Kaveri demoed, running Devil May Cry,FM2 compat.?

[mrmt]

There was a AMD GPU prototype shown a few years which had stacked DRAM(Amkor),and it is supposedly in the PS4. It will be interesting to see if one of the AMD APUs will end up with similar technology.

Charlie quoted that stacked DRAM was prototyped for Kaveri but was canned last year.

 

[R0H1T]

So what will the revised likely performance opinions for Kaveri be if GDDR5 is dropped?

Does this also suggest the product will be late?

I know some people who seemed to think Kaveri was going to be approaching 7850 performance, whereas I always doubted it would match HD5770 or HD7750 performance.

Depends on how well the memory controller is & to an extent its support for memory modules operating at 2400 MHz or above. The performance scaled nicely with faster mem on trinity, even better for richland, so with GCN & support for RAM operating at high speeds we can expect 2~4x the performance of trinity for the IGP part, again depending on how HSA pans out for them !

 

[CHADBOGA]

so with GCN & support for RAM operating at high speeds we can expect 2~4x the performance of trinity for the IGP part

AMD can't get those sorts of increases on their discrete cards with a new architecture and a full die shrink.

How the hell is it going to be 2~4x the performance of Trinity?

That is crazy talk.

 

[R0H1T]

AMD can't get those sorts of increases on their discrete cards with a new architecture and a full die shrink.

How the hell is it going to be 2~4x the performance of Trinity?

That is crazy talk.

For IGP it certainly isn't, remember VLIW4(or was it VLIW5 ) to GCN & trinity gains are limited only upto 1866 MHz IIRC whilst I'm expecting Kaveri to support atleast 2400 MHz(or above) rated memory. Beyond this is of course HSA which should bring some amount of tangible gains & then there is the 28nm half node shrink, four points that I can list right now so within reach I'd say in terms of gains at the top end !

 

[frozentundra123456]

For IGP it certainly isn't, remember VLIW4(or was it VLIW5 ) to GCN & trinity gains are limited only upto 1866 MHz IIRC whilst I'm expecting Kaveri to support atleast 2400 MHz(or above) rated memory. Beyond this is of course HSA which should bring some amount of tangible gains & then there is the 28nm half node shrink, four points that I can list right now so within reach I'd say in terms of gains at the top end !

So now instead of DDR5 it is HSA that is going to make Kaveri magically outperform a discrete card with the same number of SP? OK........whatever you say.

 

[R0H1T]

So now instead of DDR5 it is HSA that is going to make Kaveri magically outperform a discrete card with the same number of SP? OK........whatever you say.

Did I say that, stop inferring stuff that was never stated in the first place !

 

[frozentundra123456]

Did I say that, stop inferring stuff that was never stated in the first place !

You said that Kaveri would be 2 to 4 times the performance of Trinity. I believe that 2x would put it pretty much equal to 7750 and for sure 4x would put it faster than 7750. So yes, you did indirectly say that.

 

[R0H1T]

You said that Kaveri would be 2 to 4 times the performance of Trinity.

Yes, Trinity with RAM at 1866 MHz vs Kaveri with RAM at 2400 MHz is certainly what I'm waiting for. I won't claim that AMD won't botch this first ever desktop implementation of HSA but seeing as how they've had certain experience with the consoles so this time I'm cautiously optimistic, again hoping that the BD disaster will not be repeated !

I believe that 2x would put it pretty much equal to 7750 and for sure 4x would put it faster than 7750. So yes, you did indirectly say that.

Yes & it depends on the particular usage scenario, like mining bitcoins, so unless you're stating that trinity IGP is half as slow as 7750 in every way possible I'd say that Kaveri -> Trinity performance could well be in that 2~4x range. Lastly I'm looking forward to HSA, like many others, also the Kaveri IGP should benefit from it & could be at par with HD 7750 or HD 7770 in certain cases, mostly after seeing how GCN based 7790 performed.

 

[AtenRa]

For the latest DX-11 titles, Kaveri (2400MHz memory) could be up to 2x faster than Trinity(1866MHz).

 

[Vesku]

Here is a review of a 7750 DDR3 card, it also has 6670 and 6570 DDR3 cards compared. So if someone wants to go through a bit of effort they can probably get a decent estimate of a 512GCN APU.

http://ht4u.net/reviews/2012/msi_radeon_hd_7750_2gb_ddr3_test/index15.php

Find ratio of Trinity to 6670/6570 DDR3 (whichever best matches the amount of shaders) then apply that ratio to 7750 DDR3 to get an idea of how Kaveri might do. HSA isn't going to add too much, especially not games developed before the HSA standards organization was even around.

