swaaye wrote:

agent_x007 wrote:

It is clear winner in High Polygon Count (1 light) test, I give it that, but it lost in 8 lights test and barely have any leads in 4 lights test.
PS. Also TNT2 isn't a GeForce.

I think something is going wrong with 3DMark2000 on that system. The results just aren't right. Fillrate for the GeForce is way off the mark. It should not be less than any TNT2. The GeForce is only showing about 25% of its potential...

You realize he's using his super rare 64-bit SDR NV10? The TNT2 Pro has over double the memory bandwidth available.

swaaye wrote:

Also, GeForce 256's T&L is only about as fast as a low-end Pentium III at doing T&L calculations. You have a crazy fast Wolfdale CPU. You would probably not want GeForce 256 doing any T&L for you. So I think the results for the lighting tests are funky because of that. 3DMark might not be scoring any of that stuff right with such a fast system.

Well for an accurate assessment of what the GeForce 256 can do, I think one has to use its T&L unit.

swaaye wrote:

I think something is going wrong with 3DMark2000 on that system. The results just aren't right. Fillrate for the GeForce is wa […]
Show full quote

I think something is going wrong with 3DMark2000 on that system. The results just aren't right. Fillrate for the GeForce is way off the mark. It should not be less than any TNT2. The GeForce is only showing about 25% of its potential...

See here
http://www.tomshardware.com/reviews/full-revi … dia,204-10.html

Also, GeForce 256's T&L is only about as fast as a low-end Pentium III at doing T&L calculations. You have a crazy fast Wolfdale CPU. You would probably not want GeForce 256 doing any T&L for you. So I think the results for the lighting tests are funky because of that. 3DMark might not be scoring any of that stuff right with such a fast system.

I think @dexvx is right, but here's GeForce 2 Ti @ Ultra score on this platform (for comparison to that GTS)...

Fun fact about ^this platform :
CPU's RAM has more bandwidth than VRAM on GeForce 3 Ti 500 😁

I had no idea there is a 64-bit GF256. Heh.

Reputator wrote:

swaaye wrote:

Also, GeForce 256's T&L is only about as fast as a low-end Pentium III at doing T&L calculations. You have a crazy fast Wolfdale CPU. You would probably not want GeForce 256 doing any T&L for you. So I think the results for the lighting tests are funky because of that. 3DMark might not be scoring any of that stuff right with such a fast system.

Well for an accurate assessment of what the GeForce 256 can do, I think one has to use its T&L unit.

Yeah it's an interesting consideration though. What once was a performance benefit with systems of the time is probably now a bottleneck. Even back then it was talked about how GF256 was slower than the ~1 GHz CPUs for T&L. It was still beneficial because it's better to offload that stuff and have the CPU do other game systems. Maybe not with a 4 GHz Wolfdale though! 🤣

With a 4 GHz Wolfdale it might be best to just have a pure software renderer for your old D3D 7 games. 😁 Lots of SIMD power and bandwidth there.

What I gathered from one of the in-depth reviews indicated that a ~30K polygon budget is balanced for a target of 60 FPS before the T&L engine becomes a bottleneck.
Tests that synthetically stress T&L were especially showing large relative gains over CPU T&L.

I'll probably get 64-bit abomination from "Eagles" soon.

silikone wrote:

Tests that synthetically stress T&L were especially showing large relative gains over CPU T&L.

You have to remember that NV10 was released in very early 1999 as Quadro. Even Pentium 3 Katmai wasn't released at that time.

The Serpent Rider wrote:

You have to remember that NV10 was released in very early 1999 as Quadro. Even Pentium 3 Katmai wasn't released at that time.

Nvidia officially announced it in November 1999....

The Serpent Rider wrote:

I'll probably get 64-bit abomination from "Eagles" soon.

silikone wrote:

Tests that synthetically stress T&L were especially showing large relative gains over CPU T&L.

You have to remember that NV10 was released in very early 1999 as Quadro. Even Pentium 3 Katmai wasn't released at that time.

