Broadwell-M specification

[Arachnotronic]

TSMC 20nm brings 15% higher performance at same power or 30% lower power for same performance wrt TSMC 28HPM. btw this comparison is with 28HPM which is the best foundry process and not Samsung 28nm gate first which is inferior in transistor performance.

http://www.eda.org/edps/edp2013/Papers/4-4 FINAL for Tom Quan.pdf

slide 19

16FF/28HPM 16FF/20SoC

Speed @ same total power 38% 20%
Total power saving @ same speed 54% 35%

16FF 20SOC 28HPM
Speed at same Power 1.38x 1.15x 1x
Power at same speed 0.46x 0.71x 1x

So a 1.4 ghz Cyclone will draw 30+% lower power (given Apple is moving from samsung 28nm which has lesser performance than TSMC 28HPM). Also at the same power Apple can clock the A7 at 1.6 Ghz at TSMC 20nm vs TSMC 28HPM and most likely 1.7 Ghz when you compare with Samsung 28nm. This is without any power efficiency improvements to the core and any power management improvement like finer grained power gating etc. Apple also can easily have better turbo speeds if they implement a quality DVFS (Dynamic voltage and frequency scaling) system like the one found on Intel Baytrail/AMD Mullins.

By clocking the 4 cores at 1 Ghz on TSMC 20nm Apple can run at lower voltage and easily run 4 Cyclone cores within the same power as 2 Cyclone cores at 28nm. At 1.4 Ghz itself Apple has cut power by more than 30% by process transition. Add power efficiency improvements and better DVFS and Apple could even do better.

As for the GPU you are talking about PowerVR which is the leader in mobile graphics and power efficiency. Rogue 6XT will be competitive against Kepler K1. Intel graphics against Rogue 6XT is a mismatch which Apple will easily win.

What makes you think Apple will go quad core? If anything, Apple is likely to improve the core (fixing some of the glass jaws, like what cripples A7's 3DMark Physics score) and stick to a dual core solution. After all, if Apple's goal is to eventually replace Intel in the MBA/rMBP, you would expect that they would want to build cores that don't sacrifice on ST performance relative to the Intel parts

 

[raghu78]

What makes you think Apple will go quad core? If anything, Apple is likely to improve the core (fixing some of the glass jaws, like what cripples A7's 3DMark Physics score) and stick to a dual core solution. After all, if Apple's goal is to eventually replace Intel in the MBA/rMBP, you would expect that they would want to build cores that don't sacrifice on ST performance relative to the Intel parts

Apple fit 2 Cyclone cores which perform like ivybridge core on a clock for clock basis into a 100 sq mm chip manufactured at Samsung 28nm. Given the full node density increase of 80 - 90% from TSMC 20nm Apple can easily double CPU core count and GPU cores while clocking them lower (base clocks) to fit in the 5W TDP. Also improvements to power management and a robust DVFS system will allow Apple to implement a good turbo implementation which allows higher single thread performance due to much higher clocks. Apple will try to improve single thread IPC but not at the cost of perf/watt. Right now A7 has no DVFS and still is a beast. Imagine a A8 chip which can use the entire available TDP when running workloads which don't tax the GPU. for most productivity apps the GPU is lightly loaded and a good DVFS like Intel Baytrail/AMD Mullins means the CPU can easily hit 2 Ghz turbo. Apple already uses it very high IPC to finish the task asap and race to idle. A robust DVFS implementation is natural progression. Apple already have a massive core and single thread IPC. Now its about improving efficiency and tweaking the core while doubling core counts.

Apple will get to replacing Intel in Macbook Air but only when the A9 ships and they have FINFET which brings a big power efficiency boost. For the Macbook Pro it depends on how ambitious they are. Can the successors of Cyclone scale to 3.5 - 4Ghz. Apple has not proven that they can build a high frequency design which competes with Intel. They also need the software infrastructure ready for the transition to ARMv8 which is why it might not be before 2016 at the earliest.

 

[witeken]

You seriously overestimate Apple's capability to compete with Intel. Apple will have a 4GHz chip? It will have FinFETs? Great, Intel's will have been shipping Tri-Gates for 4 years and will have Germanium and III-V.

Qualcomm is the leader in market share in mobile. They probably have quite a large development team too, working on the successor of Krait. And yet they will fail in maintaining their market share and Android CPU lead, which they already lost, for the same reasons Apple will fail against Intel.

 

[jdubs03]

Intel graphics against Rogue 6XT is a mismatch which Apple will easily win.

