Apple A9X Geekbench

[Nothingness]

So what? The difference between Skylake-U and ARM tablet chips we see in this benchmark is similar to Kraken and Octane (SP4 Review). Actually it's even smaller.

When a benchmark shows something very different from other benchmarks, then one has to question it. That Speed Shift increase is a good example of that: when Kraken and Octane show no improvement, WebXPRT looks odd. If you don't question that result, you're being naive 😉

 

[ShintaiDK]

Apparently the A9X GPU is Iris Pro 5200-class:

Not entirely Apples to Apples due to FP16 and FP32. But with 50GB/sec memory it better be fast.

 

[tempestglen]

Anandtech will test SPEC06 for A9X and Skylake, plz wait!

 

[Sweepr]

When a benchmark shows something very different from other benchmarks, then one has to question it. That Speed Shift increase is a good example of that: when Kraken and Octane show no improvement, WebXPRT looks odd. If you don't question that result, you're being naive 😉

Well here's what AnandTech said:

The big win here though is WebXPRT. WebXPRT includes subtests, and in particular the Photo Enhancement subtest can see up to a 50% improvement in performance. This bumps the scores up significantly, with WebXPRT 2015 showing an almost 20% score increase, and WebXPRT 2013 has a 26% gain. These leaps in performance are certainly the kind that would be noticeable to the end user manipulating photographs in something like Picasa or watching web-page based graph adjustments such as live stock feeds.

It's probably the case that this benchmark does benefit more from SpeedShift, not necessarily (sponsored by) Intel optimizations to improve the scores in a 2-year old benchmark like you're suggesting.

 

[Nothingness]

Not entirely Apples to Apples due to FP16 and FP32. But with 50GB/sec memory it better be fast.

It could be argued that Iris Pro 5200 with its 50 (eDRAM) +25 (DRAM) GB/s is even better positioned and Intel made that claim

Intel claims that it would take a 100 - 130GB/s GDDR memory interface to deliver similar effective performance to Crystalwell since the latter is a cache. Accessing the same data (e.g. texture reads) over and over again is greatly benefitted by having a large L4 cache on package

Most likely marketing BS but heh 😛

 

[dahorns]

When a benchmark shows something very different from other benchmarks, then one has to question it. That Speed Shift increase is a good example of that: when Kraken and Octane show no improvement, WebXPRT looks odd. If you don't question that result, you're being naive 😉

I'm not sure I'd call a 2.5% and 4% improvement "no gain[,]" especially considering that we've become accustomed to ~5-10% IPC improvements per generation for Intel's Core. Your point is, of course, still valid. I'm also having a hard time understanding why Speed Shift should make such a huge difference in that benchmark. It is essentially just about getting the processor to accelerate more quickly to max speed. You'd think that outside the first 30-100 ms, Speed Shift would have no effect. I assume that WebXPRT must be a start-stop benchmark, causing the processor to fluctuate between p-states.

 

[Nothingness]

It's probably the case that this benchmark does benefit more from SpeedShift, not necessarily (sponsored by) Intel optimizations to improve the scores in a 2-year old benchmark like you're suggesting.

Without a lack of any explanation, I keep my doubt and just don't trust that result.

And BTW don't you think Intel was testing Speed Shift 2 years ago?

Anyway enough of that, we're off topic here, and as I have no concrete proof for my claims, let's stop it here 🙂

 

[witeken]

Great numbers, hopefully they will force Intel into "mobile Core2".

But looking at specs, they are expected. Apple had great core, now they have 1st class memory subsystem, 2nd to none. Cache sizes/latencies/bandwidth, memory bandwidth are awe inspiring. On top of that clock is very respectable 2Ghz+. End result is world class performance.

Sure some might argue that things are easier for Apple since they target ~2Ghz, not 4Ghz as Intel, but customers care about what they have in hand, not some esoteric gate delays.

Is there anywhere a comprehensive comparison of Twister (where does AT have that name from, actually) and SKL?

 

[coercitiv]

Twister (where does AT have that name from, actually)

I already gave you an answer on this matter in the other A9X thread, they get the names from LLVM commits.

They probably got the names the same way Swift and Cyclone were discovered.

With Swift, I had the luxury of Apple committing LLVM changes that not only gave me the code name but also confirmed the size of the machine (3-wide OoO core, 2 ALUs, 1 load/store unit). With Cyclone however, Apple held off on any public commits. Figuring out the codename and its architecture required a lot of digging. Last week, the same reader who pointed me at the Swift details let me know that Apple revealed Cyclone microarchitectural details in LLVM commits made a few days ago (thanks again R!).

