shader-based interpolation perhaps
AMD: R8xx Speculation
- Thread starter Shtal
- Start date
shader-based interpolation perhaps
That certainly must be part of it, but surely the 5830's 31% shader power advantage more than offsets it.
leoneazzurro
Regular
It is possible that halving the ROPs has some impact in the way the driver should manage them?
That XFX PCB is smaler than the 5850 by more than inch if u compare the PCI-E lenght with the rest of the card after it. I would say its more like 4850
.Its actualy a 5750 PCB :
Last edited by a moderator:
That XFX PCB is smaler than the 5850 by more than inch if u compare the PCI-E lenght with the rest of the card after it. I would say its more like 4850
Its actualy a 5750 PCB :
Maybe you should read more carefully then, since I said 5750, not 5850
PSU-failure
Newcomer
Lower count, but with a wider bus per unit (less efficient memory accesses?) and more SIMDs to collect results from.
I don't know how to classify this board...
As a test to see what balance is best? (but hey, no need to ship it then)
Or as a too much castrated chip just to not cannibalize sales of the 5850/5870? (makes more sense since what really seems to be killing performance is the halved ROP count)
By the way, was this chip/board *needed* or is it just a reference for partners to fill a gap in their price lists?
I think that one is a mock up. If you look at the review that on the Chinese site that went earlier the XFX board is more inline with the others.
What really sucks is a few months ago you could have gotten a 5850 for just $20 bucks more than this card.
http://www.newegg.com/Product/Produ...2879&cm_re=radeon_5830-_-14-102-879-_-Product
$265 5830
http://www.newegg.com/Product/Produ...1184&cm_re=radeon_5850-_-14-131-184-_-Product
$300 5850 with dirt 2 included.
That makes me laugh
Good news is that they seem to overclock well
http://firingsquad.com/hardware/ati_radeon_hd_5830_performance/page11.asp
It hit 982/1225 but considering it was already using 15watts more on load than a 5850 i'm not sure how good that is even if the higher clocks let it close the gap between itself and the 5850
http://www.newegg.com/Product/Produ...2879&cm_re=radeon_5830-_-14-102-879-_-Product
$265 5830
http://www.newegg.com/Product/Produ...1184&cm_re=radeon_5850-_-14-131-184-_-Product
$300 5850 with dirt 2 included.
That makes me laugh
Good news is that they seem to overclock well
http://firingsquad.com/hardware/ati_radeon_hd_5830_performance/page11.asp
It hit 982/1225 but considering it was already using 15watts more on load than a 5850 i'm not sure how good that is even if the higher clocks let it close the gap between itself and the 5850
Last edited by a moderator:
It seemed to make them a little more consistant, I.E 5830 > 5770 for the most part.
So Damien and Marc from Hardware.fr ran a few tests on the HD 5830, and I think they figured out the problem:
http://www.hardware.fr/articles/783-3/preview-radeon-hd-5830.html
Now it makes sense!
http://www.hardware.fr/articles/783-3/preview-radeon-hd-5830.html
Now it makes sense!
If the four L2 caches are still there could it be that they dont have 4*ROP quads but 8* dual ROPs.
So it would need 2 clocks to write a 4x FP16 instead of 1. The fillrate would be same but the write speed would be 2 times slower in those pure theoretical tests depending on bandwith.
Without AA the 5830 is clearly faster than 4890.With this they could save the L2 cache and have 256bit bus instead of half the cache and 4*ROP quads with 128bit bus. While they still maintain 256bit reads from ram.
Very interesting. Indeed those results would suggest rops are bandwidth-starved in HD 5830.
Too bad the diagrams don't really indicate how rops are connected to memory controllers / L2 cache - the pic (for instance from here, http://techreport.com/articles.x/17618) would suggest they are fully independent from both L2 and memory controllers but there might be some simplification going on...
Didn't Dave say that due fullspeed L2, clocked higher, 5830 can even end up being a bit faster on same scenarios, when compared to 5850?
Jawed
Legend
L2s are for texels only.
Intriguing idea. Need to ruminate on that, currently sceptical...
Jawed
Texture cache!=Colour(&Z) cache. Dave was talking about the texture cache, which is nice and all but doesn't indicate what's up with the ROPs. And I'd personally lean towards agreeing with mczak here, the important clues are probably hidden in the RBE-Colour/Z Cache-Memory Controller triad.
Jawed
Legend
800 versus 850MHz. Not that that explains the deficit, mind.Really strange -- in blending op's 5830 is virtually on par with its 128-bit cousin, the 5770, and fill-rates are even worse.
Jawed
Similar threads
