AMD CPU speculation... and expert conjecture

[blackkstar]

The whole Nvidia thing looks to me like Nvidia finally gives up on Tegra and goes to just make money off of other people's ARM devices via licensing.

I think they too see the end of the ARM consumer boom and they are trying to come up with a back up plan. A significant portion of the ARM industry is planning on moving to servers. But an SoC like what Nvidia has isn't what you need there.

I personally think they are in the worst position out of any of the big chip makers.

Gaming GPU is a sketchy market as far as long term stability. Nvidia also has to compete with AMD, who has GPUs in all major consoles and has Mantle. Nvidia now has to convince game developers to optimize for their architecture, but doing so harms the console ports for consoles that have AMD GPUs.

Consumer ARM SoC has never gone anywhere for Nvidia. IIRC it's been a giant money sink for them and it's never managed to yield enough design wins to make the whole thing worth while. Tegra is on its 5th version yet we have yet to see it gain traction. To put this into perspective, it's like GTX 200 to GTX 900 (300 and 800 are getting skipped or were) all being complete failures in market. Or for AMD flavor imagine HD 2000 series to HD 7000 series being failures.

For ARM server, Tegra main selling point is their GPU. ARM servers are targeted toward microservers. Servers whose primary goal is to offer the smallest CPU possible with the most connectivity options. A large GPU is useless here.

GPGPU is in trouble. CUDA is going to have to compete with HSA, HSA foundation, and AMD, Samsung, etc. They also have to fight Intel Xeon Phi. "Marry the CUDA platform for life" sounds really unappealing in the face of OpenCL/HSA/HSA Foundation and Xeon Phi running x86 code.

Their only options going forward are HPC and embedded. Via and AMD are killing it in embedded (mostly Via though). HSA (and probably Xeon Phi) is going to be a hard fight for Nvidia. Their latest move is sort of scary to me. It shows me that Nvidia is acknowledging they have a rough future ahead and their cards to play are running low. They more than likely want to land some sort of long term revenue stream with these patents. In a way I sort of feel like this move is Nvidia started to admit they are going to lose in HPC, GPGPU, consumer GPU, mobile SOC, etc and they are trying to salvage what they have left.

We are at a big phase of transition for hardware. Gaming GPUs are so powerful now that mid range cards from last generation are more than enough. AMD is losing CPU war against Intel, and even though AMD can be 50% slower in single thread performance at times, that's still not enough to make it so slow that it's not worth buying, because there aren't any single threaded programs that are demanding enough to ruin an experience on AMD's single thread performance. At worst you get a somewhat inferior experience, but it's not bad enough that you can't even run it.

I feel AMD and Intel really do have a forward looking plan with their products, but Nvidia doesn't have anything that's taking off.

 

[jdwii]

Nvidia market cap is 10B Amd is only 3.2B. Today Cuda is WAY bigger then Opencl or hsa. Nvidia already has partners in the mobile market Amd does not and probably doesn't care. Nvidia has a better name in the industry then Amd ever did. Simple Amd needs to make great products and market them well.

 

[juanrga]

According to the links given before Nvidia has been negotiating with Samsung and others but the negotiation failed. Apple and others are also in the target, but it seems that Nvidia got previous agreements with those are not in the patent complaint.

I find funny that some people that says me that AMD is needed to give competition to Intel in CPUs at the other hand expect see Nvidia to die. I never understood this double standards. We need both AMD and Nvidia, because their competition brings evolution and improvements. a world with only Intel is so boring and expensive as a world with only AMD.

Also I like to remark that Tegra is currently an important part of Nvidia business. Nvidia sell millions Tegra per quarter:

Tegra business is still growing strongly for NVIDIA. Tegra revenue was up 14% from Q1 2015, and 200% from Q2 2014 with a total revenue of $159 million for the quarter. Tegra continues to have strong demand in the automotive infotainment sector, with a 74% growth in that market year-over-year. This could be a lucrative market, with automotive systems generally locking in for at least several years compared to the mobile sector which might see a product replaced in less than a single year. The Tegra K1 has j

Nvidia has announced several HPC/server ARM-based hardware. No sure why the myth that ARM servers are targeted toward microservers continuer alive in this thread. Specially when some of the announced ARM servers are so powerful like E5 Xeons.

