AMD Beema/Mullins Performance Preview

[Kai Robinson]

http://anandtech.com/show/7974/amd-b...rmance-preview

Impressive power savings, very impressive. I can't surely be the only person to think that these Jaguar/Puma cores have far more potential than Piledriver/Steamroller?

[qcmadness]

The always crucial factors for AMD products:
1. Design win
2. Design win
3. Design win

[zalbard]

Impressive power savings, very impressive.

The preview says nothing like that. In fact,

The big unknown is how these new SoCs stack up against Bay Trail when it comes to power consumption.

Performance is definitely good, though.

[SEA]

The preview says nothing like that. In fact,

Performance is definitely good, though.

The power consumption there is all over the every page, everywhere. With numbers and graphs.
Really impressive - 4.5W chip outperforms 15W one by far in most tests, from 0% to 250%, to be more specific.

The gains in performance come while decreasing platform power. You can now have roughly the same performance as AMD offered last year in a 15W entry level notebook part, in a 4.5W TDP (2.8W SDP) tablet SKU. That’s seriously impressive.

[LightSpeed]

This is easily the largest gain ive ever seen without a process change. Impressive stuff indeed. Also, it has 80% of Trinity's single threaded performance with a 64bit memory controller and a silicon optimised for power. AMD should really scale up this architecture

[tbone8ty]

http://www.techspot.com/article/810-...iew/page5.html

I like techspots conclusion

[zalbard]

The power consumption there is all over the every page, everywhere.

Marketing speak is "all over the every page, everywhere".
TDP is no more than a marketing term these days, especially when it comes to mobile devices, with every manufacturing listing it the way they see fit. Real tests is what matters, and there are none.

It's true that these chips are often on par with 15W parts by AMD. But they are not meant to compete. Direct competitors are Atom (2W SDP, unlikely to be twice as efficient) and ARM chips.

[Darakian]

I look forward to the days after the bulldozer arch runs it's course. Hopefully amd will pull an intel and do a hybrid cat/construction equipment arch (ala the core 2).

[[XC] Lead Head]

If I were AMD, I'd be trying to scale up Jaguar to a full-sized desktop/notebook part as soon as possible. Seems to have infinitely more promise than faildozer.

[LordEC911]

This is easily the largest gain ive ever seen without a process change. Impressive stuff indeed. Also, it has 80% of Trinity's single threaded performance with a 64bit memory controller and a silicon optimised for power. AMD should really scale up this architecture

If I were AMD, I'd be trying to scale up Jaguar to a full-sized desktop/notebook part as soon as possible. Seems to have infinitely more promise than faildozer.

There is a reason why they can't do that.

[vario]

There is a reason why they can't do that.

And what that reason is ?
You know, wont and cant arent the same.
8 Core puma with a 1gb of dedicated GDDR5 at something along the lines of 2.5ghz should be damn easy for them to do.They are doing almost the same for PS4`s.
Im pretty sure its a mix of in company politics and lack of cash.But im also pretty sure they COULD and should.

[Darakian]

Not to take away from the 8 core puma (which I would love to see), but intel had a testing chip where they took 80 cores (original atom cores iirc) and stuck them on the same chip. This was done to drive home the issues of multi core chips. How big does the cache need to be? How fast of a bus? Should there be more than one bus? How do you keep private caches coherent? etc...
I forget if this thing eventually became larabee or if it was it's own project, but either way it should be clear that there are "issues" when trying to scale up something designed to work in a low power envelope.

Actually on puma, it's only got a single memory channel and that alone could be cause for a massive overhaul of the chip if they scale it up.

[vario]

Not to take away from the 8 core puma (which I would love to see), but intel had a testing chip where they took 80 cores (original atom cores iirc) and stuck them on the same chip. This was done to drive home the issues of multi core chips. How big does the cache need to be? How fast of a bus? Should there be more than one bus? How do you keep private caches coherent? etc...
I forget if this thing eventually became larabee or if it was it's own project, but either way it should be clear that there are "issues" when trying to scale up something designed to work in a low power envelope.

Actually on puma, it's only got a single memory channel and that alone could be cause for a massive overhaul of the chip if they scale it up.