You're drastically underestimating the iGPU perf of Cherry Trail. Easily win? Come on try to be a little bit unbiased. I said iGPU perf will be best in class because of 16 Gen8 EUs. 4x the amount of EUs and a significant generational increase graphically. I call it how it is not I want it to be. I have no doubt that Rogue will be a good GPU, but I seriously doubt it can challenge a chip with Haswell # of EUs and Broadwell's GPU arch.

Intel's will have been shipping Tri-Gates for 4 years and will have Germanium and III-V.

That's a big maybe there.

 

[ShintaiDK]

Apples biggest problem is that it keeps losing marketshare. Samsung is the masters of smartphones today.

Also the A7 picked a lot of low hanging fruits (Specially for benchmarking). Dont expect the A8 to do the same.

 

[raghu78]

You seriously overestimate Apple's capability to compete with Intel. Apple will have a 4GHz chip? It will have FinFETs? Great, Intel's will have been shipping Tri-Gates for 4 years and will have Germanium and III-V.

I am estimating them correctly while you are doing the underestimating. You need to just look at the last 3 Apple SOCs to see how aggressive they are. Apple beat the ARM ecosystem to 64 bit by more than 12 months. The single thread performance increase is far ahead of anything that Intel or Qualcomm offers for the mobile market.The very fact that people talk about comparing Intel's big cores to Cyclone is a testament to Apple engineering.

http://www.anandtech.com/show/7910/apples-cyclone-microarchitecture-detailed

btw they did this on Samsung 28nm which is not even as good as TSMC 28HPM and much inferior to Intel 22nm FINFET in transistor performance. by Q4 2015 Apple A9 will be shipping in iPhone 6s / ipad air 3. Apple is likely to multi source the manufacturing from both Samsung 14 LPE and TSMC 16FF+. So yeah its possible for Apple to produce a high frequency / high performance core. I would not bet against them.

Qualcomm is the leader in market share in mobile. They probably have quite a large development team too, working on the successor of Krait. And yet they will fail in maintaining their market share and Android CPU lead, which they already lost, for the same reasons Apple will fail against Intel.

Intel's Baytrail is an abject failure in phones and if not for contra revenue nobody would buy it even for tablets. So sorry if I am not too enthusiastic about Intel against Apple/Qualcomm especially when the process gap is reducing and Intel products in late 2015 and 2016 will compete against FINFET products from Apple, Qualcomm and the rest of the ARM players like Nvidia.

 

[mavere]

I'm optimistic about Logan-Denver, I think that they may have a chance even if Intel has years of experience on them. It will have higher single-thread perf than Airmont assuredly, multi-thread I doubt it. The A8 could be quite close to maybe a bit more than Logan, only because of the switch from 28nm to 20nm. uArch wise Tegra should win out this time around (against the A8).

Yea, because of Nvidia's incessant ability to churn out sophisticated CPU designs.
...
...
...
wait, what?

(I've no idea where your optimism comes from)

 

[Roland00Address]

Have a look at ATOM Z3770 (4W TDP) vs Core i5 4202Y (11,5W TDP). The ATOM is close to 70% the performance of the 11,5W TDP Core i5.

Im expecting the new 14nm ATOM to be very close to 80% or more(CPU) than 14nm Broadwell-Y at the same TDP.

Also, Haswell 11.5W TDP SKUs are priced at $280, ATOM is priced at $35. That is 8x times the price for ~30% more performance, ouch

We really do not have enough numbers to truly guess haswell y performance vs atom performance for there are barely any haswell y chips on the market. (Let alone extrapolate that info to broadwell y) For example there is a i5 4302y as well as a i5 4202y with 15% higher turbo but nobody seems to use it besides the panasonic tablet in a 7" case. Is this because the chip can't sustain that turbo long enough and thus no one wants to pay for the premium, the price is too high, intel can't make enough chips, or the oems just don't want y fanless skus just yet. Another 15% is nothing to slouch at for that makes your performance difference 50% higher.

You also picked the currently best intel atom chip (though the z3795 is coming out soon). Very few oems are using the z3770 vs the z3740 why is that? There is a difference of 30% in clock speed for turbo between the z3740 and the z3770. That does not mean there will be 30% higher total performance, but in single threaded things if turbo works correctly and they can sustain the turbo long enough you should see a difference of about 25 to 30% between the two.

Furthermore as you have pointed out it is not just the multithreaded difference but also the single threaded difference that matters.

-----

All I am doing is recommending patience and waiting for more info before making firm conclusions.

 

[tential]

I'd love a $600 Surface Pro 4.