In fact AMD Seattle is for microservers but K12 is not.

About the idea that Xeon Phi running x86 code I want to add that ordinary x86 code will run slower and recompilation will only help in some cases. In most cases obtaining performance from Phi implies rewrite of the code and the use of AVX extensions. The "it runs x86" is marketing motto.

 

[de5_Roy]

Nvidia market cap is 10B Amd is only 3.2B. Today Cuda is WAY bigger then Opencl or hsa. Nvidia already has partners in the mobile market Amd does not and probably doesn't care. Nvidia has a better name in the industry then Amd ever did. Simple Amd needs to make great products and market them well.

amd has them already. an arm a57 and a uarch license as well as x86 jaguar/puma cores, gcn igpu. lacks mainly two parts to compete in the consumer smartphone ph/tablet market: modem and a u.l.p. gpu uarch. otoh, intel and qualcomm (one competitor from each i.s.a.) have both those parts. intel even sources from imagination tech. both have much more moniez to spend, too. smartphone ph/tablet aren't profitable enough for amd as those are maturing at more accelerated rate than amd can afford(keyword here) to compete in and afford to allocate their limited r&d for 2-3 years to design another chip. the best option is to sit out till intel makes inroads then follow intel's lead and compete in x86 portables. in arm, qualcomm at the top and mediatek/allwinner/others at the bottom, there's no room for amd. it's not like amd lacks performance competitive socs or architectures, quite the opposite. the lack of lte modem has badly hurt intel for a while. when it did release a modem, qualcomm announced on-die integrated modem soon after, lol. amd cannot afford to fall behind like that at their present state. even if it could offer a large percentage of high end perf at a lower price. because mediatek et al can do the same for even lower price, and intel can ... contra revenue. amd neeeds to build a solid foundation first. seattle soc seems to have a decent chance among the people who want an arm server soc. the market and profitability may be very small, but amd seems to have the best chip among the players. if/when arm fizzles, amd can easily fall back on low clocked puma/zen cores. additionally, nvidia's campaigns might even strengthen amd-arm ties. with kaveri apus available now, and upcoming new gpus, it'll make money. overall, amd isn't lacking in terms of technology and performance. amd's shortcomings are elsewhere. for example, external limitations like fab process, choice of foundries, substrate, x86 pc market condition etc.

nvidia's design wins are nebulous (s/a ) at best. it burned quite a few bridges in the smartphone ph/tablet sector, cornered itself into selling own devices, mostly. the so-called design wins remind me of intel's "design wins". oems built devices around atom socs because intel paid them, used intel's money to build/strengthen their portfolio and p.r., then when the products got popular, they switched to cheaper parts. i think xiaomi, for example, is doing the same thing to get their phones sell outside china. nvidia seems to have good parts, although i've always thought of denver k1 as a very uh.. mellowed out one hit wonder. i'll wait a bit longer. nvidia is very good at p.r. though. always makes you think nvidia is [strike]sheen[/strike] "winning".

 

[de5_Roy]

^ nvida k1 32 yes i agree, it was late to market, softening any impact it may have had being released sooner, sound familiar ? (amd) Now the nvidia k1 64 is a different story. If reports hold up to be true k1 64 will be that game changer.

possibly. afaik, k1-denver may always be a generation or two behind it's contemporary socs due to the way it is designed i.e. may need a node shrink to be as power efficient as current ones. and it may be a deal breaker for many. is shield console, chromebooks or bulky tablets, the k1-d would be okay... but none of those really inspire oems to build around the soc. k1-d's current advantage is that it's readily available and processes 64bit arm v8 instructions. what will happen when arm catches up and google improves android's x86-64bit version? k1-d's gfx perf advantage also goes away as long as qualcomm and mediatek provides a large percentage of that performance at lower price and relatively high power efficiency. if oems can't/won't cut corners on the display or battery, the price and lower power use may be against k1-d's favor. and the hype... all that 7 way superscalar stuff became in order execution and equivalent to arm's 4 issue. this is based on what i know so far about k1-d.

edit: i mentioned nvidia's situation to compare with amd's. nvidia is more entrenched in arm ecosystem than amd, for longer, too. it also has powerful gpus. since both are players in the x86 market, nvidia's situation can be taken as a cautionary tale for amd. nvidia does have more resources yet could not stay in the competition. if amd had gone straight into those markets with it's current state, it would be worse off.

