QQ Snapdragon 810 overheating issues: delay?

[geoxile]

oobydoobydoo

It's not that easy. Even for a "stock" A15/A57, the quality of the physical implementation of the core will vary depending on the design team. A company like Qualcomm or Samsung will be able to implement the "stock" core in a much better way (higher frequencies, lower power) than a company with minimal chip design experience like LG.

That's why the LG NUCLUN clocks at just 1.5GHz for the A15; other companies have done better.

And who is to say that the problem with S810 is the CPU? That's a ~2B+ transistor SoC we're talking about and the interactions among the IP blocks within such a chip, and the fact that you have a bunch of different IPs all running at different frequencies, adds a LOT of complexity.

The CPU core is just a small part of a mobile SoC, which is why I find it perplexing that everyone cares so much about whether QCOM or whomever does a custom core or uses an ARM core...there's so much more than that to an SoC...

No, clearly the problem is the name.

NUCLUN? Dumb name.
Exynos? Cool name, it has a X in it.
Snapdragon, also a cool name but 810 just doesn't sound good anymore. They should've called it something like the Snapdragon x8100

 

[Arachnotronic]

I'll bet Intel are kicking themselves that they don't have a suitable product that would allow them to swoop in here and take advantage.

No kidding. Intel should have done another family of products on 22nm instead of betting the farm on 14nm mobile products in this timeframe.

"Goldmont" implemented in 22nm + ImgTec Series 6XT + improved ISP + H.265 encode/decode would have been a very potent combination in 2015.

 

[Nothingness]

I'll chip in if you agree to do the actual work configuring the hardware, setting up the benchmark and writing the article

SPECCPU is not that hard to get running... As long as you have PERL and a compiler. The issue would be that Intel fanboys would dismiss the results if Intel C compiler isn't used, and I would personnally dismiss any result that doesn't use the same compiler for all the platforms (even if it could be argued ARM would still be at a disadvantage due to much longer support of x86 on gcc). Also you would need access to all platforms because input data can't be redistributed (IIRC pre compiled benchmarks can be distributed).

 

[Nothingness]

No kidding. Intel should have done another family of products on 22nm instead of betting the farm on 14nm mobile products in this timeframe.

"Goldmont" implemented in 22nm + ImgTec Series 6XT + improved ISP + H.265 encode/decode would have been a very potent combination in 2015.

Hmm I thought Goldmont CPU improvements were mostly coming from higher freq (2.7GHz) which was possible only thanks to 14nm.

 

[Exophase]

SPECCPU is not that hard to get running... As long as you have PERL and a compiler. The issue would be that Intel fanboys would dismiss the results if Intel C compiler isn't used, and I would personnally dismiss any result that doesn't use the same compiler for all the platforms (even if it could be argued ARM would still be at a disadvantage due to much longer support of x86 on gcc). Also you would need access to all platforms because input data can't be redistributed (IIRC pre compiled benchmarks can be distributed).

I'm very interested in seeing this happen, I just don't have time. Or the platforms really. I assume all the benches purely driven by file I/O? Would there be any difficulty getting it running on Android? Trying to get a proper Linux distro running on all these platforms would be a pain.

Maybe results with both ICC and GCC could be done as a compromise? And for any tests that vary a lot, maybe some analysis on what ICC is doing differently. eg, generally more advanced auto-vectorization or anything that looks more like compiler breaking, or anything that legitimately relies on instructions specific to or favored on x86.

Not that I'm asking you to volunteer, of course I could ask on some other forums if anyone's interested.

 

[Nothingness]

I'm very interested in seeing this happen, I just don't have time. Or the platforms really. I assume all the benches purely driven by file I/O? Would there be any difficulty getting it running on Android? Trying to get a proper Linux distro running on all these platforms would be a pain.

I/O exists but it doesn't dominate run time. OTOH data sets can be large (several hundred MB). I've not tried to compile and run on Android, but I guess it's doable as code is portable (note I don't know how hard it would be to run FORTRAN workloads, since there's no FORTRAN compiled in the NDK and one would have to use a Linux cross-compiler).

Maybe results with both ICC and GCC could be done as a compromise? And for any tests that vary a lot, maybe some analysis on what ICC is doing differently. eg, generally more advanced auto-vectorization or anything that looks more like compiler breaking, or anything that legitimately relies on instructions specific to or favored on x86.