Eh.Puma is the designation for just cores.This architecture is very flexible , amd can add or remove whatever they want, they even offer very large buyers opportunity to do so.Thats how Xbox one (dual DDR3) and PS4 (dual GDDR5) came to be, you can add edram cache (xbox one), amd has memory controllers ready for dual ddr3, dual lddr3, single channel or gddr5.They dont have to do anything from scratch.They have all the buliding blocks ready.L2 caches are pretty much what they need to be.L3 can be added or edram , the cpu has most of the chipset funtions integrated, the cpu cores+L2 are very small ,doubling them wont cause any havok (and lets remember pretty much the same cores are in consoles already in 2 different 8 core configurations).
Puma is basically fixed up jaguar with better turbo and power management.
Im not talking about revamping it for fully fledged desktops , that would require some major overhauls in both architecture and process used, not that it would be impossible...
Entry level gaming laptops would be very interesting with this chip if amd would let it chew more power, games are becoming more and more threaded, mantle would also give this kind of a chip a big boost , dx12 will probably too.But thats just me, i think it would be a hit when properly configured, decent gaming on the cheap with low power consumption.Lets not forget intel makes an 8 core atom now, and they sell it for 178$, asrock recently made storage/home server centric mainboard with it and it goes for 350$, why not AMD ?Because AMD doesnt have such a a chip...
AMD sells MS an 8 core jaguar with 768shaders AND 32mb edram chip for 100$. Im pretty sure they could sell puma based 8 core with lets say 512 shaders and no edram for 100$ to notebook makers.

[massman]

If I were AMD, I'd be trying to scale up Jaguar to a full-sized desktop/notebook part as soon as possible. Seems to have infinitely more promise than faildozer.

AM1 Kabini has Jaguar cores

[god_43]

snip.

the cat cores dont scale very well, and if you try to they become extremely difficult design, those are low powered cores designed from the start. steamroller is still more powerful performance wise, though you still have efficiency of that performance not as big as it should be. excavator should be a good performance chip, however i am sure the chip that comes after excavator arch will be a better high performance core that puts amd back on equal footing with intel.

[LordEC911]

And what that reason is ?
You know, wont and cant arent the same.
8 Core puma with a 1gb of dedicated GDDR5 at something along the lines of 2.5ghz should be damn easy for them to do.They are doing almost the same for PS4`s.
Im pretty sure its a mix of in company politics and lack of cash.But im also pretty sure they COULD and should.

See below.

the cat cores dont scale very well, and if you try to they become extremely difficult design, those are low powered cores designed from the start. steamroller is still more powerful performance wise, though you still have efficiency of that performance not as big as it should be. excavator should be a good performance chip, however i am sure the chip that comes after excavator arch will be a better high performance core that puts amd back on equal footing with intel.

@Vario There is a reason you see different architectures, layouts and designs when moving through market segments. It isn't because they are lazy or decided to waste resources but because that is what you need to do when specific markets require different target goals, whether performance, power, die size, or a specific mix of them all.

[vario]

See below.



@Vario There is a reason you see different architectures, layouts and designs when moving through market segments. It isn't because they are lazy or decided to waste resources but because that is what you need to do when specific markets require different target goals, whether performance, power, die size, or a specific mix of them all.

Well we can agree to disagree, also im not sure were talking about the same things.Im talking about cheap low power 8 cores with a decent igpu instead of faildozers at low-mid segment.
I can also say that neither ms nor sony took the faildozer way and thats probably telling us more than the actions amd managament takes for years now.I never said engineers at AMD were lazy or waste resources.I am telling amd`s ceo's /board and upper management makes all kind of wrong decisions for years now.
Im pretty sure there would be market for 8 core puma, you disagree.I get that.No point in dragging this more.

Just one more thing to add, i always smile when people in IT invoke the holy "They KNOW what theyre doing ,and its certainly the best, they work there for years!"
That sounds logical, but anyone in little interest in technology knows that most companies make WRONG decisions when it comes to predicting the future, and most of that business people are even more wrong.Besides countless others IBM`s executive`s "World needs a handful of computers", and even that exact quote is debated, it was "obvious" then, that there is gonna be "small" market.
Also i just remembered something, it sounds so silly today , i remember Intels engineers making SO WRONG predictions about their new "netburst" architecture, 10ghz chips were supposed to be like 3 years ago :P .So even the best engineers, can make wrong predictions.And of course ,AMD expected higher frequency with dozers.