 

[h2323]

According to the links given before Nvidia has been negotiating with Samsung and others but the negotiation failed. Apple and others are also in the target, but it seems that Nvidia got previous agreements with those are not in the patent complaint.

I find funny that some people that says me that AMD is needed to give competition to Intel in CPUs at the other hand expect see Nvidia to die. I never understood this double standards. We need both AMD and Nvidia, because their competition brings evolution and improvements. a world with only Intel is so boring and expensive as a world with only AMD.

Also I like to remark that Tegra is currently an important part of Nvidia business. Nvidia sell millions Tegra per quarter:

Tegra business is still growing strongly for NVIDIA. Tegra revenue was up 14% from Q1 2015, and 200% from Q2 2014 with a total revenue of $159 million for the quarter. Tegra continues to have strong demand in the automotive infotainment sector, with a 74% growth in that market year-over-year. This could be a lucrative market, with automotive systems generally locking in for at least several years compared to the mobile sector which might see a product replaced in less than a single year. The Tegra K1 has j

Nvidia has announced several HPC/server ARM-based hardware. No sure why the myth that ARM servers are targeted toward microservers continuer alive in this thread. Specially when some of the announced ARM servers are so powerful like E5 Xeons.

In fact AMD Seattle is for microservers but K12 is not.

About the idea that Xeon Phi running x86 code I want to add that ordinary x86 code will run slower and recompilation will only help in some cases. In most cases obtaining performance from Phi implies rewrite of the code and the use of AVX extensions. The "it runs x86" is marketing motto.

Juan, Kinda off topic. Is there any solid information that AMD is going 14nm for the new cores, or is it just logic that they would considering the time frame?

 

[de5_Roy]

k164 is reported to be pin compatible to k1 32 making it a easy transition for oems, more choices. I really think the game changer will be code morphing or as nvidia calls it dynamic code optimization.

http://www.pcper.com/news/Processors/NVIDIA-Reveals-64-bit-Denver-CPU-Core-Details-Headed-New-Tegra-K1-Powered-Devices-La

http://www.extremetech.com/computing/174023-tegra-k1-64-bit-denver-core-analysis-are-nvidias-x86-efforts-hidden-within

i forgot about that 128MB cache, may be it'll be put on-package or on-die after a node shrink as LLC or L4 cache. strangely, the 128MB d.c.o. cache (in system memory) was casually kept away from die size and power measurements reveal yet counted in performance data despite it's significance .

i know very little about code morphing. may be that's why denver cores look so suspicious to me. another problems for me is the lack of independent reviews of a retail device based on the k1-d soc. i am gonna wait a bit longer for more info.

 

[truegenius]

Nvidia market cap is 10B Amd is only 3.2B. Today Cuda is WAY bigger then Opencl or hsa. Nvidia already has partners in the mobile market Amd does not and probably doesn't care. Nvidia has a better name in the industry then Amd ever did. Simple Amd needs to make great products and market them well.