I have not seen such deep study of why/how/if icc breaks/targets SPEC (libquantum has obviously been broken, and in my experience icc gives much more speedup vs gcc on SPEC code than it gives on my own code).

Here is what I found (be careful with the first, Principled Technologies are too close to Intel for them to be fair...).
http://www.principledtechnologies.com/Intel/CompComp.pdf
http://www.centtech.com/wp-content/uploads/2014/09/WP1-NanoX2-whitepaper-1-3.pdf

Also note icc is not freely available any more for non-commercial projects.
https://software.intel.com/en-us/non-commercial-software-development

Not that I'm asking you to volunteer, of course I could ask on some other forums if anyone's interested.

I could not volunteer as I don't have many platforms at hand, and those I have often are not even at engineering sample stage :biggrin:

 

[NTMBK]

This would be the ideal moment for Samsung, LG, etc. to switch to Moorefield, since there won't be any better alternative.

Not really. Samsung has their own Exynos processor built on their own 20nm process, so they aren't terribly affected by any issues at TSMC.

 

[ShintaiDK]

Not really. Samsung has their own Exynos processor built on their own 20nm process, so they aren't terribly affected by any issues at TSMC.

Well they are product wise. Since majority of chips uses in Samsung products comes from TSMC.

 

[DeathReborn]

If it's true that it's not just overheating at a specific voltage but also problems with the memory controller & GPU then Qualcomm really could be up the creek with the 810. A57 itself isn't broken, it works in other designs so there is something Qualcomm are doing that has broken it, either a design tweak or process issues.

 

[witeken]

No kidding. Intel should have done another family of products on 22nm instead of betting the farm on 14nm mobile products in this timeframe.

"Goldmont" implemented in 22nm + ImgTec Series 6XT + improved ISP + H.265 encode/decode would have been a very potent combination in 2015.

I don't think 14nm is the reason why Broxton is delayed by half a year. Intel will have plenty of cheap, bleeding edge 14nm capacity in H2'15.

 

[witeken]

Hmm I thought Goldmont CPU improvements were mostly coming from higher freq (2.7GHz) which was possible only thanks to 14nm.

Goldmont is a Tock with a new architecture, the successor of 2.7GHz Airmont.

 

[Nothingness]

Goldmont is a Tock with a new architecture, the successor of 2.7GHz Airmont.

Thanks for fixing my mistake :thumbsup: I better understand Intel17 remark now.

 

[liahos1]

It does seem from comments BK has made that Intel is reassessing their strategy in mobile.

On one hand they have Apple as a key customer in their PCG biz and I guess indirectly in their DCG biz. So it has to be a fine line to walk between aggressively going after Apple's biz with partner enablement of premium SOC skus. In addition the premium segment in aggregate hasn't really been growing. So maybe they ask themselves why waste the resources and potentially butt heads with apple with premium phone SOC's when they are likely trying to win apple's business on foundry and/or baseband.

They also have Samsung as a customer with baseband, pc's etc but also a competitor for foundry and as a competitor in SOC's (internal use vs sourcing from intel)

So they do this deal with Rockchip and Tsinghua Unigroup and go after the low end but how much money is there in that market? It's not a leading statement I really dont know.

Strategy seems to be very multipronged at this point. Go after the meat of QCOM's profits (licensing) via legal agitation, destroy them on tablet as tablets are the closet cannibalizer of PCG and take share as a viable second source for LTE. Avoid direct conflagration with Apple while finding a way to make money off them. Slowly penetrate Samsung via baseband and their collaborative work in IOT and Tizen. Destroy all the weak players already likely making little to no money in the low end of the SOC market. Strategically align in China by playing to government's wishes for domestic semi production/ip leadership. Siphon ARM out of China Tech ecosystem via direct partnership/invsetment and ultimately block any advances ARM partners could make in DCG and IOT. All the while taking share from AMD in low end PC market. The ultimate affect of all of this is making ARM and TSM's life more difficult.

I guess with that many balls in the air, a rethinking of their mobile roadmap makes sense hence the 'delays'.

 

[tviceman]

I'd like to see Tegra K1 (Denver) got a shot in cell phones. I know it'd have to be downclocked and also coupled with a discrete modem, but it should still end up plenty fast (faster than the iphone 6 in graphics, competitive in CPU).

It'd be nice to see viable ARM alternatives to Qualcomm instead of every high end Android phone being Qualcomm.

