AMD CTO Mark Papermaster Outlines Vision for 'Surround Computing'

[btarunr]

In a keynote address opening the prestigious annual Hot Chips symposium, AMD CTO Mark Papermaster delivered a vision for the coming "Surround Computing Era", and unveiled new processor architecture details, enabling technologies and design methodologies that will help drive the next era in computing. Surround computing is an extension of pervasive and ambient computing trends and describes an environment where computing technologies are completely natural and seamless parts of daily life.

"Surround computing imagines a world without keyboards or mice, where natural user interfaces based on voice and facial recognition redefine the PC experience, and where the cloud and clients collaborate to synthesize exabytes of image and natural language data. The ultimate goal is devices that deliver intelligent, relevant, contextual insight and value that improves consumers' everyday life in real time through a variety of futuristic applications. AMD is leading the quest for devices that understand and anticipate users' needs, are driven by natural user interfaces, and that disappear seamlessly into the background," said Papermaster during his opening remarks.

Papermaster explained that the Surround Computing Era will rely on robust "plug-and-play" IP portfolios including central processing units (CPUs), graphics processing units (GPUs), fixed function logic, and interconnect fabric. He also unveiled key details of AMD's upcoming "Steamroller" CPU architecture while underscoring the benefits of the industry-standard Heterogeneous Systems Architecture (HSA) that enables software developers to easily assign scalar and parallel compute workloads to the most appropriate compute units, and therefore optimize power.

"The road that leads us to the Surround Computing Era will be no less challenging and every bit as exciting as the 20-year journey in graphics processing that brought gamers from 'Pong' to today's modern game titles that feature stunning visual realism," Papermaster explained. "It will take an industry movement to complete this journey, and HSA provides the clear path forward to enable this next generation in computing."



View at TechPowerUp Main Site

 

[Hustler]

If only AMD's engineers were to the same standard as their marketing dept.

'Surround Computing'...

 

[Freedom4556]

If only AMD's engineers were to the same standard as their marketing dept.

I actually feel bad for AMD's marketing department. It must be tough to know that your CPUs are inferior to the competition in every way (which it unfourtunately is at this point) other than price but you still have to try to sell it to people. It's a wonder why we haven't heard of more of them leaving. I wish they were better.

 

[Frick]

Which is why APUs are here.

 

[Steevo]

Freedom Fabric? Who writes this shit?


I like the concept, but again, it is wasted by understating the complexities that we nerds will understand, and overstating the bullshit spin that makes us want to vomit.

 

[suraswami]

Good Dream!

 

[theJesus]

Is "Papermaster" that dude's real name?

 

[Xzibit]

Is "Papermaster" that dude's real name?

Yes and if I recall correctly hes the one that said Sea Island was tapped out back in June and that they were going BULK manufacturing from 2013 forward.

 

[hat]

I actually feel bad for AMD's marketing department. It must be tough to know that your CPUs are inferior to the competition in every way (which it unfourtunately is at this point) other than price but you still have to try to sell it to people. It's a wonder why we haven't heard of more of them leaving. I wish they were better.

Straight performance wise, no, but AMD cpus are cheaper, and their platforms are more upgrade friendly.

 

[HumanSmoke]

Straight performance wise, no, but AMD cpus are cheaper

I'm sure AMD's bean counters are over the moon with the average selling price...and if good-equals-good-enough then AMD's perceived CPU value doesn't necessarily bear close scrutiny against Sandy/Ivy Bridge Pentiums/Celerons

and their platforms are more upgrade friendly.

Helpful that AMD are happy to see the 800/900 chipsets soldier on into it's third year- not so much a marketing bonanza for motherboard makers.

I'm not entirely sure that "upgrade friendly" applies to Hudson (FM1 socket) owners either. A year from launch to what is effectively EOL.

 

[sergionography]

if you guys look at the slides especialy the ones for steamroller, you will realize amd is moving away a bit from the sharing module architecture by giving each integer core its independant decoders unlike bulldozer/piledriver that share 4 decoders. that alone should give a massive increase in scaling and ipc and latency(since now they wont need to depend on high latency to give enough time for the decoders to feed each core)
i would expect over 50% improvement from bulldozer to steamroller just based on that

http://www.techpowerup.com/img/12-08-28/154o.jpg

 

[HumanSmoke]

Is "Papermaster" that dude's real name?