amd has them already. an arm a57 and a uarch license as well as x86 jaguar/puma cores, gcn igpu. lacks mainly two parts to compete in the consumer smartphone ph/tablet market: modem and a u.l.p. gpu uarch. otoh, intel and qualcomm (one competitor from each i.s.a.) have both those parts. intel even sources from imagination tech. both have much more moniez to spend, too. smartphone ph/tablet aren't profitable enough for amd as those are maturing at more accelerated rate than amd can afford(keyword here) to compete in and afford to allocate their limited r&d for 2-3 years to design another chip. the best option is to sit out till intel makes inroads then follow intel's lead and compete in x86 portables. in arm, qualcomm at the top and mediatek/allwinner/others at the bottom, there's no room for amd. it's not like amd lacks performance competitive socs or architectures, quite the opposite. the lack of lte modem has badly hurt intel for a while. when it did release a modem, qualcomm announced on-die integrated modem soon after, lol. amd cannot afford to fall behind like that at their present state. even if it could offer a large percentage of high end perf at a lower price. because mediatek et al can do the same for even lower price, and intel can ... contra revenue. amd neeeds to build a solid foundation first. seattle soc seems to have a decent chance among the people who want an arm server soc. the market and profitability may be very small, but amd seems to have the best chip among the players. if/when arm fizzles, amd can easily fall back on low clocked puma/zen cores. additionally, nvidia's campaigns might even strengthen amd-arm ties. with kaveri apus available now, and upcoming new gpus, it'll make money. overall, amd isn't lacking in terms of technology and performance. amd's shortcomings are elsewhere. for example, external limitations like fab process, choice of foundries, substrate, x86 pc market condition etc.

nvidia's design wins are nebulous (s/a ) at best. it burned quite a few bridges in the smartphone ph/tablet sector, cornered itself into selling own devices, mostly. the so-called design wins remind me of intel's "design wins". oems built devices around atom socs because intel paid them, used intel's money to build/strengthen their portfolio and p.r., then when the products got popular, they switched to cheaper parts. i think xiaomi, for example, is doing the same thing to get their phones sell outside china. nvidia seems to have good parts, although i've always thought of denver k1 as a very uh.. mellowed out one hit wonder. i'll wait a bit longer. nvidia is very good at p.r. though. always makes you think nvidia is [strike]sheen[/strike] "winning".

64bit market is wide open and new
qualcomm charges more, mediatek offers slow, intel isn't 100% compatible, nvidia are rare to find ( atleast i India, last i saw i in tablet only )
once 64bit android is out, then 32bit soc based phones will be outdated ( marketing you know )
amd don't need to develop anything, they already have experience in arm or they can just use the 64bit cortex cores, they can borrow gpu, modem and other components from others like powervr, broadcom

there is a huge gap between qualcomm's krait and mediatek cortex-a7 ( and qualcomm cortex-a7 too, even big brands are using slow cores for high priced phones like xperia t3 crap or galaxy grand 2 junk ( over priced ) ) based phones and amd can sit in between these ( like xiaomi mi3, it is cheapest snapdragon 800 based phone which costs few bucks above moto g in India ) so amd can provide power of qualcomm krait, cheapness of mediatek and 100% app compatibility

btw i bought 6 phones in last 3-4 years out of which 4 are smartphones, because budget phones ( less than 200$ ) gave me reasons to upgrade ( budget phone which beat 2 year old flagship is good enough reason)
i upgrade my phone more often than pc ( because its cheaper to upgrade phone than already good enough pc )

 

[wh3resmycar]

funny a lot of you folks keep mentioning HSA. AMD can't even put Dynamic VSYNC in its graphics drivers. speculate if you will but please do not lose touch of reality here.

 

[juanrga]

k164 is reported to be pin compatible to k1 32 making it a easy transition for oems, more choices. I really think the game changer will be code morphing or as nvidia calls it dynamic code optimization.

http://www.pcper.com/news/Processors/NVIDIA-Reveals-64-bit-Denver-CPU-Core-Details-Headed-New-Tegra-K1-Powered-Devices-La

http://www.extremetech.com/computing/174023-tegra-k1-64-bit-denver-core-analysis-are-nvidias-x86-efforts-hidden-within

I agree on that dynamic code optimization is exciting by two reasons:

(i)
The whitepaper mentions the possibility to optimize over a 1000 instructions window for a maximum ILP greater than 7, which is a clear advantage over OoO hardware (Haswell can have 192 instructions on flight, which limits its ILP to 3.5) I asked if this means that Denver is the first KIP in the market. But I didn't receive convincing answers.

