Printable View
This has been the case since early P35 boards ASUS released the Blitz with a crosslinx chip which made about a 13% difference and then x38 came out and made it redundant. Sadly the P45 chipset has kept the 16x + 4x or worse moved to an 8x + 8x layout.Quote:
Originally Posted by m0da
Wow. So crossfire basically requires that you have an X38 or X48 chipset. Hmm. Still better than requiring an Nvidia chipset for SLI but it is a bit limiting. I didn't realize that the P35/P45 boards were only X8. I guess I've been over with the little green men so long I haven't been paying attention.
I'm quite sure 2 8x slots are much beter than 16x + 4xQuote:
Originally Posted by initialised
Post-processing will basically require such. HDR and blur affects are usally added after the frame has been completely rendered. If they had linked frambuffer, each gpu writing to mem, this is no big deal, but when having one main gpu with memorycontrol, and the rest as slaves to the first, post processing would take place entrirely on first gpu, or be very slow due to latency(final frame would be rendered to mem, then processed, would make no sense to pass it elsewhere).Quote:
Originally Posted by hurleybird
However, IMC design is something that AMD excels at, so who knows. Designing a crossbar and scheduler and using ringstop links to each gpu would work wonders, I suppose, and would offer what you are after, but again, the pcb design is what holds that back. the cost would be easily more than double two single cards...and possibly $100 more than two cards...killing the $499 launch price.
ANyway, I've thought of all of this long ago, as has AMD/ATI. It might makes sense to try it out on a gpu who's back end is pretty tight, but I don't see it possible until another gen or 2. 2 gpu's aren't enough to see a real benefit, but 4 or more starts looking good...basically the same type of computing offered by the Cell processor would be possible. one Main gpu managing all teh work and doing post processing, one gpu for vertex processing, one for shaders, one for geometry.
:shrug:
The P45s use 8x8x but it's PCI-E 2.0, so it's the same as if they were 16x16x PCI-E 1.0. It won't make any difference, performance-wise.Quote:
Originally Posted by gojirasan
I'm doing crossfire on a i975X. It has 8x8x PCI-E 1.0 and there's no performance drop.
The main problem in using the P35s and P965s for crossfire wasn't even the speed of the 4x connector, but the fact that it was connected to the southbridge instead of the northbridge (like in i975, P45, X35, 770X, etc) so there was significantly more latency.
In that case all of the texture memory can still be shared without worries. The "slave GPU" just needs to be given exclusive write access to a small portion of memory where it can do post processing on the frame it's rendering :up:Quote:
Originally Posted by cadaveca
Ok, but the clamshell mode is two-way communication, not one-way exclusive. In the end two gpus would still need to access the same ram, and this is where the problem lies. it also envolves a memory bus in that instance that would not be 256-bit(not enough, for sure), again, making pcb overly complex. it would only make real sense if multi-gpu configs were on the same substrate, a la xenos. OR custom-desgined GDDR5...which would push prices much higher than they are.
Now you could make the memory control 512-bit, with 256-bit exclusively shared...but again, the pcb complexity is what nixes any such ideas. This would probably push ATI cards to nV pricing levels.
In the end, it's price which tells us that this is not the case this gen. Yes, it's more than technically feasible, but financials is something else!
Quote:
Originally Posted by ORBR
I honestly believe that's some kind of driver bug or software issue. 8x8x PCI-E 2.0 is more than enough for crossfire.
http://www.bit-tech.net/hardware/200...has-arrived/13
http://www.elitebastards.com/cms/ind...1&limitstart=6
I wouldn't be surprised, especially with how successful the 4850 crossfire is, it'll be faster than the 280 to the leastQuote:
Originally Posted by Xello
I concur, 8x pci_e 2.0 is like 16x pci_e 1.1, I doubt any card on the market (including the gtx 280) uses all the bandwidth the pci_e lanes provide. Could be something with that p45 board specificly or perhaps has something to do in specific with the p45 nb or IC10, you never knowQuote:
Originally Posted by ToTTenTranz
Hey look what's around:
ExtremeTech RV770 Architecture Analysis
Wow, they did a LOT of changes to the architecture to fit all that within 256mm^2. 956M transistors in 256mm^2 makes me wonder why Nvidia's 1.4B seems to inefficient. Anyways, some juicy parts (and pics)
http://img229.imageshack.us/img229/2...archlx4.th.gif
http://img247.imageshack.us/img247/6...llervu5.th.gifQuote:
This may be our first hint at what is new with the dual-chip "R700" product coming in a few months—these hubs may communicate between two RV770 chips in a fashion that is more efficient than in past multi-GPU boards.