 

[Enigmoid]

Yes, Trinity with RAM at 1866 MHz vs Kaveri with RAM at 2400 MHz is certainly what I'm waiting for. I won't claim that AMD won't botch this first ever desktop implementation of HSA but seeing as how they've had certain experience with the consoles so this time I'm cautiously optimistic, again hoping that the BD disaster will not be repeated !Yes & it depends on the particular usage scenario, like mining bitcoins, so unless you're stating that trinity IGP is half as slow as 7750 in every way possible I'd say that Kaveri -> Trinity performance could well be in that 2~4x range. Lastly I'm looking forward to HSA, like many others, also the Kaveri IGP should benefit from it & could be at par with HD 7750 or HD 7770 in certain cases, mostly after seeing how GCN based 7790 performed.

HSA advantage = 0 unless the application is designed to take advantage of it.

Kaveri will not be 4x greater than trinity.

(high settings)

It would have to be pretty much 7790 territory to do that.

 

[Enigmoid]

Here is a review of a 7750 DDR3 card, it also has 6670 and 6570 DDR3 cards compared. So if someone wants to go through a bit of effort they can probably get a decent estimate of a 512GCN APU.

http://ht4u.net/reviews/2012/msi_radeon_hd_7750_2gb_ddr3_test/index15.php

Find ratio of Trinity to 6670/6570 DDR3 (whichever best matches the amount of shaders) then apply that ratio to 7750 DDR3 to get an idea of how Kaveri might do. HSA isn't going to add too much, especially not games developed before the HSA standards organization was even around.

It seems to be about 7% below a 640.

So using this chart

Using the 640 (don't know if the 6670 is GDDR5 or DDR3) it looks like 201*0.93 = 187. 187/127 = 47% over trinity. It seems to be using 33 watts too.

 

[Erenhardt]

HSA advantage = 0 unless the application is designed to take advantage of it.

If both consoles have it, than developers will use it. Not from day 1, but they will, sooner rather than later. Games are mostly console ports. Having advantages of console's arch can be hugely beneficial.

 

[AtenRa]

Here is a review of a 7750 DDR3 card, it also has 6670 and 6570 DDR3 cards compared. So if someone wants to go through a bit of effort they can probably get a decent estimate of a 512GCN APU.

http://ht4u.net/reviews/2012/msi_radeon_hd_7750_2gb_ddr3_test/index15.php

Find ratio of Trinity to 6670/6570 DDR3 (whichever best matches the amount of shaders) then apply that ratio to 7750 DDR3 to get an idea of how Kaveri might do. HSA isn't going to add too much, especially not games developed before the HSA standards organization was even around.

That review is useless for evaluating the HD7750 DDR3, the reason is that they test at high resolutions/settings and using AA Filters that makes the DDR-3 cards completely memory bandwidth starved.

 

[monstercameron]

If both consoles have it, than developers will use it. Not from day 1, but they will, sooner rather than later. Games are mostly console ports. Having advantages of console's arch can be hugely beneficial.

why wouldn't they use it from day one? isn't the main advantage that they dont have to keep redundantly copying data from cpu to gpu memory, especially when it is shared?

also could software trick the program into thinking already copied the data between the two?

 

[Vesku]

It's not as useful as I'd like but they're the only ones who published actual fps for the 7750 DDR3 version. So unless someone buys one and posts the info they're the only numbers available. Since the DDR3 6xxx series cards are hit by the same settings the ratio of the two should be a rough approximation. Then if there is a comparison of A10-5800K vs 6670/6570 DDR3 can apply the ratio and get a very rough idea whether we are looking at +20-30% or a good deal higher.

 

[Obsoleet]

Highly doubt it.

A 128 bit interface using 900 mhz DDR3 gets around 28GB/sec. 2133 mhz ram gets around 34 gb/sec but has to share that with the cpu. You are not looking at 'blow out of the water'.

5870 to a10 is probably going to be a downgrade.

It would be a downgrade.

I have a 5870 since launch day, love it. And have been waiting for APUs to catch up to this before replacing my system. Even the Kaveri A10 will match this. Better just use the APU for a backup or lightweight gaming system. I'm buying the CPU with the best GPU power for that reason.

CPU power is overpowered, even on Bulldozer (which is faster than what I'm using now), but an integrated APU is the way to go. AMD is winning.

 

[Enigmoid]

That review is useless for evaluating the HD7750 DDR3, the reason is that they test at high resolutions/settings and using AA Filters that makes the DDR-3 cards completely memory bandwidth starved.

Depends really. According to their charts the 650 is 53% better than the 640. According to Toms the 650 is 46% better than the 640 at 1080p low (which is easily playable with the 7750 GDDR5 or 650).