(ii)
Denver core can be reused to run Nvidia GPU ISA. Nvidia could use Denver cores inside future GPUs for assisting the SM. In fact this seems to be their plan.

 

[juanrga]

Juan, Kinda off topic. Is there any solid information that AMD is going 14nm for the new cores, or is it just logic that they would considering the time frame?

I have info from a friend that says that currently TSMC 16FF+ is the preferred choice over Samsung/Glofo 14nm FF, but things can change tomorrow.

There is little public information from AMD. Maybe those words from Keller about the 14/16nm process can help you

Much of the architectural work on the upcoming AMD chips is done and the company is well along in implementing them, Keller suggested. AMD is apparently already well along in the custom designs that will use a 16/14nm FinFET process that he said "looks pretty good, we're happy with it."

or maybe those other words:

Keller considers largely solved the issues dealing with multiple-patterning lithography required for the 14nm/16nm processes in which his team is designing the first K12 chips. Just adding on more metal layers is no longer an option, but "we have other proprietary tricks," he said.

For the 14nm/16nm process, his team had to scale challenges with the vertical FinFET transistors it uses. He notes of all the parts of his team, the design for manufacturing and test groups have grown the most over the years.

I don't know if I did stress this enough, but the foundries gap has reduced. Currently ARM/x86 products on 28nm planar or 32nm SOI are competing against Intel 22nm FinFET. But foundries have accelerated its strategies/roadmaps and Intel 14nm FinFET will be competing against Samsung/Glofo 14 FinFET and TSMC 16 nm FinFET.

A look to some basic node parameters show that Intel will only have half-node advantage over the rest.

 

[jdwii]

funny a lot of you folks keep mentioning HSA. AMD can't even put Dynamic VSYNC in its graphics drivers. speculate if you will but please do not lose touch of reality here.

They still don't have that i use that for ALL my games when i saw adaptive v-sync and read about it i was happy i usually just leave V-sync off. Edit just get RadeonPro

 

[wh3resmycar]

funny a lot of you folks keep mentioning HSA. AMD can't even put Dynamic VSYNC in its graphics drivers. speculate if you will but please do not lose touch of reality here.

They still don't have that i use that for ALL my games when i saw adaptive v-sync and read about it i was happy i usually just leave V-sync off. Edit just get RadeonPro

already have radeonpro. without it my 280x is broken. AMD should be paying the radeonpro guys millions imho. going back to the maximumpc interview with the nvidia guys (in answer to huddy), AMD creates new technology, make it open, then "HOPE" someone will fix it for them, much like HSA.

 

[juanrga]

already have radeonpro. without it my 280x is broken. AMD should be paying the radeonpro guys millions imho. going back to the maximumpc interview with the nvidia guys (in answer to huddy), AMD creates new technology, make it open, then "HOPE" someone will fix it for them, much like HSA.

I am curious, what has to be "fixed" on HSA according to you?

 

[de5_Roy]

^^ what's up with the hidden url? doesn't show up until i hit reply.

anywho, amd launched desktop kaveri, am1 platform then lower tdp dt kaveris and recently, the new fx cpus based on old pd architecture. compared to that, amd launched only two mobile lineups, mobile kaveri and beema/mullins.
they also launched the hawaii based cards, fully enabled bonaire based cards and the latest tonga based cards.
i don't see how that is ignoring desktop.

 

[de5_Roy]

He is a spammer testing for weaknesses in th forums. He has multiple accounts, his other name is asrdhery

sigh.. i thought those went away. iirc these copy part of legit posts, making them more difficult to detect.

 

[jdwii]

already have radeonpro. without it my 280x is broken. AMD should be paying the radeonpro guys millions imho. going back to the maximumpc interview with the nvidia guys (in answer to huddy), AMD creates new technology, make it open, then "HOPE" someone will fix it for them, much like HSA.

I am curious, what has to be "fixed" on HSA according to you?

