AMD tech Day 2014 Kaveri slides

Lorien

Lorien

Supreme [H]ardness
Joined
Aug 19, 2004
Messages
5,197
Pretty much the whole presentation here: http://imgur.com/a/9AUko
A few select slides of interest.

uU2l7Gx.jpg
Tpp2YYV.jpg


spe152L.jpg
lkBq6bF.jpg


First slide is for the integrated R7 graphics performance second is with a R9 270X discrete GPU
AW2FI6F.jpg
vJmijm1.jpg
 
Last edited:
Mantle is up to 45% faster than DirectX. Nice!
Now is that 45% more frames or is it like the Oxide demo where it is so efficient that the CPU is idling without instructions to execute. An idle CPU doesn't translate into more frames per se. But it could mean more features could be added to the game to fill in for those lost CPU cycles.

Ouc6TcZ.jpg
 
Cannot wait to see Kaveri with a good set of 2400mhz memory and its onboard GPU.

Curious though I do not see any GDDR5 performance graphs....weren't they suppose to release that?
 
GDDR5 support was canceled a long time ago. That was a part of Kaveri/Steamroller v1. They canned that and started on Kaveri 2.0/SteamrollerB at some point in 2012.
 
NaroonGTX said:
GDDR5 support was canceled a long time ago. That was a part of Kaveri/Steamroller v1. They canned that and started on Kaveri 2.0/SteamrollerB at some point in 2012.
Click to expand...

Well that sucks arse.....I mean im still excited to see Mobile Kaveri. Currenty using a A10 in my laptop now, and let me tell ya, these APU's actually do damn good. Very shocked how well they run.

Bring on Kaveri!
 
maybe they gave PS4/Sony exclusive for GDDR5 and APU use so they could focus on mantle as if the cpu does not inherently need the massive amount of bandwidth, well then DDR3 should be more then good enough given how many shader cores it has to work with?
 
dragonstongue said:
maybe they gave PS4/Sony exclusive for GDDR5 and APU use so they could focus on mantle as if the cpu does not inherently need the massive amount of bandwidth, well then DDR3 should be more then good enough given how many shader cores it has to work with?
Click to expand...

from what I understand, Kaveri will still be memory bandwidth limited with ddr3 2400mhz memory.

Now im not saying it wont suck, but GDDR5 would of made the GPU aspect of the APU faster (thats how I understand it, but I could be wrong).
 
cageymaru said:
Mantle is up to 45% faster than DirectX. Nice!
Now is that 45% more frames or is it like the Oxide demo where it is so efficient that the CPU is idling without instructions to execute. An idle CPU doesn't translate into more frames per se. But it could mean more features could be added to the game to fill in for those lost CPU cycles.

Ouc6TcZ.jpg
Click to expand...

The boost for apu and Mantle is a good idea of what we will later find with other engines.
even though I dont expect such numbers off the bat.
laptop gaming however with apu can with Mantle be a big deal.
 
dragonstongue said:
maybe they gave PS4/Sony exclusive for GDDR5 and APU use so they could focus on mantle as if the cpu does not inherently need the massive amount of bandwidth, well then DDR3 should be more then good enough given how many shader cores it has to work with?
Click to expand...

I imagined something along these lines happened.
 
Lorien said:
...second is with a R9 270X discrete GPU
vJmijm1.jpg
Click to expand...

What I'm interested to see is the before and after of Mantle games on that list, since Mantle will allow for the APU's GPU to team up to that 270x with asynchronous Crossfire.
 
Looks like empty PR - in every game you can find places where performance depends mostly on GPU so official slides need to taken with huge grain of salt.

Also Mantle seems more and more like something that will help weak amd cpus lot more than it actually helps high end gpus.
 
That firestrike score puts it just under an 8850M.

Not too shabby at all.
 
Michaelius said:
Also Mantle seems more and more like something that will help weak amd cpus lot more than it actually helps high end gpus.
Click to expand...

DING DING DING!!!!! We have a winner!

Weaker GPU's benefit more from being coded at a lower level. No different from how the console GPU's [derived from a 7800 GTX and 2900XT] can still play BF4. APU's will see very strong gains, dGPU's...not so much.
 
Michaelius said:
Looks like empty PR - in every game you can find places where performance depends mostly on GPU so official slides need to taken with huge grain of salt.

Also Mantle seems more and more like something that will help weak amd cpus lot more than it actually helps high end gpus.
Click to expand...

I'm often wondering how you come to the conclusion you post.
Nearly every time you run your mouth about stuff you actually have no clue (AMD) about, I don't mind really it is a forum but the notion that you know what Mantle does exactly is somewhat improbable

If you however did want to know what Mantle exactly does you already would have noticed like 5 or 6 videos from APU13 talking about what it exactly does.

In nearly all of them it states:
GPU are faster then what is currently possible due to DX/OpenGL.
Mantle is low level API which solves the current batch problem.

And somehow this escapes you , no where does it say you need a high end cpu or that a high end cpu would not work or not benefit throughput when using Mantle

http://hardforum.com/showpost.php?p=1040475620&postcount=10
http://hardforum.com/showpost.php?p=1040477550&postcount=23

And still you post things like this?
 
