AMD CPU speculation... and expert conjecture

[Ranth]

Juan can single handedly make this thread into an unfollowable mess.

Yeup, and it never cease to amaze me, should rename the thread to "Discuss and disagree with Juan"

 

[juanrga]

Hey Juan

I guessed you missed this one. http://www.tomsguide.com/us/tablet-dead-as-we-know-it,news-20401.html
go ahead and keep believing that the decline in advances will not affect tablet sales the same way it did with DT.

That article explains how current 32-bit tablets achieved a technical limit and explains ways how tablets will evolve. That article lacks the recent info disclosed by ARM and that is quoted in one of the links I gave before:

We've seen very strong uptake on our 64-bit processors, mainly for the tablet and mobile market, with two main ecosystems being Android and iOS

Also it is worth mentioning what AMd exec think about the future and how will be the future products of AMD

http://www.techradar.com/news/computing/the-future-of-amd-consoles-tablets-and-hybrids-1159869

http://www.tomshardware.com/news/Hybrid-Temash-Kabini-Kevin-Lensing-APU,22757.html

http://www.trustedreviews.com/news/the-notebook-is-gone-amd-exec-claims-hybrids-are-the-future

Note: "apusilicon", "Bright Side of News", "bits and chips" are not my sites. Moreover, I receive zero money by linking them or any other of the hundred sites I use in my posts.

 

[gamerk316]

Just remember: PCs are dying.

http://www.techspot.com/news/59731-nvidia-quarterly-earnings-top-wall-street-expectations-pc.html
http://techreport.com/news/27807/nvidia-rakes-in-record-quarterly-yearly-revenue

 

[king smp]

Wow I missed alot
In response to the AMD Vs Nvidia $/perf
At the time I bought my 660Ti at the $300 price point (December 2013) the 660Ti was the best buy IMHO
currently I would go with a 970 even with the controversy for bang for buck

 

[king smp]

While I have owned many AMD cards (CF 7770s,6870,5770,6670 are more recent ones) I must say that AMD is still lacking with their drivers.
Just recently I had a 6450 in my office computer and had driver crashes and issues with using a HDMI-DP adapters
i switched to a GT 630 and no issues.
one too many driver crashes over the years with AMD

 

[blackkstar]

Just remember: PCs are dying.

http://www.techspot.com/news/59731-nvidia-quarterly-earnings-top-wall-street-expectations-pc.html
http://techreport.com/news/27807/nvidia-rakes-in-record-quarterly-yearly-revenue

PC shipments from OEMs were also up around 13% from the same time last year at the last holiday season. DIY markets and OEM builds are growing. I'd wager it has a bit to do with no longer depending on an 8 year old console and ports for PC.

http://www.etforecasts.com/products/ES_pcww1203.htm
Look at sales in the 2008 to 2011 level. Xbox 360 came out in 2005. If anything, I'd say last gen consoles hurt PC sales the most. Now that we have new consoles, people have reasons to upgrade. I even want to get rid of my 7970 because I can't use MSAA in some games at 1440p because I use all my VRAM. I haven't had that problem in forever, and I bought my 7970 when it first came out.

Tablet market is collapsing, phone market is slowing down. PC and laptop is already starting to see a resurgence. If AMD can deliver a good platform in late 2015 or by mid 2016, they're going to do very well. AMD plans on moving into Asian markets too. AMD parts will absolutely dominate there. Older games are a lot more popular in those regions so they don't need overly demanding systems. High end APUs are absolutely perfect there, and if AMD can unify all consumer sockets, they'll be selling a ton of gaming chips to the Asian market while leaving them with a solid upgrade path.

Tablet market and phone market don't have demanding software that will drive upgrades. Once you have a device fast enough to browse the web and watch video without massive stuttering and issue, you have no need to upgrade. PC is far different and software is still increasing the hardware it wants, although it is just recovering from a huge lull. The only saving grace tablets and phones have for causing upgrades with this generation and next generation of hardware the battery and OEM's insistence on making sure they're not user replaceable. That and it seems like Android has something similar to Windows rot where it slows down over time. I've wiped old phones and re-installed Android and had it feel fast like a brand new phone before.

http://www.digitaltrends.com/mobile/best-buy-ceo-tablet-sales/
A tablet is plenty good for two or so years. If you have a stock gaming PC that's two years old, you're probably ready to upgrade, specially with next gen console ports coming. I was pretty surprised to see all 3GB of VRAM get eaten up so fast on my card, specially since benchmarks are so deceiving with their little bar graphs with 2GB cards. I guess when you take average frame rate it doesn't account for the huge stuttering and frame rate drops you deal with when running out of VRAM as it bounces between 80fps and 30fps. Look! It averaged to 60fps! But that's another rant for another day. My heart goes out to people who think even 4GB is going hold up, let alone 3.5GB.