[Darakian]

*snip*

You're making a lot of assumptions there.

Eh.Puma is the designation for just cores.This architecture is very flexible , amd can add or remove whatever they want, they even offer very large buyers opportunity to do so.Thats how Xbox one (dual DDR3) and PS4 (dual GDDR5) came to be, you can add edram cache (xbox one), amd has memory controllers ready for dual ddr3, dual lddr3, single channel or gddr5.They dont have to do anything from scratch.They have all the buliding blocks ready.

Well yes the ps4 and xbox1 are both semicustom designs with similar building blocks, but by the looks of it the ps4 has 3 memory controllers while the xbox1 has 2. See chip diagrams below.
xbone: http://www.chipworks.com/en/technica...-the-xbox-one/
ps4: http://www.chipworks.com/en/technica...ore-processor/

So yes the building blocks are there, but those are totally custom architectures. The OS on each device will have maybe 10 things going in the background while a normal windows/linux box could have 100 and that might be enough to hide whatever latencies are in there. I don't know as I am no micro architecture engineer, but I don't think you can claim that just because it works on a console that it will work on a general purpose computer.

L2 caches are pretty much what they need to be.L3 can be added or edram , the cpu has most of the chipset funtions integrated, the cpu cores+L2 are very small ,doubling them wont cause any havok (and lets remember pretty much the same cores are in consoles already in 2 different 8 core configurations).

How certain are you about that? More cache can increase the latency on storage and retrieval as there's just more to search for a given bit of data. Maybe the cores are tuned to expect a given latency and more cache would throw them out of wack (technical term).
Here's a good read on jaguar's memory and cache architecture if you'd like to read up on it.
http://www.realworldtech.com/jaguar/6/

Puma is basically fixed up jaguar with better turbo and power management.
Im not talking about revamping it for fully fledged desktops , that would require some major overhauls in both architecture and process used, not that it would be impossible...

Not sure where's you're getting that. Maybe, but I don't see the evidence one way or the other.

Entry level gaming laptops would be very interesting with this chip if amd would let it chew more power, games are becoming more and more threaded, mantle would also give this kind of a chip a big boost , dx12 will probably too.But thats just me, i think it would be a hit when properly configured, decent gaming on the cheap with low power consumption.Lets not forget intel makes an 8 core atom now, and they sell it for 178$, asrock recently made storage/home server centric mainboard with it and it goes for 350$, why not AMD ?Because AMD doesnt have such a a chip...

In order of bolded text:
Sure if it works well; Some are yes; Big is a vague term. 20% might be big but if 20% = 3 fps then... not so much; That's the ticket right there, we don't know what "properly configured" means or costs; Might not actually be on the cheap if development costs are huge; And they sell it as a server cpu where threads are usually the performance metric (as opposed to memory bandwidth, IPC, or whatever); You seem to be arguing that they do and that that it's in the xbox1/ps4

AMD sells MS an 8 core jaguar with 768shaders AND 32mb edram chip for 100$. Im pretty sure they could sell puma based 8 core with lets say 512 shaders and no edram for 100$ to notebook makers.

Actually MS paid AMD for the dev costs and AMD gets a cut on each chip made.

[vario]

You're making a lot of assumptions there.


Well yes the ps4 and xbox1 are both semicustom designs with similar building blocks, but by the looks of it the ps4 has 3 memory controllers while the xbox1 has 2. See chip diagrams below.
xbone: http://www.chipworks.com/en/technica...-the-xbox-one/
ps4: http://www.chipworks.com/en/technica...ore-processor/

So yes the building blocks are there, but those are totally custom architectures. The OS on each device will have maybe 10 things going in the background while a normal windows/linux box could have 100 and that might be enough to hide whatever latencies are in there. I don't know as I am no micro architecture engineer, but I don't think you can claim that just because it works on a console that it will work on a general purpose computer.


How certain are you about that? More cache can increase the latency on storage and retrieval as there's just more to search for a given bit of data. Maybe the cores are tuned to expect a given latency and more cache would throw them out of wack (technical term).
Here's a good read on jaguar's memory and cache architecture if you'd like to read up on it.
http://www.realworldtech.com/jaguar/6/


Not sure where's you're getting that. Maybe, but I don't see the evidence one way or the other.


