PCI-E 2.0 vs 1.0 impact on performance?

How about GPGPU output? Physics too.
Or stuff like craters on terrain & damage decals for persistent environmental effects etc.

 

I was assuming brit's comment was referring to the processing of HD video but maybe I misunderstood him. But yeah I'd be interested to know what exactly has to be read back across the bus as well. I previously had the understanding that an encoded bit stream would be passed from the CPU to the graphics card directly and all needed processing would be done on the GPU and outputted directly but maybe that isn't the case, or maybe it's only the case for certain cards or something. Anybody want to clarify?

 

Readback speed was always the achilles heal of the AGP slot. It's one of the big reasons for the push to PCIE. AGP was designed to deliver (at the time) massive bandwidth from main memory to the Video Card to enable texturing from main memory. Remember it was designed back when graphics memory was incredibly expensive and at the time both MS and Intel were investigating alternative (cost effective) means to texture 3D scenes.

Luckily for us the consumers, graphics memory prices took a dive and it became more and more affordable for companies to continue increasing memory on video cards. From the average of 4 MB (actually probably lower as it was designed when PCI was still the main graphics card interface) at the time to cards with up to 1 GB today.

Writeback speed from GPU to main memory wasn't even a blip on the radar at the time.

[EDIT instead of new post] - What would you need/want to be read back? Well SLI/Crossfire information for one if we want bridgeless SLI/Crossfire. GPGPU would be another biggie. And if Intel allows Havok to continue developement (doubt it) or if Microsoft makes a serious push for it, interactive Physics on the GPU rather than Physics for particles that don't interact with anything.

Regards,
SB

 

Part of the problem with AGP readback was also that nobody used it that way, so graphics driver companies never built anything particularly efficient for such things.

Later video drivers almost entirely corrected the issue, last I recall. I can't immediately find benchmarks either way, but the push for PCI-E wasn't much ado about readback speed so much as it was far better scaling and the possibility of full-duplex data operations.

 

AGP texturing is useful yet when you run out of video memory. on a voodoo5 which works as a PCI 66MHz card in an AGP slot, I'd dive to 1 or 2 fps with too heavy textures (in UT2K3/2K4 ), it's not like this with real AGP or PCIe cards.

 

Quick run with an FX5600 on SiS748, WinXP driver 163.75. AGP8x ~200MB/s. AGP4x ~190MB/s.

 

So, anyway. On the 256MB Radeon HD 3850 we've seen performance differences of up to 20% in going from PCIe-1 to PCIe-2 on titles that have high numbers of assets / create lots of off-screen buffers.

 

Wavey, I guess that's expected when your frame buffer is too small for the workload

 

Any of the horde of hd3800 reviews cares to compare both cards on both buses?
Does a pcie 2.0 hd3850 come nearer a pcie 1.x hd3870?
Im considering a system buy. Was going for pcie 1.x (p35) and hd3870, but im not feeling amused about losing 20% performance in some situations.
Is that 20% relevant or it´s just relevant under swapping, so it´s just 5 to 6 fps vs. 35 fps without VRAM swapping?

 

I'm not convinced that EITHER of those AGP interfaces are a prime example of a fully functional unit

The FX5600 was a very poor midrange part, and the SIS chipset was quite far from the best too. ATI R300-based cards did far better, and so did motherboards based on Intel 865 or NForce3 chipsets.

 

Heh, I was half expecting that kind of retort & it was kinda the whole point... If this combo can get ~200MB/s, then readback isn't as bad as the previous older tables.

Actually, I never originally got around to the FX series... I picked up an FX5600 VIVO & FX5800 for $20 a few months ago. I hate to say it, but they're not that bad... The SiS748 was pretty sorted too. VIA KT880 was quirkier. NFx was hit'n'miss. In the same retro system an R9600XT (RV360) & AIW R9800Pro (R350) yield ~120-150MB/s. I've got an i875 & P4 3.0, so may give it a go with an NV40/R420 series card.

 

Hello all, first post

Wow, 20% increase on the 3850, that is killing me, i just got the asus blitz extreme and have been waiting for this release of ati 3870 video card, almost broke down and got the 2900xt, but held out.

I am concerned for 2 reasons, to make sure i don't have to buy a whopper of a psu but also I was hoping to get a nice ati product to crossfire. BUT now i see the pci 2.0 thing and it has me worried. Blitz extreme is 8x 8x pcie lanes crossfire while x38 mobo's are 16x 16x crossfire,,,

But on the other hand i just read a review on using SLI with the 650i and 680i boards. They showed that he 680i which is true 16x and 16x and compared it to the 650i which is 16x 8x was very interesting.... ( i am looking for article now) but it said that he thruput wasn't the slowdown most times, that the gpu's were waiting for info to process and that the bottle neck were elseware in the pipeline.
So to generalize the 650i with it's 16x x 8x pcie bus speeds even beat the 16x 16x of the 680i in some tests.
So in wraping this long post up, i am hoping that with 2 great ati cards the 3870 on the blitz extreme at 8x 8x, will show not much difference then the x38 chip with it's 16x and 16x or i will have to sell the blitz, (which i have been saving for this project) and get a Maximus extreme, (really don't want to)


But we will know soon.
Best Regards
Rhinoviper
Current system.

MSI p6n Platinium
q6600 quad core - go stepping overclocled from 2.4 to 3.6 - 1600fsb
PNY 8800 gts 320MB (wish had 640) overclocked 545 to 664 gpu and 1585 to 2k memory
2Gb mushkin P6400 800mhz ram (micron d-9 chips) overclocked to 1000mhz
500 GB sata 3.0 drive, 650 W rosewill psu and rosewill case.
2 asus MW201U widescreen monitors (love them)
air cooling, everything has a fan or cooler on it! no water all air!
running xp pro sp2
Love this system, will get a second pny 8800 when prices drop
3D mark06 13329 3Dmarks under name rhinoviper

 

waiting for a crossfire matchup

I would love to see the improvemet with pci2 also with 2 8x 8x lanes vs the 16x 16x

regards
Rhinoviper

 

found the article...
http://www.bit-tech.net/hardware/2007/10/12/crossfire_comparison_intel_x38_versus_p35/8

 

This the slide?


From:
http://www.kkk.com.ua/?p=245

 

the bars start at 80% to make it look like more impressive though, and the 512MB version's scores would be greatly needed. what if scores in those conditions for 256MB 1.0/256MB 2.0/512MB are 15/17/30 fps for instance, that would not be that interesting though I agree the future is bigger and bigger datasets with virtualization and streaming stuff.