AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread

So, the weak tessellation rates in Hawaii were indeed caused by hardware bugs?


Not necessarily. TR claims Tonga supports spilling tessellation parameters to L2 which could explain the better scaling on higher tessellation levels.

Well done on the color compression and bandwidth efficiency. I love seeing innovation that doesn't rely on brute force throughput increases.

Looks like AMD had everybody fooled with respect to Tonga's characteristics.
 
Mantle is counter-productive in Thief on Tonga: http://techreport.com/review/26997/amd-radeon-r9-285-graphics-card-reviewed/7

Probably just a driver bug/issue.

HardOCP:

Right off the bat we had some performance issues with the new MSI Radeon R9 285 GAMING OC video card in BF4 running under AMD Mantle. If you look back on the second page of this evaluation there is a disclaimer warning from AMD that the current version of Mantle may not work as intended on this new GPU.

At first, we did not know this information, so we began our evaluation under Mantle as we always do. We discovered abnormally low performance, performance much lower than GTX 760 running under Mantle with the R9 285. We reached out to AMD to see what was going on and were informed to try it in DX11 mode. Running the game in DX11 mode improved performance and brought it up to now be competitive with the GTX 760. This fixed the problem.

http://www.hardocp.com/article/2014/09/02/msi_radeon_r9_285_gaming_oc_video_card_review/5
 
Mantle is counter-productive in Thief on Tonga: http://techreport.com/review/26997/amd-radeon-r9-285-graphics-card-reviewed/7

Probably just a driver bug/issue.
I suspect it's more due to the card only having 2GB of memory. The 270X (with 2GB) showed exactly the same issues in earlier Mantle tests. As such it probably is more an issue how the games using Mantle do the memory management (I _guess_ they are more reliant on everything being in vram vs. when it is managed by wddm).
 
TR speculates that there may be a 384-bit bus hiding in there. It's be quite a departure from the past for AMD to do that for a SKU that's supposedly not massively below the X version in terms of performance. It'd so indicate that the 285 is either still very BW starved or that the 285X is very compute starved. I don't think that's likely, but we'll see.

One argument against the 384 wide bus, is the large increase in transistors. My theory is that they sacrificed some of the MCs to make room for the L2 and the color compression was sufficient to compensate the BW reduction.
 
Oh... so back to my original position then of "bad advert for Mantle"
Yeah, you'd think they'd have tried to show up with their best game, but no.

Since this is a Tahiti replacement, they probably have a +15% across the board performance increase in store 6 months from now. /s
 
can someone explain me the meaning of Tonga? from the point of view of performance / watt improves very little compared to Tahity pro, it's not a test chip for a new production process, it seems to be bigger than Tahity despite the smaller bus, so I assume that it doesn't costs less to produce, has less vram ....
 
It's a nice perf/$ improvement (for AMD.) It's not exactly a product to get all excited about for consumers, not ground braking in any way.

But it's interesting from a technical point of view and I love that it's 28nm so we can make apples-to-apples comparisons as soon as the full die is released. Though I doubt that that was a motivation for AMD to create it. :wink.
 
Very nice! I wonder if they also improved the depth compression not to require decompression before being used as a texture? Always seemed like a weird thing to do.
Oh I always thought only up to r5xx had that problem, but apparently (looking at the open source linux driver) you are right! I think from a hw pov this was not really weird, since the ROP caches were outside the ordinary cache hierarchy. Though yes it seems rather strange that you have this great bandwidth saving feature but you'll need a full read/write pass (and write without compression) for the whole depth buffer for shadowmaps for instance. Pretty sure if they can do it for color now they could also do it for depth - the slide only mentions "compressed data" which would probably indicate it's not limited to color.
Maybe that's some indication ROPs are now also using unified L2 cache? That was on my wishlist for next GCN iteration ;-). It would most likely also help increase bandwidth efficiency quite a bit, which this chip definitely has done (despite the rather low-clocked 256bit bus, overclocking memory doesn't really seem to do much if anything, at least not in the computerbase.de tests).
 
can someone explain me the meaning of Tonga? from the point of view of performance / watt improves very little compared to Tahity pro, it's not a test chip for a new production process, it seems to be bigger than Tahity despite the smaller bus, so I assume that it doesn't costs less to produce, has less vram ....
Less ram and memory bus width (less traces, less pcb layers) will make it somewhat cheaper to produce, despite the equally sized die.
 