In order of bolded text:
Sure if it works well; Some are yes; Big is a vague term. 20% might be big but if 20% = 3 fps then... not so much; That's the ticket right there, we don't know what "properly configured" means or costs; Might not actually be on the cheap if development costs are huge; And they sell it as a server cpu where threads are usually the performance metric (as opposed to memory bandwidth, IPC, or whatever); You seem to be arguing that they do and that that it's in the xbox1/ps4


Actually MS paid AMD for the dev costs and AMD gets a cut on each chip made.

When i was Saying "all the builidng blocks are there" i meant, AMD has it all figured out, no need to R&D whole new solutions.
There are couple of ways 8 core jaguar/puma could land on normal pc. Youre right that i dont know how much of a hassle it would be, you also dont .And from few solutions the one with 8 cores on regular DDR3(just 2133 like in xone) with no edram or gddr5 is pretty much ALMOST for sure easy as hell for them to do.
They are NOT TOTALLY custom architectures, amd has all the BLOCKS ready, and their partners choose the way they want it to be.Again simple double jaguar on normal DDR3 shouldnt be no problem.But of course there could be some more innovative ways to do that, dedicated gig of GDDR5 for example just for GFX ops.They HAVE GDDR5 controller for use ready...
This wave of consoles and specially xbox one, are just a pc, x86-64 pc, ms runs a derivative of win8 on it with a dx derivative tailored just for one pc configuration.Of course i CANT claim FOR SURE it will work, but there is no logical reason why it wouldnt, jaguar IS used on tablets pc's notebooks, everything, just half of the xbox one chip cpu wise.Jaguar cores used in consoles are basically the same as desktop ones! Just 2 "blocks" of them.
About the cache, there is 2M cache L2 for 4 cores ,so 4M for 8 cores or more probable 2x2M and each core has its own L1, and there is no reason why it would work any worse.
As for the OS`s , theyre not PS2/xbox(first one) , theyre doing multiple things at once , theyre pretty much at desktop PC levels, and while sony with its ONLY GDDR5 is certainly different and probably harder, XONE obviously was made with easier to program OS in mind ,its a PC now doing multiple things at the same time, thats the reason they went with DDR3 and had to resort to edram to boost GFX performance.Look at the xbox ONE die shot from the link you gave me, there are 2 distinct modules with each 4 cores and 2m of cache and ddr3 controller.
If you would chop off edram you would essentially have normal jaguar chip just with 2 modules glued together and more powerful GPU (same one tho! just more CU`s).
I read all the articles you gave, thanx.
As for the 100$ figure, i read an article with BOM`s , it said ms pays 110$ for each apu and sony pays 100$ (as its a lil smaller).But maybe it just round up cost of whole agreement divided per cpu.
As for R&D costs, there would be miniscule , because its already done chip wise, they would only need R&D needed for a cpu launch (but thats mostly done too on jaguar/puma launches).
In its simpliest form, it could be two separate dies from regular pumas on one pcb, just like in the olden days it was done.

[Olivon]

AMDee

Dis is good but give us moar enthusiats processorz plz.
U wont let Intel alone, no ?

[[XC] Lead Head]

See below.



@Vario There is a reason you see different architectures, layouts and designs when moving through market segments. It isn't because they are lazy or decided to waste resources but because that is what you need to do when specific markets require different target goals, whether performance, power, die size, or a specific mix of them all.

Think about the Pentium M - it started out as a low performance, low power single core CPU that ran at ~1.6 GHz, and eventually became the foundation for the Core, Core 2 and Core i-series CPUs. Jaguar has IPC easily exceeding that of Streamroller.

[raghu78]

the cat cores dont scale very well, and if you try to they become extremely difficult design, those are low powered cores designed from the start. steamroller is still more powerful performance wise, though you still have efficiency of that performance not as big as it should be. excavator should be a good performance chip, however i am sure the chip that comes after excavator arch will be a better high performance core that puts amd back on equal footing with intel.