Yup. I'm not complaining if devices start at $600. I'd take a Surface Pro like produce that starts well under $1000. That's great compared to what we have now.

 

[tential]

Have a look at ATOM Z3770 (4W TDP) vs Core i5 4202Y (11,5W TDP). The ATOM is close to 70% the performance of the 11,5W TDP Core i5.

Im expecting the new 14nm ATOM to be very close to 80% or more(CPU) than 14nm Broadwell-Y at the same TDP.

Also, Haswell 11.5W TDP SKUs are priced at $280, ATOM is priced at $35. That is 8x times the price for ~30% more performance, ouch

I agree with you and sadly, we never see the designs we really want to see. I thought the Z3770 would be in a lot of different notebooks by now as it's a GREAT processor, yet we normally see i3/i5s or something else instead.

Atom isn't used in nearly as many different types of devices as it could be. I usually only see the lower specced Z series and only in tablets.

The Surface Pro for example could have used the Z3770 and easily released a great performing cheaper model.

 

[jdubs03]

Yea, because of Nvidia's incessant ability to churn out sophisticated CPU designs.
...
...
...
wait, what?

(I've no idea where your optimism comes from)

Companies can adapt and produce higher quality products. Past performance is not an indicator or future results.

Nvidia has been working on this since 2011-2012, and has already shown it running on a device.

Logan 64-bit is likely to be in the Nexus 9, which will be the launch device for Android L. The 32-bit quad-core version has already produced some venerable benchmarks.

 

[ninaholic37]

I agree with you and sadly, we never see the designs we really want to see. I thought the Z3770 would be in a lot of different notebooks by now as it's a GREAT processor, yet we normally see i3/i5s or something else instead.

Atom isn't used in nearly as many different types of devices as it could be. I usually only see the lower specced Z series and only in tablets.

I see Bay Trail-M everywhere in laptops now (labeled as Celeron N28xx/N29xx and Pentium N35xx).

 

[SPBHM]

Have a look at ATOM Z3770 (4W TDP) vs Core i5 4202Y (11,5W TDP). The ATOM is close to 70% the performance of the 11,5W TDP Core i5.

from the link you posted for example
CB10 ST 32bit
2900 vs 1200 from average

and looking at the IGP
3dm cloudstorm
3700 vs 1200 for highest

cb 11.5 OGL
14 vs 4 for highest


so the IGPs are not even worth comparing, and the ST performance difference is huge, which will probably reflect on basic usage experience, considering how bad the Atom is for ST.

 

[xpea]

iGPU perf should beat everything else on the market. It's CPU perf that may not (Logan 64bit).

well I think the opposite. Kepler mobile GPU will have no equivalent this year from intel in the sub 10W TDP. but on the CPU side, I'm not sure Logan 64bit will compete...

 

[jdubs03]

well I think the opposite. Kepler mobile GPU will have no equivalent this year from intel in the sub 10W TDP. but on the CPU side, I'm not sure Logan 64bit will compete...

16 Broadwell EUs depending on clocks will outpace Haswell HD4600, surpassing TK1. Logan is dual-core, that is going to smash Atom. In terms pf CPU perf. Logan could be competitive with Broadwell-Y. Time will tell.

 

[xpea]

16 Broadwell EUs depending on clocks will outpace Haswell HD4600, surpassing TK1. Logan is dual-core, that is going to smash Atom. In terms pf CPU perf. Logan could be competitive with Broadwell-Y. Time will tell.

but what will be the tdp of 16 EUs outpacing HD4600 ? It means that intel could beat nvidia GPU experience in one generation, when last gen they were crunched by a factor of 2 or 3 ? comon be serious... no way intel can show such jump in performance/efficiency in such short time...

 

[witeken]

but what will be the tdp of 16 EUs outpacing HD4600 ? It means that intel could beat nvidia GPU experience in one generation, when last gen they were crunched by a factor of 2 or 3 ? comon be serious... no way intel can show such jump in performance/efficiency in such short time...

-4x as many EUs
-new process node
-new much improved architecture

 

[jdubs03]

but what will be the tdp of 16 EUs outpacing HD4600 ? It means that intel could beat nvidia GPU experience in one generation, when last gen they were crunched by a factor of 2 or 3 ? comon be serious... no way intel can show such jump in performance/efficiency in such short time...

the gpu arch change is very significant, you can't put bay trail in the same boat as cherry trail, its nowhere even close; I don't think bay trail even has gen7, cherry trail will have gen8, which is much improved from haswell (gen7.5). the switch to 14nm freed up the space to add all the extra EUs with some wiggle-room for energy savings.