequency by 30% and the A12 will also be hitting 10 watts per core - the end of that graph is going real vertical real fast.

Since Zen 2 and A12/A13 are all on the same TSMC process this shouldn't be that surprising...

> I expect at least a 50% per core performance increase

Based off of what evidence? That'd be an unheard of improvement. TSMC isn't even claiming anything close to that at a pure transitor switching frequency for 5nm? They are predicting 15% frequency gain (at the same complexity and power) or a 20% power reduction (at the same frequency and complexity) over their 7nm process.

> They are predicting 15% frequency gain (at the same complexity and power)

My intuition is that 50% might be overoptimistic. But going from iPad to Laptop thermal constraints, you'd expect a big increase in frequency just from clocking the thing higher, no?

Apple would lose their edge over Intel because most of the efficiency gains come from the lower frequency.

> Apple would lose their edge over Intel because most of the efficiency gains come from the lower frequency.

Presumably that would only apply when the cores are actually running at full-throttle though. For casual use there could still be considerable gains if the processors are better at power management (which they most likely are, as they've had to hyper-optimised for this for phones).

I meant cumulatively. I'm assuming a "A14x+" on 5 nm. Comparison to A13.

So a A14x+ (15 W+)

30% frequency increase vs A13, because of higher TDP +

10% IPC increase (larger caches, design tweaks) +

15% Frequency increase, due to 5 nm.

A i7-1065G7 costs Apple something like $400. The A13 costs something like $60 to manufacture. Apple is highly motivated.

Yes. I saw that there were already other responses to your comment, but I'd like to add my own, quoted from the conclusion of the article:

"But the biggest surprises and largest performance increases were to be found in the A13's GPU. Where the new chip really shines and exceeds Apple’s own marketing claims is in the sustained performance and efficiency of the new GPU. Particularly the iPhone 11 Pro models were able to showcase much improved long-term performance results, all while keeping thermals in check. The short version of it is that Apple has been able to knock it out of the park, delivering performance increases that we hadn’t expected in what's essentially a mid-generation refresh on the chip manufacturing side of matters."

One of the key phrases is "...Apple has been able to knock it out of the park..."

The rest of the article is pretty clear - Apple gets it, and beats their competitors pretty soundly.

Apple is definitively competent at chip design but the end result won't be leagues ahead. They might be something like 10% ahead in terms of IPC and another 10-20% just because they get early access to 5nm compared to whatever "ancient" process Intel is using.

Everyone seems to think about cost and speed all the time.

I think that's only part of the story, which also needs to include the ability to add features and control the entire feature set across chips on all of their devices.

Encryption, ML, graphics, power management, security, etc. are all things that Apple can now add or remove as needed.

The level of optimization they can do is now well beyond just speed and price.

Right now absolutely everyone else is ahead of Intel in the NM race, with some currently shipping chips two process nodes ahead and several announced chips going so far as being twice as small as Intel's current process node. This is partially why AMD has been able to offer laptop CPUs that rival Intel's desktop offerings for less cost.

Reply on news.ycombinator.com