 

[Nothingness]

Is there anywhere a comprehensive comparison of Twister (where does AT have that name from, actually) and SKL?

IIRC Andrei got that name from a crash screen.

 

[Nothingness]

I already gave you an answer on this matter in the other A9X thread, they get the names from LLVM commits.

Last time I looked (2 weeks ago) only Cyclone was mentioned in LLVM code.

 

[raghu78]

There is a poster here, raghu78, who basically was predicting that the A9X would force Intel to increase its core count in Y/U series processors.

It's pretty clear that Intel had the right idea with its low power processors: two cores that run as quickly as possible.

I am surprised that Apple went back to 2 cores for A9X. But that 2.2 Ghz clockspeed and resulting single thread performance is massive. Combined with 128 bit LPDDR4 for a massive 51.2 GB/s bandwidth and massive high speed L2/L3 caches this SoC is a freaking beast. Heck the single thread performance is equal to the 2x the single thread performance of A8 and > 2x the single thread performance of A7. Thats screaming performance. GPU performance is also double that of A9.

I am sure that Apple has SMT running on A10/A10X. Those are going to be single thread and multithread monsters. Apple could make the jump to 3C/6T at TSMC 10nm which brings 2.2x increase in density and 35-40% lower power at same performance. Right now I am sure Apple will be the company driving Intel the most to increase core count and improve performance. Apple have some of the best CPU design talent and loads of money through their cash cow iPhone. Apple rocks. :thumbsup:

 

[Andrei.]

There's a discrepancy against what AT says: the GB entry claims 2.16 GHz while Joshua says it's 2.26 GHz. My understanding is that GB on iOS reported frequency always is accurate.

GB frequency on A9 fluctuated for quite a while in the database.

We actually measure/estimate the frequency of the cores, I suppose GB has a similar method but it's just slightly off.

 

[ShintaiDK]

I am sure that Apple has SMT running on A10/A10X. Those are going to be single thread and multithread monsters. Apple could make the jump to 3C/6T at TSMC 10nm which brings 2.2x increase in density and 35-40% lower power at same performance. Right now I am sure Apple will be the company driving Intel the most to increase core count and improve performance. Apple have some of the best CPU design talent and loads of money through their cash cow iPhone. Apple rocks. :thumbsup:

A9/A9X is still far from Skylake. Plus they are bigger than Skylake 2+2 parts. This is only acceptable as long as Apple can charge "any" money for the parts.

A10 will be made on 16FF.

Even Apple says now in unquestionable actions that "moar cores" doesn't work.

 

[Abwx]

Geekbench is complete rubbish, they dont weight the scores, individual scores are added the one with the other and then they divide by the number of sub tests...

Here how it work, let s use two CPUs and two tests :

CPU A

Test 1 500 pts
Test 2 100pt pts

Score 600/2 = 300 pts


CPU B

Test 1 400 pts
Test 2 200 pts

Score 600/2 = 300 pts

Yet when tests are weighted the CPU B is 60% faster.

 

[witeken]

I am surprised that Apple went back to 2 cores for A9X. But that 2.2 Ghz clockspeed and resulting single thread performance is massive. Combined with 128 bit LPDDR4 for a massive 51.2 GB/s bandwidth and massive high speed L2/L3 caches this SoC is a freaking beast. Heck the single thread performance is equal to the 2x the single thread performance of A8 and > 2x the single thread performance of A7. Thats screaming performance. GPU performance is also double that of A9.

I am sure that Apple has SMT running on A10/A10X. Those are going to be single thread and multithread monsters. Apple could make the jump to 3C/6T at TSMC 10nm which brings 2.2x increase in density and 35-40% lower power at same performance. Right now I am sure Apple will be the company driving Intel the most to increase core count and improve performance. Apple have some of the best CPU design talent and loads of money through their cash cow iPhone. Apple rocks. :thumbsup:

TSMC 10nm brings a 0.52x improvement, according to themselves.

 

[Nothingness]

Geekbench is complete rubbish, they dont weight the scores, individual scores are added the one with the other and then they divide by the number of sub tests....

Come on, they don't average. Copy and paste into a spreadsheet and play with the numbers. Hint: use GEOMEAN...