 

[juanrga]

Wow I missed alot
In response to the AMD Vs Nvidia $/perf
At the time I bought my 660Ti at the $300 price point (December 2013) the 660Ti was the best buy IMHO
currently I would go with a 970 even with the controversy for bang for buck

Toms has a nice article about the 970

http://www.tomshardware.co.uk/nvidia-geforce-gtx-970-specifications,news-49637.html

The conclusion is that the incorrect specs controversy doesn't change anything of interest. The 970 card perform the same than the first day and concludes "we will continue to recommend it until there is a better-performing option for the price."

 

[noob2222]

Juan, your website is linked on every single post.

As for 64 bit arm, is that going to solve the main problem with tablets today? Crappy implementation of mobile websites such as toms here dumbing down the resolution of a 1080p phone to 360x640 are driving people away from tablets and back to desktops since the settings are not user accessible without "rooting". For the average user, thats not an option. Its getting worse with more "mobile friendly" websites.

 

[juanrga]

Juan, your website is linked on every single post.

As for 64 bit arm, is that going to solve the main problem with tablets today? Crappy implementation of mobile websites such as toms here dumbing down the resolution of a 1080p phone to 360x640 are driving people away from tablets and back to desktops since the settings are not user accessible without "rooting". For the average user, thats not an option. Its getting worse with more "mobile friendly" websites.

Learn what rel="nofollow" means and then learn what blogspot is because you mix stuff.

As for 64-bit, didn't you read the links that I gave claiming that the future of AMD is on consoles, tablets and hybrids? Did not you read the AMD exec claiming that notebook is gone?

64-bit means more performance and more memory. More memory means more complex OS and applications. AMD didn't develop any 32-bit tablet but has plans for future 64-bit tablets.

 

[-Fran-]

64-bits in tablets is just a FAD by most accounts, really. More memory won't improve what's really important: responsiveness of complex applications. What matters is from the improved ops you get in ARMv8; in particular the refresh to FP ops (from what I've read) and how they up the performance bar in a lot of scenarios.

I know that is implied, but let's make it explicit: having more addressable space in a Tablet means squat. Tablets need more *brute* performance at their current TDPs and get more complex software working inside of them (sounds like a familiar problem?) to make them work again as more than just fashion objects (or mail-reading appliances). Extend that to smartphones and wearables.

I really wonder if the tablets will suffer from the exact same issue PCs have today: enough processing for stuff -> no need to upgrade. Now, tablets are more "fashionable" just like smartphones, but if you think about regular use, they're far from being useful yet.

Cheers!

 

[gamerk316]

So...is now a good time to mention ARM is loosing efficiency as it's scaling up?

http://www.extremetech.com/computing/199187-first-samsung-cortex-a57-a53-chips-arrive-with-significant-performance-boosts

 

[-Fran-]

After reading Anand's article, I think most of the issues could be assigned to 20nm being "green" more than flat out stating ARMv8's A57/A53 designs are bad and they lost all the traction ARM was building around efficiency.

I think they're just getting more bloated (the ISA) and following the steps of X86 when MMX showed up; making the uArchs fatter than needed. Hope I'm wrong though.

Cheers!

 

[8350rocks]

While I have owned many AMD cards (CF 7770s,6870,5770,6670 are more recent ones) I must say that AMD is still lacking with their drivers.
Just recently I had a 6450 in my office computer and had driver crashes and issues with using a HDMI-DP adapters
i switched to a GT 630 and no issues.
one too many driver crashes over the years with AMD

You are talking about driver issues on a 5+ year old card? Really?

Try a newer card. The drivers are quite good these days.

 

[8350rocks]

So...is now a good time to mention ARM is loosing efficiency as it's scaling up?

http://www.extremetech.com/computing/199187-first-samsung-cortex-a57-a53-chips-arrive-with-significant-performance-boosts

Called it long ago...

You cannot get to the point you can compete with the big boy ISA (x86-64) in the big pond, even on the lower end, without becoming so much more complex than what ARM was when it was ridiculously efficient. ARM is going to have learn lessons that x86 did years ago, and take a crash course doing it. They have the luxury of hindsight, but they also have the misfortune of being a different architecture as well.