I think it's an amalgam of AMD strengths ( Paper Tiger and Power Point Master)

 

[Recus]

It looks good on paper, but in reality it's just wood.

 

[Steevo]

if you guys look at the slides especialy the ones for steamroller, you will realize amd is moving away a bit from the sharing module architecture by giving each integer core its independant decoders unlike bulldozer/piledriver that share 4 decoders. that alone should give a massive increase in scaling and ipc and latency(since now they wont need to depend on high latency to give enough time for the decoders to feed each core)
i would expect over 50% improvement from bulldozer to steamroller just based on that

http://www.techpowerup.com/img/12-08-28/154o.jpg

50% improvement, with just that? Wow, if only engineers would have consulted with you earlier.



AMD had an idea, and if it would have happened when Phenom I did, it would have been great, but it didn't instead it has been a slow motion belly flop at the olympics diving competition.

None of their own slides which are marketing and investment seeking mention anything anywhere the 50% mark, 20% I can believe through process improvements, and some tuning of the cores. Is 20% enough? Depends on what 20% they are going to improve, if they think throwing more cores into the mix and the resulting 20% increase is enough they are sadly mistaken. If they are talking about a 20% IPC per core increase on highly dependent branched execution I will applaud them for making it.

 

[xenocide]

Straight performance wise, no, but AMD cpus are cheaper, and their platforms are more upgrade friendly.

Cheaper and Upgrade Friendly from what I've seen means you spend more over a longer time line for the same level of performance. My 2500k will easily last me until Haswell, but if I had gone BD I would need to upgrade immediately to PD and then probably Steamroller to get the same level of performance for what I use my computer for.

 

[hat]

Such is the enthusiast world... could you really say you wouldn't be happy with BD's performance? The only reason I went up from my Phenom II 550 to the i7 920 is because a bunch of friends happened to have a CPU and a motherboard between them and I got a good deal. I recall the 550 lagging behind in some situations where more cores would have helped out a lot... not as much as the increased single thread performance. Sure that's a benefit that doesn't go without enjoyment, but without these deals available to me I would have dug up a PII x4 or x6 and called it a day.

 

[sergionography]

50% improvement, with just that? Wow, if only engineers would have consulted with you earlier.



AMD had an idea, and if it would have happened when Phenom I did, it would have been great, but it didn't instead it has been a slow motion belly flop at the olympics diving competition.

None of their own slides which are marketing and investment seeking mention anything anywhere the 50% mark, 20% I can believe through process improvements, and some tuning of the cores. Is 20% enough? Depends on what 20% they are going to improve, if they think throwing more cores into the mix and the resulting 20% increase is enough they are sadly mistaken. If they are talking about a 20% IPC per core increase on highly dependent branched execution I will applaud them for making it.

lol well true 50% was probably too optimistic, but dont rule it out.
trinity is already 15% faster per clock than bulldozer, by what ur saying then ur pretty much saying steamroller will only be 5% faster than piledriver. we all know the biggest bottleneck for bd is the shared front end and with dedicated decoders that problem is solved, so there goes the 20% multithreading scaling penalty and the core starving for resources
So yes 50% over bulldozer or 30% over piledriver is very much likely, that's what is required to put it on par or a bit faster than ivy. But that is based on theory lets hope amd can make it happen. remember 2 cores in bulldozer used 2 share 4 decoders so thats like 2 decoders per core which is pretty wimpy(bobcat has 2 decoders per core)
The slides do mention 30% better ops per cycle and 5-10% better efficiency in another slide. But we will wait and see

 

[Steevo]

I will believe it when I see it. Where has Trinity been shown to be 15% higher IPC than bulldozer, which is slower in IPC on many tasks than my 1100T?


We all know the biggest bottleneck, who is we, Bulldozer thrives on memory intensive multithreaded applications. It fails hard at single threaded and branching instructions due to mismanaged data scheduling between "cores".


AMD has always been behind in cache hit efficiency, and unless they pull apart a I series and copy it they are not likely to gain too much ground there.


All told I'm sure there might be one benchmark that it is 50% faster at, but in real world situations it is only going to be 20-30%* faster than current design.