He probably means support hardware without software is the most useless crap, Kinda like OpenCL and how its still not as big as Cuda sadly. Nvidia should have close to nothing since Amd and Intel and others(VIA-Snapdragon-ect) can use OpenCL. Amd is hoping others jump in and actually start making mainstream software that supports HSA i mean technically a A10 7850K could be creaming a 8350fx(even a I7) all the time if the software used the hardware to its full potential.
The problem is Amd is still the main one meaning no one cares until Intel does it.

Edit off topic
I think they should do an article to start proving these console fans wrong. The article should be about 3 year old hardware playing modern "next gen" games . That is the HD6900 series(580 series) generation and sandy-bridge generation. I'm actually willing to bet that the HD5800 series and I7 920(4 year old hardware) would still be able to play most next gen games at decent settings. Going further back wouldn't work over directx 11 compatibility. If only they can also include how much cheaper steam is and that fact that you can play older games as well without keeping the old hardware(unless you play games with 16bit code on a 64bit os).

 

[-Fran-]

From the times of "3D Now!", AMD has been putting great tech ahead, but with little to no adoption.

I think the only thing AMD was successful enough to put "ahead of its time" was 64 bits through x86-64, but that is mostly thanks to Microsoft I guess.

Anyway, to create stuff that actually takes advantage of the full HSA mambo jambo, the HSA foundation has to put incentives ahead. Samsung has been really silent about HSA when they have the money to do stuff. In any case, I think AMD still has some momentum left, so I hope we see things sooner than later.

Cheers!

 

[jdwii]

AMD is GARBAGE!!!! When I saw this video of a STOCK 2500k killing a 8350FX @4.8ghz in BF3 all I could do was laugh!
I can just imagine how badly my 4690k @4.7GHZ smokes the 9590FX or w/e they call their newest 8 core failure!
Watch this 2500k with a single 660 kill a 8350FX on a 7970GHZ! https://www.youtube.com/watch?v=HX6N9uJtzPA

This is obviously a troll video all you showed was how you come to false conclusions so quickly

 

[-Fran-]

Ignore the trolls, please.

Cheers!

 

[jdwii]

Ignore the trolls, please.

Cheers!

Its something the mods will take care of when they come back its sad they act like that and can't even find reliable information i work at a school and they teach kids how to do that now.

 

[wh3resmycar]

@juan

what jdwii said

 

[cdrkf]

To be fair Amd only outdid intel for 6 years or so 1999-2005. Before and after that intel has been superior with Amd having to sell more for less money than intel ever since making their CPU division unprofitable.

Who ever is on top will always get to pick the prices on their competition.

Who ever is on top will always get to pick the prices on their competition.[/quotemsg]

Actually AMD also outperformed Intel through the 286 and 386 era. It was around then that Intel managed to get the rules changed so that they no longer had to give their designs to AMD and only had to share the ISA (when AMD was first brought in as a second supplier, they were given the full Intel core design, then tweaked it to make it faster- AMD 286 and 386 chips were all better than Intels as there were effectively a 2nd revision!).

AMD kept up well during 486 as well, then were behind during the Pentium / Pentium II era before catching up again with the first Athalon.

 

[de5_Roy]

i have a suspicion that we might not see amd cpu/soc related official news until mid q4 or december...

s/a tests both fx8370 and it's 95w version
http://semiaccurate.com/2014/09/08/amds-fx-8370-8370e-30-watts-good/
vortez
http://www.vortez.net/articles_pages/amd_fx_8370e_review,4.html
lanoc
http://lanoc.org/review/cpus/6934-amd-fx-refresh-fx-8370-fx-8370e-fx-9370-and-fx-9590?showall=&start=4
runs cooler than centurion, too.

 

[-Fran-]

In any case... Now that the DDR4 train has started moving officially, I really hope AMD brings some news soon about a new socket for the APUs.

The speed for DDR4 are going up really fast and that's a good thing, I guess.

If they announce a new socket, they would need to change the package, right? Or they could have the same 938 pin socket config, but with a different layout? AM3++? haha.

Cheers!