 

[juanrga]

So...is now a good time to mention ARM is loosing efficiency as it's scaling up?

http://www.extremetech.com/computing/199187-first-samsung-cortex-a57-a53-chips-arrive-with-significant-performance-boosts

No. That article is based in incorrect data and incorrect conclusions from Anand review. Not to forget that Anand did only run A32-based benchmarks when most improvements are from the new A64 ISA.

 

[con635]

Alot of doom and gloom in here recently, anyone seen reviews of Evolve? Time for toms to adjust the gpu hierarchy chart? The real 'win' in the console wins I think is showing in the last few ports with amd cards faring quite well vs maxwell and beating out Kepler, seems as if by magic the 7970ghz is now on par with 780 if you play the latest big multi platform games even nvidia gameworks titles.

 

[palladin9479]

All uArch's lose efficiency as they scale up, it's the natural consequence of adding more transistors that do important control work. I mentioned it long long ago that the vast majority of transistors in a modern x86 CPU aren't for doing math but for controlling the flow of data in order to feed the high powered engines.

Let me state this for everyone to know, binary math is not hard nor complex or power intensive. Binary math is so fast and easy to do that it scales linearly with clock speed. The problem with complex powerful process designs isn't about the speed of the math but getting enough information to actually do the math in the first place. Engineers have invented all sorts of cool and interesting methods to speed the control of the math up, everything from extremely-fast local storage of data (L1 cache), layered secondary and tertiary storage (L2/L3 cache) to building an electronic crystal ball to have what you need present before you need it (branch prediction and prefetch). The improvements from these tricks are so significant that the engineers now spend more time / energy creating and applying these tricks then actually building a faster math machine.

AMD and Intel both do binary math the same way, it's so simple to do that there is no difference in how the math is executed, and yet Intel commands a strong lead in raw efficiency. This isn't because they sacrifice kittens to some dark satanic digital god, it's because their engineers have found better tricks and have build a better electronic crystal ball then AMD's engineers. Massive vector processors (GPUs) operate under the same principles, nearly all the advances are now in how to feed more data efficiency to the processors rather then making the processors faster. We are now devoting large swaths of silicone to the control and input / output functions over adding more execution functions.

So yes ARM, like PPC, SPARC, MIPS, and x86 will experience the same scaling issue whereby they make the math units faster but need to then add exponentially more control logic to keep them active.

 

[-Fran-]

All uArch's lose efficiency as they scale up, it's the natural consequence of adding more transistors that do important control work. I mentioned it long long ago that the vast majority of transistors in a modern x86 CPU aren't for doing math but for controlling the flow of data in order to feed the high powered engines.

Let me state this for everyone to know, binary math is not hard nor complex or power intensive. Binary math is so fast and easy to do that it scales linearly with clock speed. The problem with complex powerful process designs isn't about the speed of the math but getting enough information to actually do the math in the first place. Engineers have invented all sorts of cool and interesting methods to speed the control of the math up, everything from extremely-fast local storage of data (L1 cache), layered secondary and tertiary storage (L2/L3 cache) to building an electronic crystal ball to have what you need present before you need it (branch prediction and prefetch). The improvements from these tricks are so significant that the engineers now spend more time / energy creating and applying these tricks then actually building a faster math machine.

AMD and Intel both do binary math the same way, it's so simple to do that there is no difference in how the math is executed, and yet Intel commands a strong lead in raw efficiency. This isn't because they sacrifice kittens to some dark satanic digital god, it's because their engineers have found better tricks and have build a better electronic crystal ball then AMD's engineers. Massive vector processors (GPUs) operate under the same principles, nearly all the advances are now in how to feed more data efficiency to the processors rather then making the processors faster. We are now devoting large swaths of silicone to the control and input / output functions over adding more execution functions.

So yes ARM, like PPC, SPARC, MIPS, and x86 will experience the same scaling issue whereby they make the math units faster but need to then add exponentially more control logic to keep them active.

Counterpoint: The ARMv8 ISA is not big enough (yet) to be as bloated as X86 and even more, they drop ARMv6 compatibility, so this ISA is not carrying dead weight like X86 is. I'm not disagreeing with your statements, but do they apply to ARMv8 till the point we can safely assume "man, they F'ed up big time"?

Like I said, I'm willing to put more blame on the half cooked 20nm from Sammy first than the actual implementation in silicon. I mean, the performance went up by a lot, but then so did power consumption. The SoC *is* more powerful and uses more power, but the efficiency still went up a bit; not as hoped by the reviewers, but still.

