Dedicated PhysX GPU performance dependence on PCI-E bandwidth
Using a mid- or low-end GPU as dedicated PhysX card is quite popular idea today (even among ATI owners), while PCI-E bandwidth requirements for dedicated PhysX GPU are not absolutely clear. Everyone knows, that x8 and even x4 would be enough, but what about PCI-E x1 ?
PCPOP.com has published article today, that investigates PhysX performance of GTX295 + 9800GT setup, with 9800GT used in PCI-E x1/x4/x8 cases (unnecessary pins were taped).
And, surprisingly, even PCI-E x1 usage isn’t affecting performance as much as it should at a glimpse.
You can read rest of the article here
BTW, users with old mobos (or lack of free full scale PCI-E slots) are asking sometimes – “What if I just put my old 9600GT in PCI-E x1 slot ? Would that be enough for PhysX ?“. Now they have some clarity. Of course, mechanical “adjustment” will be needed to make it fit, that can result in GPU/mobo malfunction, so we don’t recommend even to think about it until you are clearly sure.
Thanks this helped answer my question:
I have a Gigabyte extreme x58
Two 16x pcie
one 8x pcie
In order to give my 285gtx breathing room, I have my 8800gt in the bottom 8x slot.
How much of a performance hit does 8x physx have over 16x physx?
So, again thanks for the answer.
I kept it in 8x the heat generated by the two gfx cards in such close proximity isn’t worth it, especially cause I can just oc the 285 with the lower temps.
Amyn
27 Jan 10 at 9:45 am
I just bought a PCI Geforce 8400GS 512MB how bad of a performance hit would I have if I use a PCI card rather than PCI-E. My machine is an HTPC so the x1 PCI-e is being used by a TV Tuner. Any help would be appreciated.
Richard
3 Mar 10 at 8:52 am
I tried a similar test using a asus p5n 570i (pci-e v1 switchable x1 or x8) and msi 750i (pci-e v2 x8) with an e5200, ATI 4830 and NV 8800gs.
Generally speaking, my experience mirrors the article. However, the v1×1 did have noticeable, but tolerable, bandwidth issues. Both x8 slots were essentially equal however v2×8 was more fluid, I believe this is due to the halved latency of v2 and not the doubled bandwidth. Increasing the pci-e v1 bus to 111MHz reduced the lag on the v1×8 and noticeably boosted the v1×1 slot performance.
My guestimation is that a v2×1 or better would fully satisfy any physx needs. Unfortunately, Intel chip sets typically have pci-e v1 x1 slots and even worse, they are on the south ICH bridge which introduces even more latency compared to most NV and AMD chips.
Ron
22 Apr 10 at 11:22 am
hi,
did all tests they performed on v1 pci-e or v2, cause in article x58 chipset have both standards from different x58 and ich10 hubs? i guess it was v2 but either assume it could be v1 from south bridge too.
[nv40]
5 Aug 10 at 12:36 am
Am I susposed to read chinese now??
that link goes to a chinese language website.
jose
2 Sep 10 at 6:50 pm
jose
Am I susposed to read chinese now??
You don’t need to read – graphs are sayng more than enough
Zogrim
3 Sep 10 at 12:46 am