Apple A8x

[witeken]

A8X single threaded performance is getting stupidly fast. With a ~4W chip at 1.5 GHz they are hitting 70% of the performance of a 2.6 GHz 15W i5? Put that in a Mac already!

Your 2.6GHz number is the boost clock speed. The base speed is closer to 2GHz, so in short workloads, the chip will be a lot slower than your 1.4x number. And because it's made on a planar process, it will consume a lot more.

Broadwell and Skylake will simply widen the gap.

BTW, comparing TDP of a chip to compare performance per watt is ludicrous. People seem to accept it blindly, but when AMD did the same (performance/TDP) with its Puma+ chip, that was suddenly not fair, interestingly enough.

 

[Enigmoid]

Is it the CPU Physics score that is still holding back A8X in 3dmark Ice Storm Unlimited overall score? The third CPU core should have significantly improved the physics score. The Graphics score should be close to mobile Kepler, but based on a combined score of 21,659, it appears that the Graphics score is less than 30,000...

Apple's memory controller sucks at random memory performance.

http://www.futuremark.com/pressrele...results-from-the-apple-iphone-5s-and-ipad-air

 

[asendra]

Do we know anything wether Anandtech got a review unit when they were at the event? I suppose so, but not sure, and they haven't posted a review yet.

 

[mavere]

Apple's memory controller sucks at random memory performance.

http://www.futuremark.com/pressrele...results-from-the-apple-iphone-5s-and-ipad-air

Are there game developers or general software/benchmark developers who have struggled with that supposed issue?

Futuremark's response sounded reasonable then, but another year has passed with an entire ecosystem of developers coding for the A7 (and now A8). If they did so with few issues, when can we call out Futuremark for depending on bad code?

Edit: Apple should definitely improve their memory controllers, but a benchmark called 3DMark should also represent the capabilities and direction of game engines and should not be held back by opaque, special-snowflake bottlenecks affecting nobody else.

 

[Eug]

http://www.apple.club.tw/ipad-air-2%E7%A2%BA%E5%AF%A6%E6%98%AFa8x-2gb-ram/

So this would be the A8X to be (presumable) released this Thursday.

There are two ELPIDA dram chips next to the A8x and with some
imagination we might read the (very blurry) part number right below
the ELPIDA logo as "FA164A..." which would make them 16Gb LPDDR3
parts each with a 2x32 bit databus.

That would make a decent 4GB dram in total with a wide 128 bit bus.
If I'm reading the "GD-F" on the 3rd line correctly then that should make
them DDR3-1600 parts giving the A8X a desktop like bandwidth of
25.6 GB/s.

Well, I guess Thursday, 16th we'll find out the details with the introduction
of the new iPads.

iFixit's teardown:

https://www.ifixit.com/Teardown/iPad+Air+2+Teardown/30592

We misread the memory. It is not FA164A... (16 Gb). It is F8164A3MD (8 Gb) x 2 chips. Now the leaked image makes more sense. There are indeed two chips, but both are 1 GB.

Apple APL1012 A8X 64-bit Processor
Elpida/Micron Technology F8164A3MD 8 Gb (1 GB) RAM (1 GB x 2 = 2 GB total)
SK Hynix H2JTDG8UD1BMR 128 Gb (16 GB) NAND Flash
NXP 65V10 NFC Module (as found in the iPhone 6 and 6 Plus)

It has NFC???

 

[jpiniero]

It does have NFC but not the antenna apparently. Not sure why Apple included it. I'd love to see people use Apple Pay with an iPad, that would be hilarious.

 

[Eug]

Apparently it is needed for Apple Pay with an iPad, but without the NFC functionality part of Apple Pay.

 

[deathBOB]

Apparently it is needed for Apple Pay with an iPad, but without the NFC functionality part of Apple Pay.

Secure Element?

 

[III-V]

Are there game developers or general software/benchmark developers who have struggled with that supposed issue?

Futuremark's response sounded reasonable then, but another year has passed with an entire ecosystem of developers coding for the A7 (and now A8). If they did so with few issues, when can we call out Futuremark for depending on bad code?

Edit: Apple should definitely improve their memory controllers, but a benchmark called 3DMark should also represent the capabilities and direction of game engines and should not be held back by opaque, special-snowflake bottlenecks affecting nobody else.