Cheers!

 

[palladin9479]

Counterpoint: The ARMv8 ISA is not big enough (yet) to be as bloated as X86 and even more, they drop ARMv6 compatibility, so this ISA is not carrying dead weight like X86 is. I'm not disagreeing with your statements, but do they apply to ARMv8 till the point we can safely assume "man, they F'ed up big time"?

Like I said, I'm willing to put more blame on the half cooked 20nm from Sammy first than the actual implementation in silicon. I mean, the performance went up by a lot, but then so did power consumption. The SoC *is* more powerful and uses more power, but the efficiency still went up a bit; not as hoped by the reviewers, but still.

Cheers!

Has nothing to do with the number of instructions needing to be supported, especially on a load / store architecture. L/S uArch's, otherwise known as RISC, don't really suffer from "instruction bloat" because their instructions already have extremely predictable execution times, adding support for more instructions doesn't slow things down. And even since x86 went super-scalar with decoders it no longer matters for them either. The "but x86 is such a bloated / blah / blah ISA!" is an old argument from a long time ago, it lost it's relevance when the Pentium Pro was created. Intel and AMD both utilize RISC uArchs internally, they are Load / Store designs with an instruction decoder that converts the x86 macro-ops into several RISC L/S mini-ops according to a set of preprogrammed microcode. That decoder has been integrated with the predictor and scheduler and is a key element to abstract and distribute work across multiple integer execution engine. When ARM implemented their own super-scalar out-of-order design, they had to create the same sort of instruction decoder themselves to abstract the incoming ARM code and distribute it across multiple internal integer units.

To be perfectly honest, supporting "legacy" ISA's is no longer an issue unless your developing software with such a small footprint that it runs on a watch, or the ECU inside a car (not to be confused with the media system). We're talking code with a footprint measured in kilobytes here.

 

[-Fran-]

Then it's right what I'm saying; although not exactly the way I said it, the problem doesn't lie in ARMv8 itself.

If there's a diminishing return that gives such a big difference in what you projected and the end result, the problem might lie somewhere else, right? 😛

Cheers!

EDIT: After giving it some thought, I came to my senses. The problem lies in OoO inside the new uArch. Since they targeted more performance, they had to go OoO a while back, so scaling up for performance makes each passing uArch way more complex. Maybe it's lineal, but they are getting diminishing returns on each gen. That's the short explanation and now I finally got it, haha.

 

[gamerk316]

So...is now a good time to mention ARM is loosing efficiency as it's scaling up?

http://www.extremetech.com/computing/199187-first-samsung-cortex-a57-a53-chips-arrive-with-significant-performance-boosts

No. That article is based in incorrect data and incorrect conclusions from Anand review. Not to forget that Anand did only run A32-based benchmarks when most improvements are from the new A64 ISA.

Call me when theres ARM64 software.

Oh wait, right back to that little problem. And devs hate maintaining two builds. Until ARM64 becomes the majority, pretty much everything is going to run ARM32. Just look how long Win64 took to even START to take off for reference.

I'll say it over and over: You can come up with the best HW in existence, but you need to convince people like me its worth their time and money to use it. And when 99.99999% of ARM platforms run ARM32, and .00001% runs ARM64, do you really think people like me are going to bother?

 

[Cazalan]

It is true that AMD will be limited to 4GB whereas Nvidia will not, but the explanation has nothing to do with 2.5D vs 3D stacking. In fact, it is false that Nvidia will use 3D stacking on the GPU. It is technically impossible. Both AMD and Nvidia use 2.5D stacking for high-end GPUs.

AMD is not limited any more than NVidia is when it comes to HBM. When HBM 2 is available AMD can use it just as soon as NVidia can.

HBM itself is a 3D stack so the resulting GPU is 3D regardless.

 

[truegenius]

is it typo or did they really mean fiji on "28nm"

I believe it is not a typo. 😉

if not then it looks they are not much behind nvidia in efficiency
and its coming from [strike]amd hater[/strike] you means good 😀

I believe this issue was settled before. One wouldn't confound the efficiency of the graphics architecture with the efficiency of the memory subsystem.