 

[raghu78]

A9/A9X is still far from Skylake. Plus they are bigger than Skylake 2+2 parts. This is only acceptable as long as Apple can charge "any" money for the parts.

A10 will be made on 16FF.

Even Apple says now in unquestionable actions that "moar cores" doesn't work.

A9X is definitely right up there with Skylake in single thread and multi thread performance. We will see anandtech run SPEC 2006 on iPad Pro as it has 4 GB RAM. So we will have confirmation soon of Twister's IPC and A9X performance.

As for GPU performance A9X wrecks Surface Pro 4 running core i5 6300u (15w) in Manhattan Offscreen GPU performance. close to 50% faster.

http://www.anandtech.com/show/9727/the-microsoft-surface-pro-4-review-raising-the-bar/5
http://www.anandtech.com/show/9780/taking-notes-with-ipad-pro/2

Apple with A9X is embarassing Intel. Apple with their aggressive die sizes is telling Intel to up the game. My guess is A9X is closer to 150 sq mm. Sure Apple needs more die size than Intel but who is stopping Intel from increasing die size on their mobile chips. Right now its clear that Intel has gotten complacent due to lack of competition from AMD. Hopefully Intel gets the message from Apple and starts going for 140-150 sq mm mobile chips.

As for the process for A10/A10X its most likely 16FFC (non shrink version) which is expected to bring further process improvements for mobile. A10 is also expected to be manufactured using TSMC InFO (Integrated Fan Out) which is their Wafer level chip-scale packaging technology. InFO brings significant power and size reduction.

https://www.semiwiki.com/forum/cont...xt-bite-apple-chip-business.html?new_comment=

http://www.tsmc.com/uploadfile/ir/quarterly/2015/16cBq/E/TSMC 1Q15 transcript.pdf

"Now let me move to our InFO business update. The schedule to ramp up the InFO in second quarter next year remains unchanged. We expect InFO will contribute more than $100m quarterly revenue by next year, fourth quarter next year, when it will be fully ramped.

Right now we are building a new facility in Longtan, that's a city very near to Hsinchu, where our headquarters are, for ramping up InFO. Today a small product line is almost complete and it's ready for early engineering experiment. This pilot line will be expanded to accommodate the high-volume ramp in year 2016."

http://www.tsmc.com/uploadfile/ir/quarterly/2015/3C2bO/E/TSMC 3Q15 transcript.pdf

"Now let me update InFO. We have completed the construction of the new facility in Longtan and are ready for InFO's volume production. The manufacturing equipment move-in is on schedule and also we target volume ramp-up at second quarter next year.

Compared to existing package scheme, TSMC's InFO can bring greater-than-20% reduction in overall package thickness, 20% speed gain in performance and 10% better in thermal performance for power dissipation. So we are -- now InFO technology is capable and well positioned to enable next-generation mobile applications. So right now we continue to work
with major customers on completion of their product qualification. Meanwhile we are developing the next-generation InFO process, as Mark just said, for the future application.

Our expectation of InFO contributing more than $100m quarterly revenue by 4Q 2016 remain unchanged."

So Apple has a lot of TDP headroom to be gained next year. A10 / A01X will be huge improvement in multithread performance as I expect Apple to implement SMT next year in their custom ARM core. Add to that process and packaging improvements and we are going to see a monster SoC. 😎

 

[raghu78]

TSMC 10nm brings a 0.52x improvement, according to themselves.

its well known that TSMC has made clear statements on density improvements from 16FF+ to 10FF.

http://community.cadence.com/cadenc...0nm-is-ready-for-design-starts-at-this-moment

"Due to aggressive scaling, the 10nm FinFET (10FF) process node increases logic density by 2.1X compared to the TSMC 16nm FinFET Plus (16FF+) process node. Compared to 16FF+, the 10FF node can offer a 20% speed increase at the same power, or more than 40% power reduction at the same speed. TSMC has demonstrated a fully functional 256Mb SRAM in 10FF technology with die size scaling close to 50% of 16FF+. "

 

[StrangerGuy]

Okay, so here's Intel's R&D spending for 2014, although there's got to be a mistake somewhere since it was...

An amazing, astonishing, mind-blowing >20% of revenue.

http://fortune.com/2014/11/17/top-10-research-development/

Also, tell me if you find Apple somewhere 😀.

Yeah, you are proud of outspending Apple on R&D yet they still got the better mobile chip without owning a single fab.

