VOGONS


First post, by MaRCer

User metadata
Rank Newbie
Rank
Newbie

I bought a GeForce 7800GS from an obscure Chinese brand. It uses a reference PCB and features a GPU taken from a 7800GT (with a 20/7 configuration). The card seems to run fine—no artifacts or freezes—and it overclocks to 480/1500 MHz. It scores higher in 3DMark than the 6800 I had previously, though only by a small margin: the 6800 managed 4,100 points in 3DMark2005, while the 7800GS hit 4,500 points (or 5,100 when overclocked).
My system consists of an E5800 CPU at 3.5 GHz, a GA-8I865GME-775-RH motherboard, 2x1 GB of DDR1 RAM (running at 440 MHz, 3-4-4-8 timings), Windows XP SP2, and ForceWare drivers (versions 92 and 175), all powered by a 450W FSP power supply. Through some experimentation, I discovered that the card’s vertex shaders perform very poorly; in the "Feature Set" test, the pixel shaders run twice as fast as those on the 6800, but the vertex shaders are 2–3 times slower. Using RivaTuner, I tried disabling some of the vertex processors: leaving only one enabled causes a further drop in speed, but enabling three of them results in only a 10% performance loss compared to having all seven active. I also tested the card on another motherboard—one with a SiS651 chipset (it’s the best alternative I have)—and the situation remained the same. I even tried flashing the BIOS from an XFX card; although that version is designed for Samsung memory (while mine uses Hynix), the card’s behavior didn't change at all. So the problem lies specifically with the graphics card—but what kind of strange behavior is this, where the vertex processors are the ones lacking power or memory bandwidth, while the pixel processors have everything they need?

Reply 1 of 6, by agent_x007

User metadata
Rank Oldbie
Rank
Oldbie

In general, Vertex test will be somewhat CPU limited (since it is the CPU's job, to deliver data for Vertexes - if it or RAM/AGP bus can't keep up, score won't go up with more Vertex units).
That said, it's still hard to tell if your CPU is the issue when we have no other scores to compare against.

So, which of these three scores is closest to your card score in Vertex test ? (ideally, you would have to clock match to each card to be fair - including GPU driver)

The attachment 3DMark 05.PNG is no longer available
The attachment 3DMark 05.PNG is no longer available
The attachment 3DMark 05.PNG is no longer available

Reply 2 of 6, by MaRCer

User metadata
Rank Newbie
Rank
Newbie

Pentium Dual e5800 at 3.5 GHz is not the weakest processor. In any case, on the same system, if you insert 6800 in Vertex Shader Simple, you get 42Mvertices/s, but if you replace it with 7800, it drops to 15Mvertices/s. And in the CPU test, the results drop by a factor of 2.5.

The attachment IMG_20261121_175538_144.jpg is no longer available
The attachment IMG_20261121_175538_143.jpg is no longer available

Reply 3 of 6, by agent_x007

User metadata
Rank Oldbie
Rank
Oldbie

Not 1:1 with your GF 6800 LE, but should be close enough :

The attachment 3DMark 05.PNG is no longer available

Unless your Vertex units are fused off (unlikely)... I think this is just a driver issue.
Maybe try older driver on 7800 GS AGP (you can clear the old one with DDU, v17 should still work fine under WinXP).

Reply 4 of 6, by MaRCer

User metadata
Rank Newbie
Rank
Newbie

I have a 6800LE unlocked to 12/5. It doesn’t matter.
No, it’s not the driver. I tried 93 and 175, I tried installing Windows 7 with a driver that seemed to be 120. I installed XP from scratch on an SIS651 system, installing AGP drivers — but it’s still the same: when you switch from a 7800 to a 6800, performance in CPU and vertex tests increases noticeably.

Reply 5 of 6, by Ozzuneoj

User metadata
Rank l33t
Rank
l33t

Just a thought...

Have you removed the cooler and looked at the GPU die itself? Just to 100% verify that the markings are correct and that it isn't a different GPU relabeled or an engineering sample or something.

Obviously if it really has a 256bit memory bus and that part is performing correctly, then that would rule out any lower-tier 7000 series GPUs below the 7800GS, but... this is really perplexing and is starting to seem like a hardware issue. I have never heard of vertex processors that were detected, could be toggled on\off, didn't create graphical errors but also barely affected performance. Super strange issue, and I would love to know what is causing it.

I guess it's also possible there is something different or wrong related to the AGP bridge, so I would look at that too. Since it's an off brand card they could have implemented that in a way that creates a bottleneck between the CPU and the GPU. Not sure if that would manifest in this way... but it's something else to check. This article has some nice close up shots of the bridge chip.

I wouldn't mind seeing some pictures of your card and how it looks with and without the cooler, just so we have a record of this if anyone else runs into it in the future.

I see that some cards like this say they are 24/8 configuration, so there seems to be some variation out there among the Chinese brand cards.

Now for some blitting from the back buffer.

Reply 6 of 6, by MaRCer

User metadata
Rank Newbie
Rank
Newbie
The attachment avito-1795340359971-shot.jpg is no longer available

Here is a photo of the GPU. The AGP bridge also appears to be undamaged.