[pointertovoid]

Hello tinkerers!

As graphics cards continue to increase their computing power, their onboard Ram keeps more or less the same throughput...

Is that a limit? Or do they now have an integrated cache memory, which the drivers and DirectX uses efficiently? Did you notice a speed difference from different Ram options?

Thanks!

[pointertovoid]

Found some bench comparisons:

http://www.xtremesys...z-vs-4870/page4
http://www.hardwares...ard-Review/1272
http://www.hardware....ddr5-utile.html (Soʁʁy foʁ ze lãguage)

These are synthetic figures, meaning that Ram throughput will often have no importance but sometimes make the game lag.
Also interesting: more Ram throughput is sometimes more important at higher resolution, sometimes at lower resolution.

Apparently, recent drivers or dX can cope with a still slow memory by using the internal cache. This must be delicate programming.

More inputs?

[submix8c]

I say now and say again... FSB FSB FSB (bus speed)!!!! It's the difference between AGP 1x->2x->4x->8x->?x

A MOBO with PCI-e will likely be faster than AGP just like AGP is faster than PCI... (note that the Mobo's are better through each phase.)

Wiki that will help understand-

Quote

...including higher maximum system bus throughput, lower I/O pin count...

Yes, more RAM is a definite help but if you put any same-type-card with more RAM onto the same PC, the bus speed does not change... It's the same difference as adding more RAM to a given PC - the CPU/FSB speeds didn't change but you reduced the "choke" of Paging. Note that a GPU=CPU, so yes it will increase BUT the limit will be the FSB (that all-encompassing MOBO speed). Obviously proper drivers will help.

So... to alleviate that "limit" - Upgrade!!! You could take CardA off of MoboA and put in MoboB (provided compatibility w/CardA) and the "better" Mobo (CPU/FSB/RAM) will undoubtedly give better results.

Please note the Test Results are based upon using the SAME HARDWARE CONFIGURATION between the Two Cards being compared.

More input found here...

[pointertovoid]

submix8c, on 30 June 2012 - 08:41 AM, said:

FSB... AGP... more RAM...

The thread is about the throughput of the Ram on the graphic card.

[submix8c]

I realize that... but the limitation is still based on the MoBo, no matter how you slice it (in my feeble understanding).

As I said (AFAIK), change MoBo's to a better one (FSB, CPU, RAM, etc.) and watch throughput increase....

Quote

Is that a limit?

Quote

More inputs?

So, your point is? I gave my input...

[pointertovoid]

I've received an Hd 6570 to replace the Gf 8600gt and tried it with Colin McRae 2005.
The Hd 6570 has more computing capacity than the Gf 8600gt but about the same Ram throughput.
It also has a PciE 2.0 instead of 1.0, and 1GB instead of 256MB.

While I can increase the resolution to 1440*900 - but not to the screen's 1920*1200 - I must keep the "medium" texture definition. This adjustment has much heavier consequences than the image definition.

I believe to understand that the texture definition has no effect on the demand on vertex and pixel shading operations.
It would demand more throughput from the disks and PciE, and possibly the Cpu in between, BUT this game is from 2005, so I suppose an E8600 and an X25-E are more than enough for it, and 1GB of video Ram can hold the complete textures anyway.
It does also demand more throughput at the video Ram, so I suppose this is a real limit of this unbalanced card.

Your opinion?

[pointertovoid]

I've compared again the Gf 8600gt and the Hd 6570, this time on Xp, with Ski Challenge 2009 and Mountainbike Challenge 2010, at 1920*1200 pixels.

The Hd 6570 is a bit worse than the Gf 8600gt despite claiming brutally better computing capability.

My only explanation is that the video Ram throughput is the real limit of both cards: ddr3 128 bits.
Well, maybe more recent games use on-chip cache better.

[pointertovoid]

I've tried a 55nm Gf9600gt as well, with 288GFlops (8600gt: 113GFlops, Hd6570: 624GFlops). While the two previous cards perform similarly, the Gf9600gt outperforms both ones on games. I understand its Ram throughput makes the difference.

To put figures on that, I used 3DMark 2001se. It's a dX8 test, about as old as the games I play, and I put its resolution at 1920*1200*32 which is my target at games. I noted the smallest Fps (and the mean Fps) at the tests "Car Chase High", "Lobby High" and "Nature". These gave me the same ranking as the games.

I also varied the Gddr3 frequency a lot at the Gf9600gt, because the initial 1000MHz were abnormally low. The result is convincing, with some test improving nearly as much as the Ram frequency. Please note that on the graph here (log in to see, click to magnify), the Fps of "Lobby high" are halved.

 9600gt_MHz_Fps_Series.png (20.47K)
Number of downloads: 2  9600gtEtc_FiguresMHzFps.png (4.56K)
Number of downloads: 2

From these measurements, we see clearly that the Ram determines the speed, more so the minimum Fps than the mean ones.
- Measured at 1024*768*32 with the dX9.0 test 3DMark 2003, the Ram throughput is still very important. Anyway, recent cards have to address big screens.
- All these cards are dX10 or dX11 with less efficient unified shaders, while my dX9 games and tests use vertex and pixel shaders. Maybe the translation by the drivers demands more bandwidth.
- It could be (or not) that dX10 games take a better advantage of the on-Gpu caches.
- Multi-Gpu cards and Sli improve the bandwidth, but they may also need more if each Gpu requests the full amount of data in the Ram. It depends on driver programming.

Marc Schaefer, aka Pointertovoid