 

[Arachnotronic]

I'd like to see Tegra K1 (Denver) got a shot in cell phones. I know it'd have to be downclocked and also coupled with a discrete modem, but it should still end up plenty fast (faster than the iphone 6 in graphics, competitive in CPU).

It'd be nice to see viable ARM alternatives to Qualcomm instead of every high end Android phone being Qualcomm.

NVIDIA and Broadcom are done in smartphones. Really, the only high-end players left vying for that market (as merchant chip vendors) are Qualcomm, MediaTek, and Intel.

 

[Thala]

NVIDIA and Broadcom are done in smartphones. Really, the only high-end players left vying for that market (as merchant chip vendors) are Qualcomm, MediaTek, and Intel.

I don't see how Intel is in a better position than NVidia. It would be Cherrytrail, which would be competing with Tegra K1...when it is available.
And neither NVidia nor Intel had some major design wins in Smartphones.

 

[Arachnotronic]

I don't see how Intel is in a better position than NVidia. It would be Cherrytrail, which would be competing with Tegra K1...when it is available.
And neither NVidia nor Intel had some major design wins in Smartphones.

Intel has publicly committed to staying in this market. NVIDIA publicly announced that it wants nothing to do with "mainstream smartphones."

And, say, when was the last time NVIDIA announced a cellular modem? NVIDIA also doesn't have in-house connectivity chips (Wi-Fi, Bluetooth, NFC, etc.) which essentially locks it out of the vast majority of the phone market.

I'm not saying Intel will definitely take tons of market share in phones, but they're investing so that they have a fighting chance.

 

[Thala]

Intel has publicly committed to staying in this market. NVIDIA publicly announced that it wants nothing to do with "mainstream smartphones."

I understand this, as they only invest in high-end as the Tegra K1 is pretty much a high-end only solution. They consciously were choosing the wording here.

And, say, when was the last time NVIDIA announced a cellular modem? NVIDIA also doesn't have in-house connectivity chips (Wi-Fi, Bluetooth, NFC, etc.) which essentially locks it out of the vast majority of the phone market.

There is a huge market for modems like xmm7260 to sell into high-end. In addition there are lots of IP vendors for connectivity solutions.
I see no problem in chosing a modem from a different vendor than the AP.
There is also no integration advantage for Intel, as SoFIA is low end only.

 

[soccerballtux]

I don't see anything wrong with big.LITTLE. But what is the point of have 4+4?

2+4 with aggressive scaling on the big 2 would probably be good.

you know, that's probably what apple will do next, but it'll be 2+2

 

[soccerballtux]

I don't see how Intel is in a better position than NVidia. It would be Cherrytrail, which would be competing with Tegra K1...when it is available.
And neither NVidia nor Intel had some major design wins in Smartphones.

how did Nvidia end up in this position? They're not idiots, but they've been acting like idiots with everything ARM related for quite some time, no?

 

[mrmt]

No kidding. Intel should have done another family of products on 22nm instead of betting the farm on 14nm mobile products in this timeframe.

I don't think 22nm is up to the task of fighting ARM, and Intel will have to pay for the 14nm transition any time in the future. So between spending on contra-revenue in order to keep bay trail on the mobile market until 14nm arrive and spending on a new 22nm design that would also need contra-revenue plus transitions to 14nm later in the future, Intel went with the former.

 

[Elixer]

Hmm, you know, it is possible for AMD to help them...they have working ARM + GPU, so, this could be just the big break they need.

 

[Idontcare]

Hmm, you know, it is possible for AMD to help them...they have working ARM + GPU, so, this could be just the big break they need.

On what process node and at which foundry?

 

[geoxile]

On what process node and at which foundry?

I imagine their 20nm A57 APU for skybridge will be at TSMC. But I'm pretty sure AMD's skybridge is a server platform, so I'm not sure it'll be well suited for tablets and phones

 

[tviceman]

Intel has publicly committed to staying in this market. NVIDIA publicly announced that it wants nothing to do with "mainstream smartphones."

And, say, when was the last time NVIDIA announced a cellular modem? NVIDIA also doesn't have in-house connectivity chips (Wi-Fi, Bluetooth, NFC, etc.) which essentially locks it out of the vast majority of the phone market.

I'm not saying Intel will definitely take tons of market share in phones, but they're investing so that they have a fighting chance.

Mainstream phones = low cost phones with modem integrated in SoC. Phones like HTC ONE is not a mainstream phone, it's a high end phone. And I'm not sure, but I'd think discrete Qualcomm modems would work just fine with SoC's other than their own.