[WSJ] TSMC is shipping 20nm SOCs to Apple

[witeken]

Are you guys sure what is coming out is not the A7x but the A8?

Why can't it be the A7x with;
Quad-core Cyclone?
GX6650?

We only know it's a 20nm node SoC.

 

[NTMBK]

Because it will be substantially worse than competing SoCs (Core M).

Are you really certain of that? Cyclone is already a Haswell-class CPU design, and Broadwell doesn't change the CPU design in any significant way. They're both getting a node shrink. And Apple has a vertical stack from hardware to software, so they can do a lot of optimization work to squeeze the most out of their SoC. Will that be enough to overcome the slight process advantage Intel has? I don't know! Should be an interesting fight to watch

 

[NostaSeronx]

Hoping for A7x with Quad-core Cyclone CPU-side and "Rogue XT" class GPU-side, this year.

Hoping for A8 with Quad-core Cyclone2 CPU-side and "Wizard" class GPU-side, next year.

 

[ShintaiDK]

You are wastly overating the cyclone performance on a few benchmarks that is largely cache or specific instruction dependent.

 

[NTMBK]

I'm wondering whether Apple will bring out their own GPU this time around. Its the only piece they're missing for a "fully in house" SoC.

 

[witeken]

Are you really certain of that? Cyclone is already a Haswell-class CPU design, and Broadwell doesn't change the CPU design in any significant way. They're both getting a node shrink. And Apple has a vertical stack from hardware to software, so they can do a lot of optimization work to squeeze the most out of their SoC. Will that be enough to overcome the slight process advantage Intel has? I don't know! Should be an interesting fight to watch

CPU design doesn't matter. Haswell/Broadwell is designed to scale up to 140W. There is a ton of headroom. Broadwell-Y has already a better IPC and goes up to 2.6GHz while Cyclone tops out at half of that. But that doesn't matter in the end because efficiency is key here. 20nm planar simply can't compete in any way with 14nm Tri-Gate.

You probably have a point with software optimization; iOS isn't yet optimized for Core. They'd better do that if they don't want to get behind.

 

[NostaSeronx]

20nm planar simply can't compete in any way with 14nm Tri-Gate.

That is up to the designer not the foundry.

 

[witeken]

That is up to the designer not the foundry.

Cyclone must be substantially more efficient than the Haswell architecture to be faster in this thermally constraint environment with 20nm vs 14nm. I'd say this is basically impossible.

 

[itsmydamnation]

You probably have a point with software optimization; iOS isn't yet optimized for Core. They'd better do that if they don't want to get behind.

you realise how much of a crack pipe shill you sound like right? ... intel has sold how many ~600 -700 USD phones? apple has sold what 400 million 600 -700 USD iphones in total.......

every generation we hear the same thing....

just like how HD4000 is "good enough" A7 ( yet alone everything above it) is "good enough"

Personal attacks will not be tolerated
-ViRGE

 

[NostaSeronx]

Other than Intel's 256-bit vector instruction set. Cyclone is much more efficient and much more faster than Haswell.

The better case study is Apple versus AMD.

Apple A7 - 2W TDP
http://browser.primatelabs.com/geekbench3/585335

AMD Puma HP - 15W TDP
http://browser.primatelabs.com/geekbench3/632049

AMD Puma ULP - 4.5W TDP
http://browser.primatelabs.com/geekbench3/569477

All three are on Common Platform 28-nm nodes. It really shows the power of the designer.

 

[NTMBK]

CPU design doesn't matter. Haswell/Broadwell is designed to scale up to 140W. There is a ton of headroom. Broadwell-Y has already a better IPC and goes up to 2.6GHz while Cyclone tops out at half of that. But that doesn't matter in the end because efficiency is key here. 20nm planar simply can't compete in any way with 14nm Tri-Gate.

You probably have a point with software optimization; iOS isn't yet optimized for Core. They'd better do that if they don't want to get behind.

The way that Broadwell and Haswell scale to 140W is by putting extremely high numbers of cores on a die. Haswell EP goes up to 18 cores in one socket!

And CPU design makes a huge difference to efficiency. Just look at Pentium D vs Core 2 Duo on the same process node. (Not that I am saying that Haswell is a power hog, just wanted to point that out.)

Anyway, as I said, we shall see. Apple's SoCs will benefit from a more optimised software stack and a more efficient instruction set. And we have absolutely no idea how costs and yields will compare. Going to be fun to watch.

 

[coercitiv]

Other than Intel's 256-bit vector instruction set. Cyclone is much more efficient and much more faster than Haswell.

The better case study is Apple versus AMD.

I'd like to see the worse case study, can we see where is Cyclone both much faster and more efficient than Haswell?

 

[NostaSeronx]

I'd like to see the worse case study, can we see where is Cyclone both much faster and more efficient than Haswell?

Found Intel's comparable setup to the Apple A7.

i3-4010Y / 11.5W TDP / 6W SDP
http://browser.primatelabs.com/geekbench3/401702
(heh, heh, model 69)

Apple A7 / 2W TDP / <1W SDP
http://browser.primatelabs.com/geekbench3/585335

So, nevermind on AVX2. It does not help Haswell.