Yeah. AMD is working on a new core microarchitecture for its high end big cores. There is a difference between the cat cores and the big cores. The design goals are different. There are lot of tradeoffs to achieve those design goals.You cannot scale a Jaguar or Puma+ to run at 4+ Ghz speeds without significant redesign of the circuits. In fact even the semiconductor process node is one of the design factors. You need a High performance process like Intel 22/14 FINFET or GF 28SHP or TSMC 16FF+ to run at 4+ Ghz clocks.

Excavator will improve efficiency over Kaveri but the problem is the Bulldozer core foundation is a extremely bad design. Its major weaknesses

1. Very poor single thread performance (given its die size)
2. Very poor FP performance
3. Very poor cache performance , especially L3 cache.
4. Very poor perf/sq mm and perf/watt.

Puma+ based Mullins and Beema are competing with Baytrail even with a process node disadvantage. This is what an efficient core can allow you to do. Bulldozer does the reverse. It widened the gap with Intel's high end core sandy. ivy and Haswell have done a good job maintaining that gap against Piledriver and Steamroller.

Jim Keller returned to AMD in Aug 2012 and in an interview last year said AMD is on track to catch up at the high end.

http://www.rage3d.com/articles/hardware/amd_worldcast/

"AMD are on track to catch up on high performance cores"

- Jim Keller, Corporate Vice President and Chief Architect of AMD's Microprocessor Cores

The best part of interview came in a nice little tidbit about core performance while discussing how much market and application awareness plays a role in core design. Many things are incremental, one of which is legacy performance on new designs. Jim confidently stated AMD are on track to catch up on high performance core, a function of design improvements. We couldn't pin down a timeline for this, but with a time scale of two years core design and one year build and test, it's not going to be immediate. My expectation is 2015.

My expectation is that the new high end core will be fabbed at TSMC 16FF+ or Samsung 14LPP and will come in late 2015 or early 2016.

[AliG]

Y
Jim Keller returned to AMD in Aug 2012 and in an interview last year said AMD is on track to catch up at the high end.

http://www.rage3d.com/articles/hardware/amd_worldcast/

"AMD are on track to catch up on high performance cores"

- Jim Keller, Corporate Vice President and Chief Architect of AMD's Microprocessor Cores

The best part of interview came in a nice little tidbit about core performance while discussing how much market and application awareness plays a role in core design. Many things are incremental, one of which is legacy performance on new designs. Jim confidently stated AMD are on track to catch up on high performance core, a function of design improvements. We couldn't pin down a timeline for this, but with a time scale of two years core design and one year build and test, it's not going to be immediate. My expectation is 2015.

My expectation is that the new high end core will be fabbed at TSMC 16FF+ or Samsung 14LPP and will come in late 2015 or early 2016.

I wouldn't trust anything out of AMD's corporate mouth just yet. They've said a lot of things that turned out to be blatantly untrue (i.e. the notorious "IPC increases" from JFAMD). We'll know soon enough whether they have a high end chip up their sleeve, though my money says they're putting their efforts towards mobile markets.

[LordEC911]

Think about the Pentium M - it started out as a low performance, low power single core CPU that ran at ~1.6 GHz, and eventually became the foundation for the Core, Core 2 and Core i-series CPUs. Jaguar has IPC easily exceeding that of Streamroller.

Yes, a very basic underlying foundation. Look at how vastly different each of those architectures are.
They didn't simply scale up Pentium M, like many are suggesting AMD should do with their kitty.

[raghu78]

I wouldn't trust anything out of AMD's corporate mouth just yet. They've said a lot of things that turned out to be blatantly untrue (i.e. the notorious "IPC increases" from JFAMD). We'll know soon enough whether they have a high end chip up their sleeve, though my money says they're putting their efforts towards mobile markets.

Jim Keller's reputation is outstanding. As the designer of the Athlon 64 and Apple Ax SOC including the Cyclone core, he has delivered two of the best non-Intel CPU cores of the the last decade. I don't believe he would put his credibility at stake by making false statements. JF-AMD was a marketing guy. what do you expect ? But from a industry veteran CPU designer who is AMD's Chief Architect of Microprocessor cores we can expect a measured response. So I am looking forward to what AMD has to offer after Excavator. AMD has shown with Puma+ and Mullins/Beema that they can compete wuith Intel even with a process node lag. Its a matter of time before they come up with an efficient high performance core.