That's not what the situation is. They have a specific test that sees no improvement, whereas the other tests mentioned do fine.

Also, to call it an issue is odd. Simply because one performance metric doesn't improve does not mean it's an issue.

Frankly, you're also being incredibly rude.

 

[mavere]

TIL questioning whether a benchmark's scaling is representative of real gaming development is "incredibly rude".

 

[III-V]

TIL questioning whether a benchmark's scaling is representative of real gaming development is "incredibly rude".

It's not, but being purposefully illiterate is.


Personal attacks are not allowed here
Markfw900

 

[xpea]

Great job, Apple's silicon team(s)!

really ?
it takes them 3 billion transistors on 20nm process, with 128-bit RAM bus, TBDR arch and fp16 precision to match TK1 in GFXbench (~1b tranies on 28nm, 64-bit bus and fp32)
sorry, but color me unimpressed... or in other words, NV did an incredible job with TK1 and I can't wait to see what TM1 Erista will do...

 

[ancientarcher]

really ?
it takes them 3 billion transistors on 20nm process, with 128-bit RAM bus, TBDR arch and fp16 precision to match TK1 in GFXbench (~1b tranies on 28nm, 64-bit bus and fp32)
sorry, but color me unimpressed... or in other words, NV did an incredible job with TK1 and I can't wait to see what TM1 Erista will do...

Next year in Sep/Oct we will probably get to see both Apple A9/9X and Nvidia Erista, with the Apple chip on Samsung 14FF (from what I have heard) and the Nvidia chip on TSMC 16FF. Intel will also be on 14nm FF. With all of the major players (including Qualcomm hopefully) on 14nm FF we will get to see their design chops. An even playing field in manufacturing for the first time ever...

 

[xpea]

Next year in Sep/Oct we will probably get to see both Apple A9/9X and Nvidia Erista, with the Apple chip on Samsung 14FF (from what I have heard) and the Nvidia chip on TSMC 16FF. Intel will also be on 14nm FF. With all of the major players (including Qualcomm hopefully) on 14nm FF we will get to see their design chops. An even playing field in manufacturing for the first time ever...

Erista has been taped out last July on 20nm. It will be launched at CES (January) or MWC (February) for an availability in first devices around July 2015 (same schedule as TK1).

16nm Tegra (Parker) part will be in devices in H2 2016, not before

No idea about Apple A9/9X schedule

 

[Eug]

really ?
it takes them 3 billion transistors on 20nm process, with 128-bit RAM bus, TBDR arch and fp16 precision to match TK1 in GFXbench (~1b tranies on 28nm, 64-bit bus and fp32)
sorry, but color me unimpressed... or in other words, NV did an incredible job with TK1 and I can't wait to see what TM1 Erista will do...

I'm curious which is more power efficient.

Also, I'm still wondering if A8X is sort of overkill for the 2048x1536 iPad. I was fully expecting 2 GB RAM, and perhaps an upgraded GPU, but definitely not the third core, esp. at >3 billion transistors and 20 nm.

I wonder if this is destined to be a 14 nm part for a 12" MaxiPad or even an ARM Retina MacBook. Furthermore I wonder if the smaller battery and thinner design of the iPad Air 2 is built more with the 14 nm part in mind as opposed to the 20 nm part.

P.S. Don't forget that iOS 8.1 includes the option for 3X screen elements for iPads. No such iPads currently exist. All are 1X or 2X.

 

[sontin]

TSMC's 20nm process gives a 30% power saving advantage over their 28nm process.

However, Tegra K1 with Denver is the only SoC which comes close or even beat Apple's A8X. And power consumption is more a result of the display and not primarily of the SoC, so i think most people dont care about runtime under 3D load.

 

[Mopetar]

I'm curious which is more power efficient.

Also, I'm still wondering if A8X is sort of overkill for the 2048x1536 iPad. I was fully expecting 2 GB RAM, and perhaps an upgraded GPU, but definitely not the third core, esp. at >3 billion transistors and 20 nm.

I think Apple is really trying to push full-blown productivity apps on tablets. The GPU doesn't need to be just for prettier graphics or games because with their Metal API it can also be used for compute.

It may also be because Apple wanted to ship iOS 8 with some functionality to run multiple apps side-by-side and have the processing power such that they don't have to share resources and have performance suffer. However, the software probably wasn't ready to ship, but by that time they had already committed to the hardware which couldn't be changed.