HBM reduces power consumption compared to GDDR5. Also I believe the memory stacks will use 20nm which will add to increase the overall card efficiency.

only if you would have read my reply correctly then you wouldn't have confused my memory power consumption figure with gpu efficiency

as you can see below, i was talking about how they could get 4096 gcn in 28nm under 300w limit of atx,
and i showed that they can because they reduced the power consumption of memory and some other tweaks

but i see that you purposely erased the part of my comment where i was talking about possibility of 4096gcn in 28nm and then mixed my first part of reply with word "memory" of my second part

if not then it looks they are not much behind nvidia in efficiency
and its coming from [strike]amd hater[/strike] you means good 😀

i currently tested hd7950 for power consumption (using hwinfo64) and found that gpu took 125w and vram took 42w max (tried furmark, heaven, 3dm11, gpuviewer opencl opengl, mix of all to get this reading)
the card was running at stock settings (same as reference), and max power consumption was 167w, thus fairly below 200w ceiling [strike](i guess nvidia would have rated this card at 150w tdp i mean card power, just like gtx980)[/strike] 😗
so it looks like 4096 gcn on 28nm under 300w is possible by using stacked vram (cutting power usage here), reduced computation performance, aggressive voltage/clock setting, and maybe more mature 28nm ?

Juan, your website is linked on every single post.

As for 64 bit arm, is that going to solve the main problem with tablets today? Crappy implementation of mobile websites such as toms here dumbing down the resolution of a 1080p phone to 360x640 are driving people away from tablets and back to desktops since the settings are not user accessible without "rooting". For the average user, thats not an option. Its getting worse with more "mobile friendly" websites.

exactly, and due to this crappy implementation on other sites too i am not able to read extremetech link shared by gamer below in opera mini
tried to open it in chrome and it showed me mobile site, then i manually selected "request desktop site" in chrome menu and it still showed me mobile site, clicked on the www.extremetech.com link given on that page and then it showed me desktop site for 1 second and then directed me to mobile site again. I use anandttech's bench section very much but it doesn't work correctly in opera (though it works with chrome) so i switched to chrome and this still causes inconvenience (switching back and forth between browsers) but tolerable.
after this i will never click on the extreme tech search results or links when using mobile/opera
i read this post of yours in mobile but couldn't reply because there was no button for quote, so no on the go toms browsing (without pain)

and now sub 80 IQ holder analysts will say that people prefer mobile sites in mobile, i request to analysts like this to quit their job it will be much much more helpful than their analysis
what else people will do when they can't access desktop site after applying all tricks they could

So...is now a good time to mention ARM is loosing efficiency as it's scaling up?

http://www.extremetech.com/computing/199187-first-samsung-cortex-a57-a53-chips-arrive-with-significant-performance-boosts

 

[Embra]

All uArch's lose efficiency as they scale up, it's the natural consequence of adding more transistors that do important control work. I mentioned it long long ago that the vast majority of transistors in a modern x86 CPU aren't for doing math but for controlling the flow of data in order to feed the high powered engines.

Let me state this for everyone to know, binary math is not hard nor complex or power intensive. Binary math is so fast and easy to do that it scales linearly with clock speed. The problem with complex powerful process designs isn't about the speed of the math but getting enough information to actually do the math in the first place. Engineers have invented all sorts of cool and interesting methods to speed the control of the math up, everything from extremely-fast local storage of data (L1 cache), layered secondary and tertiary storage (L2/L3 cache) to building an electronic crystal ball to have what you need present before you need it (branch prediction and prefetch). The improvements from these tricks are so significant that the engineers now spend more time / energy creating and applying these tricks then actually building a faster math machine.

AMD and Intel both do binary math the same way, it's so simple to do that there is no difference in how the math is executed, and yet Intel commands a strong lead in raw efficiency. This isn't because they sacrifice kittens to some dark satanic digital god, it's because their engineers have found better tricks and have build a better electronic crystal ball then AMD's engineers. Massive vector processors (GPUs) operate under the same principles, nearly all the advances are now in how to feed more data efficiency to the processors rather then making the processors faster. We are now devoting large swaths of silicone to the control and input / output functions over adding more execution functions.

So yes ARM, like PPC, SPARC, MIPS, and x86 will experience the same scaling issue whereby they make the math units faster but need to then add exponentially more control logic to keep them active.

Would APU's see much of a performance boost with L3 cache? If so, is it a space/price issue?

 

[noob2222]

Ok juan, you caught me. I didnt read your article links, but now that i did, im laughing hysterically.

The toms article talks about the prediction made by someone (dont feel like looking at the name atm) stating 2 years ago that tablet fad will fade in 5 years. Toms stated that it only took 2 years to start to come true.

The articles you linked to defend arm tablets ... ironically ... are from 2 years ago ... roflmao.