*Based on Simulated designs, and simulated workloads done by AMD, of course, for the purpose of keeping their jobs.



I speculate by simulating a simulation of simulations that I will run faster after my morning movement. See how that works, plausible deniability if I don't or they don't since it was based on simulated results obtained from a simulation.

 

[Super XP]

Interesting. Blame AMD's past BOD for ridiculous choices regarding CPU design decisions. What got done is done, scrap the automation that killed Bulldozers performance and now make it work.

I love the colourful pictures in those slides.

 

[HumanSmoke]

@Super XP

TRUE STORY: 5 present AMD board members have held their position since 2006 (i.e. predating BD) back when AMD's board comprised 7 members. AMD just keep adding board members...I think they are up to eleven - nine of which comprise AMD's governance....so, 71% of the deadwood ( the "ridiculous choice makers" as you put it) that signed off on BD still man the helm at the company.

 

[sergionography]

I will believe it when I see it. Where has Trinity been shown to be 15% higher IPC than bulldozer, which is slower in IPC on many tasks than my 1100T?


We all know the biggest bottleneck, who is we, Bulldozer thrives on memory intensive multithreaded applications. It fails hard at single threaded and branching instructions due to mismanaged data scheduling between "cores".


AMD has always been behind in cache hit efficiency, and unless they pull apart a I series and copy it they are not likely to gain too much ground there.


All told I'm sure there might be one benchmark that it is 50% faster at, but in real world situations it is only going to be 20-30%* faster than current design.


*Based on Simulated designs, and simulated workloads done by AMD, of course, for the purpose of keeping their jobs.



I speculate by simulating a simulation of simulations that I will run faster after my morning movement. See how that works, plausible deniability if I don't or they don't since it was based on simulated results obtained from a simulation.

http://www.tomshardware.com/reviews/a10-5800k-a8-5600k-a6-5400k,3224-2.html

here it is, a couple benchmarks but thats all we have untill piledriver comes out i guess
still piledriver will be atleast 10% faster just due to the stock clock speed which is 4.0ghz, so no need 2 get all pessimistic. as far as i can tell AMD has went a far way in terms of efficiency.
now if anything amd just need 2 get their cores more beefy and it seems thats exactly what they plan to do with steamroller, and if anything they are moving away with wat ruined bulldozer. tho i would argue they sure can benefit from making the pileline shorter. i believe they will keep it at 20 stages. but who knows maybe they plan on doing so with excavator
or maybe sticking to higher clocks was their solution to complement their slower cache subsystem, aslong as they can work around it im good i guess.

 

[Steevo]

Seems like a company tried this already. Oh yeah, anyone premember prescott? I havd one and it sucked. Shorter pipeline helps with branch mispredictions but only up to a point.


I remember too well what Intel tried and it wasn't so much that AMD had superior technology under the hood at the time as much as it was that Intel screwed the pooch. Now we are reversed, AMD keeps trying to spin this flop as if it were somehow great, then they have the balls to try and spin shit they should have been doing 5years ago and alot of us knew it as if it were some rainbow and magic pixie dust idea that no one has ever had before.


Insulting on multiple levels, and to really top it all off, I bet they flop the "fabric" setup and it dies a stillborn death like half of their great ideas.

 

[sergionography]

Seems like a company tried this already. Oh yeah, anyone premember prescott? I havd one and it sucked. Shorter pipeline helps with branch mispredictions but only up to a point.


I remember too well what Intel tried and it wasn't so much that AMD had superior technology under the hood at the time as much as it was that Intel screwed the pooch. Now we are reversed, AMD keeps trying to spin this flop as if it were somehow great, then they have the balls to try and spin shit they should have been doing 5years ago and alot of us knew it as if it were some rainbow and magic pixie dust idea that no one has ever had before.


Insulting on multiple levels, and to really top it all off, I bet they flop the "fabric" setup and it dies a stillborn death like half of their great ideas.

yea so true, they have amazing recipes but always come up with something half baked, it has to do with the budget I believe, Intel can afford extensive testing and sampling prior to release while amd it seems usconsumers test its prototypes until they improve an architecture. Phenom I for example, had it been tested properly it wouldn't have had a bug and whatnot.