Last edited:
Everyone already knows that Mantle scales up with the amount of CPU power and cores available for it to use. He's trolling like usual. Don't feed the trolls. :)
 
Pieter3dnow said:
I'm often wondering how you come to the conclusion you post.
Nearly every time you run your mouth about stuff you actually have no clue (AMD) about, I don't mind really it is a forum but the notion that you know what Mantle does exactly is somewhat improbable

If you however did want to know what Mantle exactly does you already would have noticed like 5 or 6 videos from APU13 talking about what it exactly does.

In nearly all of them it states:
GPU are faster then what is currently possible due to DX/OpenGL.
Mantle is low level API which solves the current batch problem.

And somehow this escapes you , no where does it say you need a high end cpu or that a high end cpu would not work or not benefit throughput when using Mantle

http://hardforum.com/showpost.php?p=1040475620&postcount=10
http://hardforum.com/showpost.php?p=1040477550&postcount=23

And still you post things like this?
Click to expand...

All you need is basic understanding of cpu and gpu bottlenecks and basic rules of company PR.

Mantle isn't some kind of cure it all drug that will help in every situation. It will help in certain scenarios and since Direct X inefficiencies can be brute forced through usage of faster CPUs then by basic logic the biggest benefits will be observed on systems with slower cpus.

Now since the job of PR department is making sure you see their product in best possible light then we can safely assume 45% is best possible scenario they could find.

cageymaru said:
Everyone already knows that Mantle scales up with the amount of CPU power and cores available for it to use. He's trolling like usual. Don't feed the trolls. :)
Click to expand...

And then you run into GPU bottleneck
 
Last edited:
Michaelius said:
All you need is basic understanding of cpu and gpu bottlenecks and basic rules of company PR.

Mantle isn't some kind of cure it all drug that will help in every situation. It will help in certain scenarios and since Direct X inefficiencies can be brute forced through usage of faster CPUs then by basic logic the biggest benefits will be observed on systems with slower cpus.

Now since the job of PR department is making sure you see their product in best possible light then we can safely assume 45% is best possible scenario they could find.



And then you run into GPU bottleneck
Click to expand...

When an 8 year old powerpc chip can perform better than a current i7 chip in this "bottleneck situation, I'd say the brute force method is simply not working, not all design problems can be fixed with more power. Some things simply don't scale well with Mhz.
 
Michaelius said:
Mantle isn't some kind of cure it all drug that will help in every situation.
Click to expand...
Assuming there are no glaring design defects, and assuming a certain level of developer competence, Mantle should help in almost every situation. The degree to which it's going to help is going to vary wildly based on the system, the rendering paradigm and the quality of implementation.
 
Michaelius said:
All you need is basic understanding of cpu and gpu bottlenecks and basic rules of company PR.

Mantle isn't some kind of cure it all drug that will help in every situation. It will help in certain scenarios and since Direct X inefficiencies can be brute forced through usage of faster CPUs then by basic logic the biggest benefits will be observed on systems with slower cpus.

Now since the job of PR department is making sure you see their product in best possible light then we can safely assume 45% is best possible scenario they could find.



And then you run into GPU bottleneck
Click to expand...

I think you need to watch/read more of the mantle technical dive's.
 
This (Mantle) is an interesting way to get gains after a plateau in general processor performance at set TDPs. Both with AMD (Trinity>Richland) and Intel (SB>IB>Haswell) remaining generally static in overall performance despite IPC improvements within generations and revisions. This just reaffirms to me that general computing has been in a period of both a sweet spot (a quadcore is great just about at every price point) AND stagnation (a quadcore bought in 2009 is going to be about the same price for the same performance today)... very surreal to see diminishing returns hit that brick wall so soon.
 
Mantle can scale with CPU, GPU, or both. It is up to the dev to decide how, when, and where to use it.

It seems to scale very well with something like Kaveri to your eyes only cause it is a cpu and gpu in one so compared to say DX it does very well, but with the high end discrete gpu it scales that much better due to them being more powerful, they have shown demos in this context where they can pretty much shut the cpu off as it really is not needed for the game to function correctly.

So I think the high end gpu have that advantage maybe just with the sheer amount of shaders or whatever they get the horsepower advantage and due to this, also specific advantages at this point that APU will not.
 
maboblivion said:
any hybrid graphics slide out there? (igpu+gpu)
Click to expand...

Probably no different than current hybrid cross-fire.

Probably a hell of a lot smoother due to mixing GCN/GCN cards, vs richland VLIW with GCN.

And probably only truly useful with mantle or perfectly matched dgpu (512spu/ddr3). Otherwise you get the typical crap of async-xfire.
 
With mantle, I'm pretty sure you can look at the whole of GCN cores and allocate tasks to 'partitions' of cores. Thus if the code can recognize the available GCN cores on the iGPU, than it can utilize the power asymmetrically, like the iGPU handling post-processing or physics.
 
From here: https://forums.robertsspaceindustri...eveloper-summit-on-line-event-mantle-api-info


Interesting, mantle will finally allow asymmetric use of GPUs, allowing to combine APUs with whatever dedicated GPUs you have. Mantle drivers will simply give access to all available GPU resources and developer can use them as they please. Lots of power, lots of control, lots of responsibility.