Does there exist some list of all Quadro cards along with their "desktop variants"?

Tetrium wrote:

The Serpent Rider wrote:

I'll probably get 64-bit abomination from "Eagles" soon.

silikone wrote:

Tests that synthetically stress T&L were especially showing large relative gains over CPU T&L.

You have to remember that NV10 was released in very early 1999 as Quadro. Even Pentium 3 Katmai wasn't released at that time.

Does there exist some list of all Quadro cards along with their "desktop variants"?

I do know that when you look at gpureview.com for example for fx5800 you get all the specs. Then click on the nv30 gpu in that list and you'll get all other cards based on that gpu including the quadro versions.
Very handy when searching for good alternatives 😎

The Serpent Rider wrote:

You have to remember that NV10 was released in very early 1999 as Quadro. Even Pentium 3 Katmai wasn't released at that time.

This chart provides some really good information. This Quake 3 map pushes polycounts at about the half of the first six digit number.
What's interesting to note is that even with a 1GHz CPU, it remains the the sole bottleneck on both the 256 and the 2, and the acceleration still makes a huge difference compared to a non-GPU. The typical scenario of Ati being more harsh on the CPU is also evident. The fact that the Geforce 2 has no performance advantage despite its improved T&L shows that the 256 already has ideal off-loading as long as its boundaries are not crossed.

Q3 doesn't really use HW T&L. All lighting and model transforms are calculated in software then batched.

I remember a video of Brian Hook claiming that they leveraged hardware transformation. Is it the virtual machine stuff that limits its benefit?

For a proper T&L test, try the High Polygon count test in 3DMark2001.

Scali wrote:

For a proper T&L test, try the High Polygon count test in 3DMark2001.

I got results from it as well 😀

I overclocked VRAM on my GF256 64-bit to TNT2 Pro speed.
I think 8 Light test with TNT2 Pro is in correct... (or should it be not supported by TNT2 ?).

agent_x007 wrote:

I think 8 Light test with TNT2 Pro is in correct... (or should it be not supported by TNT2 ?).

That's 3DMark01's software T&L in action, which as indicated by your results works pretty darn fast on a Wolfdale, as Swaaye was talking about.

I think my long search has ended. Found a badly labeled item on eBay while I was browsing for other stuff just by chance. Price was reasonable. Looked up the part number, and it is a GeForce 256 DDR. Hopefully the seller doesn't send me the wrong one or some sh1t.

dexvx wrote:

I think my long search has ended. Found a badly labeled item on eBay while I was browsing for other stuff just by chance. Price was reasonable. Looked up the part number, and it is a GeForce 256 DDR. Hopefully the seller doesn't send me the wrong one or some sh1t.

Would you have a link to the listing?

Has someone got the Hercules/Guillemot 3D Prophet DDR-DVI? Chips are marked SGRAM, with both sides populated. My card is pretty beat up, and I need some high resolution photos to ease the restore attempt.

agent_x007 wrote:

I think 8 Light test with TNT2 Pro is in correct... (or should it be not supported by TNT2 ?).

The test uses per-vertex lighting, so all light values are pre-calced in the T&L phase. The actual triangle rendering is the same, regardless of the number of lights.
You render the same number of triangles, and in both cases you just perform gouraud shading (interpolating diffuse and specular light values between vertices). The only difference is in how the per-vertex light values are calculated.
So in the case of the TNT2, where you use a very fast CPU for the T&L, you see that the performance is virtually the same, which is expected when the rendering is mostly limited by the rasterizer.

If you use a period-correct CPU, then the TNT2 would probably be slower than the hardware T&L cards, especially with 8 lights, because the CPU would be the bottleneck.

Is the application of eight lights not largely variable? Local lights should be more demanding than directional sources, and the amount of geometry being affected by each light is also to be considered.
The 3D Mark test seems to simulate the worst case scenario, with the maximum number of hardware lights all present locally, blending together on a high-poly mesh. I'd imagine that in more practical usage, eight lights would be feasible on this period of hardware with excellent results.