(Just noticed the SHA1 results, someone fix Geekbench)
http://browser.primatelabs.com/geekbench3/compare/401702?baseline=585335

Well if your mining for coins those SHA1/SHA2 results sell those A7s.

 

[mikk]

Geekbench is nonsense.

 

[NostaSeronx]

Geekbench is nonsense.

Geekbench is actually pretty good as long as you're looking at 3.1 scores.

 

[blackened23]

Comparing A7 or A8 to android or windows products is 100% meaningless. Apple competes on a device level , not a SOC level. When Apple sells their A7 or A8 chips to anything than iOS use, it might matter, but now it doesn't.

It is literally a waste of time to even talk about benchmarks of a product that doesn't compete as an android SOC. Benchmarking in iOS is different and more optimized than benchmarking in a general OS (android). It's like comparing benchmarks from Windows to Mac OS - they're not the same no matter whether the benchmark shares the same name or what have you. Apple does not sell their SOCs, they sell devices. They compete with other devices. They do not directly compete with other SOCs, rather (AGAIN) they compete on a device (higher) level.

Now I personally think Apple products are fantastic, but comparing A7/A8 to other SOCs? Waste of time, apple is not competing as a SOC Vendor, period.

 

[NostaSeronx]

Apple's A7 is a desktop class SoC. So, comparing it to a desktop or laptop chip is okay.

The optimization between iOS, Windows, Android, *Unix. Does not matter in benchmarks as the whole AMD vs Intel dilemma has proved.

 

[witeken]

Now I personally think Apple products are fantastic, but comparing A7/A8 to other SOCs? Waste of time, apple is not competing as a SOC Vendor, period.

Not when you're on a site like AnandTech .

Apple's A7 is a desktop class SoC. So, comparing it to a desktop or laptop chip is okay.

We'll see if they use it in the next Mac Pro.

 

[Lepton87]

Assuming 10mm to 10mm die (100mm^2) then a 300mm wafer will have ~600 dies.
Lets say they have 60% yields, then that makes 360 dies per wafer.

Lets say they need 5M 20nm SoCs, then they will need ~13888 300mm wafers.
For 10M SoCs they will need ~27777 wafers and so on.

that's wrong math, using your math there would be no penalty to having a big die instead of a small one. Just use a hypothetical 650mm2 chip and do the math. 92* dice per wafer with perfect yields and with 60% yields you would get 55 functional dice, that's just wrong.

that's assuming your math is correct for the number of 100mm2 dice in a 300mm wafer.

 

[Arachnotronic]

CPU design doesn't matter. Haswell/Broadwell is designed to scale up to 140W. There is a ton of headroom. Broadwell-Y has already a better IPC and goes up to 2.6GHz while Cyclone tops out at half of that. But that doesn't matter in the end because efficiency is key here. 20nm planar simply can't compete in any way with 14nm Tri-Gate.

You probably have a point with software optimization; iOS isn't yet optimized for Core. They'd better do that if they don't want to get behind.

Transistor to transistor, 14nm should be quite a bit ahead of 20nm. However, I would be very careful in assuming that the underlying architecture doesn't matter.

It does matter, and you need only look at Gen. 7 GPU v.s. Power VR Series 6 to see how the world's best transistors don't mean much if your architecture is sub-par.

 

[BarkingGhostar]

I don't get this. I thought "CPUs and Overclocking" was about people buying processors for computers they made, and maybe for those that OC them. I see no one in the wild being able to buy one of these CPUs for their computer, let alone OC.

Should move to the "All Things Apple" forum.

 

[Lepton87]

Apple's A7 is a desktop class SoC. So, comparing it to a desktop or laptop chip is okay.

The optimization between iOS, Windows, Android, *Unix. Does not matter in benchmarks as the whole AMD vs Intel dilemma has proved.

desktop class? You're swimming in apple's kool-aid I guess simply drinking too much of it is not enough. Desktop? Really? They would need to triple the frequency or double the frequency and the number of cores for that to be true but only for CPU part. I don't think its GPU part compares favorably to crystalwell or 512 GCN cores.

 

[Qwertilot]

I think that's normally taken to mean that they built a CPU core which could have that done to it in principle, then clocked it down quite low etc to fit into a mobile power space. That's a rather different approach to a lot of the mobile SoC's.

Seems to have worked quite nicely, although you imagine part of the reason they did it that way might have been at least being able to threaten to move their macs over.

 

[dahorns]

I think that's normally taken to mean that they built a CPU core which could have that done to it in principle, then clocked it down quite low etc to fit into a mobile power space. That's a rather different approach to a lot of the mobile SoC's.

Seems to have worked quite nicely, although you imagine part of the reason they did it that way might have been at least being able to threaten to move their macs over.

That remains to be seen. I don't know that we've seen any evidence that the A7 is effective, stable, or efficient at higher clock speeds. Hopefully we learn more with the A8.

 

[itsmydamnation]

Not when you're on a site like AnandTech .

i think you broke your quote, also when i say A7 i dont mean apple A7 i mean cortex A7