It's a nice perf/$ improvement (for AMD.) It's not exactly a product to get all excited about for consumers, not ground braking in any way.

But it's interesting from a technical point of view and I love that it's 28nm so we can make apples-to-apples comparisons as soon as the full die is released. Though I doubt that that was a motivation for AMD to create it. :wink.

Less ram and memory bus width (less traces, less pcb layers) will make it somewhat cheaper to produce, despite the equally sized die.


so basically the only advantage is for Amd to recover the costs of Tahiti, and propose a product cheaper to produce, and therefore more profitable, in a segment of the market as important as that of the $ 250 one

from consumer point of view, who games at fullhd doesn't have advantages compared to 280x (50$ less but 1gb vram/128bit bus less too)


p.s.: excuse my english, i hope it's clear what i mean
 
From a consumer standpoint: higher performance, TrueAudio, better encode/decode, XDMA, latest IP, FreeSync capabilities, etc.
 
I suspect it's more due to the card only having 2GB of memory. The 270X (with 2GB) showed exactly the same issues in earlier Mantle tests. As such it probably is more an issue how the games using Mantle do the memory management (I _guess_ they are more reliant on everything being in vram vs. when it is managed by wddm).
Yeah, but they fixed this issue through combinations of game patches and driver fixes, so it shouldn't return just because they changed a few things in the graphics core.

Maybe it's the compatibility thing we've all being speculating about, which means the more future chips depart from the original GCN design, the more Mantle will have to be reconfigured.
 
Yeah, but they fixed this issue through combinations of game patches and driver fixes, so it shouldn't return just because they changed a few things in the graphics core.
There were some bugs that's true, but afaik it was never really fully fixed, nor do I think a driver update can fix this (I believe only the affected games could fix this).
E.g. you can see some BF4 numbers here:
http://www.tomshardware.com/reviews/amd-radeon-r9-285-tonga,3925-6.html
Both the 2GB 270X and 285 do significantly worse with Mantle, but the 3GB 280 does better (actually just barely, not even worth mentioning). It is true that the 285 seems to suffer even more than the 270X with Mantle but the trend is the same.
 
From a consumer standpoint: higher performance,
Yes, per mm2. But not all that spectacular compared to existing offerings. It's a replacement for 280X and the full version will slot a bit below the 290.

TrueAudio, better encode/decode, XDMA,latest IP, FreeSync capabilities, etc.
I understand that these kind of features are a marketer's dream to print on a box, but, let's face it, nobody is going to sell their 280X for one of those.

TrueAudio: a waste of space and effort IMHO.
Encode/decode: is this still a thing? I've long cancelled my DVD Netflix subscription in favor of streaming. Maybe still important in non-USA regions, but even there, aren't we in the territory of 32x vs. 40x speed CDROM drives?
XDMA: Yay! Finally something that has a benefit for gaming.
Latest IP: Wow, you really had to scrape the bottom of the barrel for that one.
FreeSync: Let's wait and see.
 
From a consumer standpoint: higher performance, TrueAudio, better encode/decode, XDMA, latest IP, FreeSync capabilities, etc.

in many games at fullhd it begins to be limited to 2 gb of ram, right up until yesterday when you had 3 in the same range (and 384bit bus for filters)

everything else you are referring to it's secondary,or rather does not represent real added value on the 280x, imho
 
I understand that these kind of features are a marketer's dream to print on a box, but, let's face it, nobody is going to sell their 280X for one of those.

Would you expect anyone to? 7870 (or equivalent <$300 card) is a different thought process though,

TrueAudio: a waste of space and effort IMHO.
Encode/decode: is this still a thing? I've long cancelled my DVD Netflix subscription in favor of streaming. Maybe still important in non-USA regions, but even there, aren't we in the territory of 32x vs. 40x speed CDROM drives?
XDMA: Yay! Finally something that has a benefit for gaming.
Latest IP: Wow, you really had to scrape the bottom of the barrel for that one.
FreeSync: Let's wait and see.
Lots of personal opinion there, mileage will vary dependant on the user.
 
Back
Top