 

[III-V]

really ?
it takes them 3 billion transistors on 20nm process, with 128-bit RAM bus, TBDR arch and fp16 precision to match TK1 in GFXbench (~1b tranies on 28nm, 64-bit bus and fp32)
sorry, but color me unimpressed... or in other words, NV did an incredible job with TK1 and I can't wait to see what TM1 Erista will do...

I highly doubt he was talking about the GPU, which isn't even Apple's work anyway. But really, the last thing I'd be with TK1 is "impressed." Relieved is a bit more like it. It took them this long to treat mobile as a first class citizen?

There's also the question of power efficiency, as mentioned above.

 

[ams23]

Bringing a fully featured flagship desktop/notebook Kepler GPU architecture (with associated software stack) down to the ultra mobile space is no easy task (and FWIW, TK1 32-bit variant has already been on the market for several months now). It would be akin to Intel bringing their Core architecture to the ultra mobile space.

 

[witeken]

Next year in Sep/Oct we will probably get to see both Apple A9/9X and Nvidia Erista, with the Apple chip on Samsung 14FF (from what I have heard) and the Nvidia chip on TSMC 16FF. Intel will also be on 14nm FF. With all of the major players (including Qualcomm hopefully) on 14nm FF we will get to see their design chops. An even playing field in manufacturing for the first time ever...

The name being the same doesn't mean anything. Samsung will barely have its first FinFET transistors while Intel has their second generation.

Intel has a 3.5 year lead.

 

[jdubs03]

The name being the same doesn't mean anything. Samsung will barely have its first FinFET transistors while Intel has their second generation.

Intel has a 3.5 year lead.

I don't think their lead is that high anymore. Their delays have certainly took at least half a year chunk out of whatever the lead was. Realistically I think Intel's lead is on the order of 2-2.5 years when considering that TSMC's 16FF and Samsung 14LPE/P have higher densities than 14nm.

However, when not considering that, Samsung will be about 9 months to a year behind Intel when it comes to having a product on the shelf.

@xpea

In regards to Nvidia Parker, I suspect you're right, though it would be great if they got a nice jump going from 28nm to 16FF for Erista next year, and then 2016 having 16FF+, but unfortunately it'll likely be 20nm Erista.

 

[witeken]

I don't think their lead is that high anymore.

You don't have to take my word.

Ivy Bride: April 2012.
16FF: October 2015?

 

[dahorns]

Realistically I think Intel's lead is on the order of 2-2.5 years when considering that TSMC's 16FF and Samsung 14LPE/P have higher densities than 14nm.

I thought one of Intel's selling points was that their 14nm would (really for the first time in a while) have better density than the comparable node for the foundries.

Anyway, this is off topic.

It'll be interesting to see how often that third core really gets used (and how). Until now IOS devices were limited to 2 cores. Is it something that existing software could even use? Or will it primarily be used for multitasking?

 

[Hans de Vries]

iFixit's teardown:

https://www.ifixit.com/Teardown/iPad+Air+2+Teardown/30592

We misread the memory. It is not FA164A... (16 Gb). It is F8164A3MD (8 Gb) x 2 chips. Now the leaked image makes more sense. There are indeed two chips, but both are 1 GB.

Apple APL1012 A8X 64-bit Processor
Elpida/Micron Technology F8164A3MD 8 Gb (1 GB) RAM (1 GB x 2 = 2 GB total)
SK Hynix H2JTDG8UD1BMR 128 Gb (16 GB) NAND Flash
NXP 65V10 NFC Module (as found in the iPhone 6 and 6 Plus)

Nice work from iFixit. Now hoping for a die shot from ChipWorks

 

[Eug]

I got my 64 GB Gold WiFi, iPad5,3
Geekbench says 1.51 GHz, 1.94 GB RAM.

Ran all these benchmarks twice

Kraken 1.1
4423.2 ms
4538.6 ms

Sunspider 1.0.2
329.2 ms
312.3 ms

Octane 2.0
6672 (did not restart browser as recommended)
8349 (restarted browser, as recommended)

Geekbench 3.2.2
1797/4496
1811/4482

Basemark X 1.1.1 Medium (default)
41603
41454

GFXBench 3.0.20
Didn't seem to work.
Went through all the tests and the tests seem to run OK, but then the test results were empty.

Anything else to run?