The 650 is clocked at 1058 vs 900 for the 640. Normalizing the clocks the 650 is 28% faster due to the GDDR5. If we use the scaling of +GDDR5 = 28% faster (which we really shouldn't since AMD seems more sensitive to Bandwidth than Nvidia; in that 'useless' review the 650 is 32% stronger than the 640 clock for clock while the 7750 DDR5 is 53% stronger than the DDR3 7750) the the kaveri version of an apu will be 218% or 218/127 = 72% faster clock for clock. However, again looking at the fact the AMD seems to need more bandwidth I doubt we will see that kind of scaling.

My bet is around 50% at most using DDR3 (because its shared with the CPU and higher performance = higher settings, resolution and AA where performance will drop more).

 

[galego]

why wouldn't they use it from day one? isn't the main advantage that they dont have to keep redundantly copying data from cpu to gpu memory, especially when it is shared?

It seems you confound HSA with hUMA.

 

[galego]

HSA advantage = 0 unless the application is designed to take advantage of it.

Sure, but the next consoles have HSA and games will use it.

 

[Enigmoid]

Sure, but the next consoles have HSA and games will use it.

Except that currently intel and AMD are the most equal they have ever been in mobile and on the desktop amd is a scant ~30-35% better than intel (one generation). Intel supports similar features on their igp (and almost every single intel consumer chip ships with an igp) and already have intel exclusive features in a game (grid 2 cloud/dust-pixelsync).

And PC exclusives may not use HSA/HUMA (just like many pc exclusive games, SC2, Dota, LOL, Gw2, tend to still be highly singlethreaded).

And the fact the AMD's top CPUs don't have any igp (won't help FX series one lick).

And that intel seems to be doing very well in compute.

And the fact that the consoles won't be releasing for a while and games that take advantage of HSA/HUMA will take even longer to be released.

 

[galego]

Except that currently intel and AMD are the most equal they have ever been in mobile and on the desktop amd is a scant ~30-35% better than intel (one generation). Intel supports similar features on their igp (and almost every single intel consumer chip ships with an igp) and already have intel exclusive features in a game (grid 2 cloud/dust-pixelsync).

And PC exclusives may not use HSA/HUMA (just like many pc exclusive games, SC2, Dota, LOL, Gw2, tend to still be highly singlethreaded).

And the fact the AMD's top CPUs don't have any igp (won't help FX series one lick).

And that intel seems to be doing very well in compute.

And the fact that the consoles won't be releasing for a while and games that take advantage of HSA/HUMA will take even longer to be released.

HSA/hUMA is already being used in games under development. E.g. killzone shadow demo already uses a poll of shared memory and CUs for physics compute. I don't wait this to be maintained in first PC ports because there are not PCs with HSA/hUMA now. But once developers are familiar with the technology and use it in games they could reuse it in future PC ports when chips as kabini are ready.

We know that AMD will be substituting the FX line by APUs, we don't know exactly when and how. Why do you believe that Steamroller cores will be appearing first in Kaveri and not in FX line? Because the APU line is more important for AMD that the 'traditional' FX line.

 

[Enigmoid]

HSA/hUMA is already being used in games under development. E.g. killzone shadow demo already uses a poll of shared memory and CUs for physics compute. I don't wait this to be maintained in first PC ports because there are not PCs with HSA/hUMA now. But once developers are familiar with the technology and use it in games they could reuse it in future PC ports when chips as kabini are ready.

We know that AMD will be substituting the FX line by APUs, we don't know exactly when and how. Why do you believe that Steamroller cores will be appearing first in Kaveri and not in FX line? Because the APU line is more important for AMD that the 'traditional' FX line.

??????

Killzone is an exclusive so how exactly is it relevant to PC ports?

Kabini is way too weak to play any xbox 360/ps3 port very well (768p low 35fps avg) so a ps4 port is going to really suck

Doubt AMD will get rid of their FX line. Their apu cpu section is simply too weak to really stand alone (compared to 8350 or i5/i7).

Kaveri is probably a smaller chip (easier to manufacture) similar to how the 8350 is bigger than the a10-5800k. It can also be used for mobile as well a desktop and a greater portion of amd sales are probably from APU vs CPU only.

 

[inf64]

I think we will have FX migrated to FM2(+) platform next year. Of course FX will lack iGPU on-die (except maybe 3M Kaveri which now disappeared from AMD's internal manuals and roadmap for 2013).

 

[Abwx]

And PC exclusives may not use HSA/HUMA (just like many pc exclusive games, SC2, Dota, LOL, Gw2, tend to still be highly singlethreaded).


And the fact that the consoles won't be releasing for a while and games that take advantage of HSA/HUMA will take even longer to be released.

You re making some confusion on theses features that works
at completely different levels.

HSA is software implementation dependent while
hUMA is hardware implemented in SR APU , as such
hUMA is transparent to the software and will work
whatever the code.