Was anyone waiting for a review on the 4870 1 gb? wellread here.
review
Printable View
Was anyone waiting for a review on the 4870 1 gb? wellread here.
review
http://www.hardwarezone.com/img/data...ysis-xp-AA.gif
the extra mem seems to come in more as you go up the res. chart...with 4xAA
That comparison's no good....the powercolor card is overclocked to 800/925 which doesn't show that its just the effects of the extra memory.
yeah not much different to the 750MHz 512 4870....
http://www.hardwarezone.com/img/data...s-vista-AA.gif
except at 1920x1440 4xaa, with unplayable fps anyway...
i believe the graphs clearly indicate that the extra memory increases the fps at higher res.
50MHz isnt going to boost your frames that much
Unless driver is written to properly support the extra vram, don't expect much. And seeing how AMD has not released an official 1gb card, other than the 4870x2's dual 1gb gpus, I'm not going to hold my breath waiting for a suitable driver for thees cards.
I think the reviewer is somewhat skewed. Or perhaps it's the testing methodology? How come a single 4870, even with 1GB of VRAM, outperforms GTX 280 in Crysis? Also, how is it possible, shoddy scaling or not, for a single 4870 to beat dual 4850's? Compare to a review here:
HWC Review
Just a heads up.
Please explain, I was not aware that Xfire support is FSB-related
Not yet, and 4870 X2 is still faster than GTX 280 non-OC
another review
Hardspell
each gpu to gpu transaction goes through the NB so its card to card bandwidth and latency is directly effected by the cpu FSB (on intel)
1. GPU to GPU also uses the bandwidth of CrossFire connectors together with PCI-E bandwidth. Connector-less crossfire is only effective on lower-end Radeons as they don't require the extra bandwidth provided by the connectors.
2. FSB speed is in NO way related to PCI-E transfer latency/bandwidth. FSB connects the CPU to NB, and that's it. PCI express lanes are FSB-independent
Please check here:
http://en.wikipedia.org/wiki/Front_side_bus
========================
This thread is about 4870 1GB vs 512MB performance, please post any further findings here. Looks like 512MB is not quite enough for 4870, but 1GB is probably overkill. Too bad ~750GB is not possible.
your idea would work if going through the NB the NB could send it back but it has to go card->NB->CPU->NB->card, and the tatal bandwidth that the NB can take is directly related to the FSB since they are synced, so if u have a low FSB say 266 then u have a max IO bandwidth of 8.5GB/s (core fsb * 4*64/8) while a 400mhz fsb gives u 12.8GB/s and 500 gives u 16.0GB/s so to be able to feed the cards u need enough bandwidth for each card to exchange data and to transfer from and to the system memory so its an almost linier increase in performance until u have enough IO bandwidth
and the total latency is the pci-e/NB and the cpu L2 cash so u can almost cut that cpu side in half from 266x10 to 400x9
i would like to see some real game benchmarks like a UT3 based game, COD4, DMC4, and then some maya, like what hard has. it looks all gpu clock to me
I see your point, but could you give me a link to anywhere it shows that between-GPU communication goes through CPU and not card-to-card via PCI-E lanes with driver arbitrating the process. I'm not sure. Thanks