The memory stuff sounds exciting as well, the ability to reuse memory as you please, essential just giving an object a pointer instead of creating object and allocating each time new memory should be fantastic if properly used. With mantle you have to ability and the responsibility of a driver developer. Though iirc they mention as well that they have an abstraction layer in between the game developer and the real memory. Each application has to stay within its own memory block and has no access to memory of a different (GPU) applications.

Oh btw, this means that if you are using 2x 290 that your mantle games have access to 8GB of ram and not 4 like in crossfire or SLI. You don´t have to duplicate everything.



I did not know about that last statement I put in bold. So each card adds more memory to the equation, and you can mix and match cards - 290x + 270 + APU + 280. That's a lot of cores...
 
That makes it sound like it's HUMA'ing the vram of the dgpu's.

Which doesn't seem right.

But would be cool to have 16gb ddr3 + 4-8gb GDDR5 in the same memory pool.
 
Nice slides. If some manufacturer design a board with GDDR5 integrated memory that will be nice for kaveri.
 
Twingo said:
Nice slides. If some manufacturer design a board with GDDR5 integrated memory that will be nice for kaveri.
Click to expand...
There is no GDDR5 memory controller on Kaveri. :p The source of that rumor was BSN (heavy on the first 2 letters) and was simply incorrect.
 
no GDDR5 support. only DDR3 were being used here in the platform lab
 
CombatChrisNC said:
Oh btw, this means that if you are using 2x 290 that your mantle games have access to 8GB of ram and not 4 like in crossfire or SLI. You don´t have to duplicate everything.



I did not know about that last statement I put in bold. So each card adds more memory to the equation, and you can mix and match cards - 290x + 270 + APU + 280. That's a lot of cores...
Click to expand...

Compute wise, yeah but if you are doing CF/SLI and they are rendering the (almost) same exact scene, using AFR, then they will need the same textures loaded... so essentially nullifying that benefit.
 
I wonder if an FM2+ board could be DDR4 compatible if the next gen aftter Kaveri gets a DDR4 controller.

That is, are they electrically compatible (only differing in protocol) or does DDR4 require new sockets and new leads to the CPU socket...
 
It really is good to see, I do hope AMD come back they have been lagging behind for far too long.
 
omega-x said:
I fear that won't last long once intel drops 14nm
Click to expand...

Is that related to Jesus Christ our saviour because with all the new things AMD comes with I'm hearing from people it will somehow be not so relevant because of 14nm?
 
Last edited:
Darakian said:
Maybe not, but that's nearly a year out.
Click to expand...
As widely reported about 2 months ago, volume Broadwell production was delayed by a quarter from Q4'13 to Q1'14, according to Intel's CEO. There have been no further announcements of delays. That suggests a launch in late Q1 or the middle of Q2 (typical for May releases). Not sure what bathroom walls the other rumors are coming from. :p It was already demonstrated at the last IDF in September.

Intel CEO Brian Krzanich said during his company's third-quarter earnings call that Broadwell yields are now back on track but an extended "defect density issue" with the 14nm process node has pushed volume production out for another quarter.

"While we are comfortable with where we are at with yields, from a timing standpoint, we are about a quarter behind our projections. As a result, we are now planning to begin production in the first quarter of next year," Krzanich said.

He later elaborated on the specific issues affecting Broadwell production, as recorded in a transcript of the call by Seeking Alpha:

"It was simply a defect density issue ... As we develop these technologies, what you do is you are continually improving the defect densities and those result in the yield, the number of dies per wafer that you get out of the product. And what happens as you insert a set of fixes in groups, you will put four or five, maybe sometimes six or seven fixes into a process and group it together, and run it through, and you expect an improvement rate.

"Occasionally as you go through that, the fixes don't deliver all of the improvements you thought. We had one of those. So why do I have confidence now? Because we've gone back now and added additional fixes, gotten back onto that curve. So we have confidence that the problem is fixed, because we have actually data that it's fixed. And so that gives us the confidence that we are going to keep moving forward now. And that happens sometimes in these development phases like this, so that's why we are moving [volume production] out a quarter."
Click to expand...
tl;dr tweaks have already been made to improve the yield at that stage in production to get back to the expected

Anyways, Iris Pro graphics are very rarely used by manufacturers and it's not really a good place to point out Intel's GPU performance. Almost everything ships with HD 4400 or HD 4600 (GT3 HD 5000 & 5100 are pretty rare) as the top integrated GPU for Intel CPUs and it's hard to say how much Broadwell will improve performance on mainstream processors. AMD seems willing to dedicate over half the CPU die to a GPU, so I wouldn't expect most Intel parts to challenge the top end Kaveri GPUs in gaming performance. Computational performance though may tip in Intel's favor due to much better cache subsystems.
 
Last edited:
That's cool and all, but in production != in channel. I look forward to those chips to but HSA and kaveri have my attention for the next 6-8 months. Take a look at intels roadmaps and you don't see broadwell until after summer.
 
You must log in or register to reply here.
Back
Top