Truly mind-blowingly lousy R&D ROI indeed 😀

 

[mikk]

Apparently the A9X GPU is Iris Pro 5200-class:

This is ~10% faster than HD 530 @1150 Mhz.

 

[dark zero]

RIP Core M Broadwell. Also, seems that Apple finally has potential to enter on Desktop tier. And that is WITHOUT HT and only on Dual Core.

Also it has the best GPU avaliable on Mobile. Only few GPU like nVidia Maxwell and AMD Radeon or Intel Iris Pro can curbstomp it.

I wonder when they will release a Macbook Pro with ARM A10X Quads... That would be epic.

 

[Abwx]

Come on, they don't average. Copy and paste into a spreadsheet and play with the numbers. Hint: use GEOMEAN...

A geometric mean change nothing since the bases are not weighted, the contribution of the score with the highest absolute value will keep being exagerated.

Using my previous exemple :

CPU A

Test 1 500 pts
Test 2 100pt pts

Score 600/2 = 300 pts


CPU B

Test 1 400 pts
Test 2 200 pts

Score 600/2 = 300 pts

Let s use the geometric mean :

CPU A score = sqrt(500^2 + 100^2) = 509.9

CPU B score = sqrt(400^2 + 200^2) = 447.2


That s even worse, the CPU B is now significantly slower, the GM just inflated the higher subscore contribution.

 

[Arachnotronic]

Apple with A9X is embarassing Intel. Apple with their aggressive die sizes is telling Intel to up the game. My guess is A9X is closer to 150 sq mm. Sure Apple needs more die size than Intel but who is stopping Intel from increasing die size on their mobile chips. Right now its clear that Intel has gotten complacent due to lack of competition from AMD. Hopefully Intel gets the message from Apple and starts going for 140-150 sq mm mobile chips.

I don't understand why the view is "OMG INTEL SUX" rather than "Wow, Apple did a damn fine job" with its CPU?

Intel is a merchant chip vendor that is fundamentally in a different business than what Apple is in. Apple makes a lot of money selling fully functional devices; SoC cost is an important but relatively small part of the bill of materials of an Apple product.

SoC cost is literally everything for Intel; it needs to make sure that the die sizes of its products are cheap enough to produce in order to support many customers offering products at many different price points.

Think of the A9X as a custom ASIC designed by Apple for Apple products (which is actually what it is). This gives Apple certain freedoms that a merchant chip vendor like Intel doesn't have.

Also remember that Apple releases precisely two SoCs per year -- Ax and AxX. Intel has to spend its efforts building chips/platforms for many different market segments, so it can't focus in the way that Apple can.

Apple has achieved something truly excellent here with the A9 and the A9X -- these are wonderful chips and are the product of first class engineering and a laser-like focus on a handful of great products. But to criticize Intel because Apple managed to deliver an SoC with competitive single-threaded performance doesn't seem fair to me.

That being said, I think Intel can maybe stop with its claims that Core m is "twice as fast" as any tablet out there because that claim is now very clearly and patently false.

 

[dahorns]

I don't understand why the view is "OMG INTEL SUX" rather than "Wow, Apple did a damn fine job" with its CPU?

Intel is a merchant chip vendor that is fundamentally in a different business than what Apple is in. Apple makes a lot of money selling fully functional devices; SoC cost is an important but relatively small part of the bill of materials of an Apple product.

SoC cost is literally everything for Intel; it needs to make sure that the die sizes of its products are cheap enough to produce in order to support many customers offering products at many different price points.

Think of the A9X as a custom ASIC designed by Apple for Apple products (which is actually what it is). This gives Apple certain freedoms that a merchant chip vendor like Intel doesn't have.

Also remember that Apple releases precisely two SoCs per year -- Ax and AxX. Intel has to spend its efforts building chips/platforms for many different market segments, so it can't focus in the way that Apple can.

Apple has achieved something truly excellent here with the A9 and the A9X -- these are wonderful chips and are the product of first class engineering and a laser-like focus on a handful of great products. But to criticize Intel because Apple managed to deliver an SoC with competitive single-threaded performance doesn't seem fair to me.

That being said, I think Intel can maybe stop with its claims that Core m is "twice as fast" as any tablet out there because that claim is now very clearly and patently false.

I think it is also worth while to point out that Apple can gear its chips to mobile workloads and ignore other workloads that may be more useful in things like servers. If Intel wants to remain competitive, its one Core to rule them all philosophy may need to change.