Maybe all that talk about ATI really employing wizards is true.Quote:
Last but not least, we have the render back-ends, commonly called ROPs. There are still 16 ROPs in the RV770 chip, but ATI has focused on improving their performance degradation when using anti-aliasing—another real sore spot of the RV670 chip. Depth/stencil ops are doubled per ROP, which effectively doubles the fill rate with 2x or 4x AA to a full 16 pixels per clock (up from 8 in the RV670). 8x AA is 8 pixels per clock, again double the RV670. All 16-bit per component color modes have doubled AA resolve, and depth/stencil only rendering passes are of course doubled as well. The RV770 ROP is now closer in capabilities to that of Nvidia's G92 and GTX 200 GPUs.
Of course, with all these additional texture units and stream processors, ATI had to roughly double the intra-chip communication bandwidth to make sure it doesn't all get bottlenecked. It's really rather impressive that the company was able to accomplish all these architectural changes and fit so many more texture units, stream processors, and improved ROPs in a chip that is still on a 55nm process but manages to be only around 33% bigger, and still manages to be over 20% smaller than the G92 GPU.
800SP's, 40 TMUs, 16 ROPS (at twice the clocks of before), and a new hub
Now I wonder what the new hub does for R700... interesting
Do u remember the time when NH showed the slide which said RV770 has 800SPs and some great users here proofed it wrong (of cause a fake) and the knew it all better cause RV770 will have 480SPs + a shaderdomaine? Sometimes i think back and smile :D
eg:
http://www.xtremesystems.org/forums/...postcount=1524
and
http://www.xtremesystems.org/forums/...postcount=1551
Haha I see the sig got changed.
I'm still in awe that AMD somehow crammed 2.5x ALU's, 2.5x TMU's, redid the internal chip communication, distributed L2 caches, beefed up the ROPS/RBES in just 30% more die space...
FYI 4850's are on the egg now
Good stuff.Quote:
The entire array shares 16KB of local data share memory. The individual shader units themselves have been optimized, with more aggressive clock gating to reduce power consumption, and added integer bit-shift operations on every single unit (RV670 only supported such operations on a few of its shader units).
Now the million dollar questions are: Is HD4850 being overly limited by it's low bandwidth and does R700 have tricks up it's sleeve. If either one of those questions is a "yes" then NVIDIA hasn't been in this much trouble since 9800PRO. If the answer to both questions is yes, then God help them.
it's all about the memory controller. Ringstop controller was transistor expensive. If you look back in the thread, I mentioned first to Perkam that this chip was near 1bil transistors, but had 800SP's, and better TMUs. Noone mentioned that the transistor count didn't add up unless something else changed.Quote:
Originally Posted by zerazax
True it looks like they did away with the unnecessary part of it and put it local and global data shares connected to a crossbar
They also stopped using ringstops for pci-e and crossfire communication...
Funny thing is when u know the specs but cant say anything because of ur NDA and u just reason w/ other users. U try to convincen them as u know whats real but u cant break the NDA in a place this public.
If i have vista and dx10 i can't play games with dx9?
DX9 apps work in Vista.
U can play DX9 & DX10.
Haha yeah with the "unaligned text" :clap:Quote:
Originally Posted by w0mbat
Since the 4850's are now on newegg a week early does that mean the 4870 will be a week early to????
http://www.newegg.com/Product/Produc...HD+4000+series
I believe the 4870's have actually been delayed until something like the 2nd week in July.