Assuming my math is correct (often it's not) that gives a maximum of 145FPS and 60FPS would be ~55MB/s at 800x600x16.

5u3 wrote:

TheLazy1 wrote:

Now I'm wondering if you could read the Voodoo2's framebuffer and display it on a separate window.

Yes, there even were Voodoo(2?) cards which did away with the analog pass-through and copied the framebuffer over the PCI bus.

Some games even did this on the plainest vanillaest Voodoo2s available. Rayman 2 for example, can Glide in a window on a Voodoo2. How it happened it freaking shocked me 😳

I appear to have amassed a small collection of AGP cards, bottoming out at the lowly ATi Rage. Feeling bored tonight, maybe I'll benchmark them and post the results! Not as diverse a collection as in the OP, but here tis:

nVidia:

GeForce 256 DDR
GeForce 3 Ti200
GeForce 4 MX400
GeForce 4 MX440
GeForce FX 5200
GeForce FX 5600

ATi:

Rage 128 Pro Ultra (Ex Dell)
Radeon 9200
Radeon 9200 SE
Radeon 9550
Radeon 9700 Pro
Radeon 9800XT

So stay tuned!

I have some of those too but I stopped at GF256 because the later boards are too fast for the resolutions the older cards can handle. 😁 You'll have hundreds of fps and quickly be CPU limited.

I'm testing them on a 2.4GHz Northwood, the benchmarks are 3DMark 2001 SE, 3DMark 2003, and AquaMark. These all bench at 1024x768 32bit.

Tonight I'll go back and run 3DMark 99, as per the OP. I stayed up till 1AM dicking around with drivers and stuff, so I only got to bench three cards using the above benchmarks. Today it's all set up so I can get some actual testing done. 😀

Blue = AquaMark
Purple = 3DMark99
Yellow = 3DMark01
Cyan = 3DMark03

Wow, what a tedious process that is!

I'm stuffed, I'll do the nVidia cards next weekend, as well as post the GPU-Z grabs from all the cards.

BTW, there's no errors here, the 9200 was ever so slightly faster than the 9550 in 3DMark99, no matter what I did with drivers, control panel options, etc.

*phew*

It's interesting to see 9200 beat the 9550. But I know that R200 level cards do a lot less work to make a frame because they have much lower quality texture filtering. They also have a lot of bugs so maybe something wasn't even being properly rendered.

R300 cards also tend to be somewhat less efficient with older games. Or maybe I should say that old games don't leverage the new design's strengths at DX9. GeForce FX was actually very competitive with them in this way.

SquallStrife wrote:

Wow, what a tedious process that is!

Yeah a little bit of benchmarking several cards gives one a nice perspective on the work the hardware reviewers put in. 😀

BTW you forgot to identify what each bar color represents.

swaaye wrote:

BTW you forgot to identify what each bar color represents.

Haha! I did too! Oops!

From left to right, the bars are AquaMark, 3DMark99, 3DMark01, 3DMark03.

Task for this weekend is to bench the nVidia cards.

I don't know if it has already been discussed here, but Kyro II's performance in 3DMark99 always puzzled me. Not only it owns its main competitors of that time (GeForce2 and Voodoo5), but it also beats more advanced cards like Radeon 9000 or GeForce4 4200. 3DMark99 scores should be taken with a grain of salt.

batracio wrote:

I don't know if it has already been discussed here, but Kyro II's performance in 3DMark99 always puzzled me. Not only it owns its main competitors of that time (GeForce2 and Voodoo5), but it also beats more advanced cards like Radeon 9000 or GeForce4 4200. 3DMark99 scores should be taken with a grain of salt.

It is Kyro's rendering technique that makes it fast in benchmarks. It discards any information that you won't actually see on screen and only renders what it has to to create the scene. This discarding of all unnecessary information speeds it up a lot. Having to draw objects that are fully or partially hidden from view wastes a lot of power that could be applied to other things. Unfortunately, Kyro never had the raw power to really maximize it's benefits and pull ahead of the other video card makers. That's why it stumbles in actual gameplay even though it does well in benchmarks.

The hidden surface removal the Kyro was hyped for didn't help the PCX-2 all that much, and in real games it's hardly ever useful as the CPU still has to deal with the vertices you can't see. GTA3 (a game infamously slow by excessive overdraw) is still slow, so no miracles there.

PowerVR probably did the usual 'game profile' tweaking to compensate for more 3dmark99 performance.

Well it did get them a lot of performance for their relatively tiny fillrate. It was another way to get similar performance.

I read that their tiling architecture made T&L and shaders more difficult to implement. One or both of those.

Not sure if this is still private property, but here's NVIDIA's take on the KYRO2:

h**p://www.russell-crow.com/nvidia_on_kyro.pdf
(most of this was taken from public reviews and common sense anyway)

sliderider wrote:

Unfortunately, Kyro never had the raw power to really maximize it's benefits and pull ahead of the other video card makers. That's why it stumbles in actual gameplay even though it does well in benchmarks.

Kyro II was not that bad in actual gameplay. I got 120-130 FPS in Quake III @ 1024x768x32, a significant performance boost over the 90-100 FPS I got with my previous card, a highly overclocked and modded GeForce DDR. It could be achieved by activating a very useful feature in games that allowed it: forcing both S3TC compression and trilinear filtering enabled what PowerVR called 'fast trilinear filtering', a trick that almost doubled framerate without any noticeable quality loss (you can still detect the mipmap transitions if you carefully look at some detailed textures, but they were almost invisible most of the time).

With this trick, Kyro II could make more expensive cards bite the dust in real games like Serious Sam FE and SE (being Serious Engine one of the most optimized and customizable graphics engines ever made, IMHO). These games could really take advantage of some otherwise underrated cards with unused features, like Radeon 7500's third texturing unit, Radeon 9000's single pass 6-layer multitexturing, and of course Kyro II's single pass 8-layer multitexturing.

swaaye wrote:

Well it did get them a lot of performance for their relatively tiny fillrate. It was another way to get similar performance.

I read that their tiling architecture made T&L and shaders more difficult to implement. One or both of those.

That's what I meant when I said they didn't have the power to leverage the advantage of tile based rendering. If they had a comparable fillrate to nVidia's cards, they would have been even faster. Of course their lack of other features that gamers looked for as pointed out in the doc that was linked also weighed against them.

I can't help but wonder if the tiling deferred renderer has its own baggage that prevented them from scaling up to the same level of fillrate with the same manufacturing technology / chip design techniques.

Although if you look at the linked NV PDF above they say that PowerVR's design partners weren't exactly the most skilled/experienced. So maybe they were lucky to pull off what they did manage.

I actually find PowerVR stuff to be pretty interesting. They are still around big time too, just look at most of the phones and tablets out there. These little handheld devices that put off very little heat can play Quake 3 well and have OpenGL support similar to DirectX 9.

BTW, Intel's 845G to GMA 950 (and some of the newer non "HD" GPUs) use a form of deferred rendering called zone rendering that has similar overdraw benefits. It makes sense to do something like this when you have absolutely pathetic memory performance (IGP world).

swaaye wrote:

I actually find PowerVR stuff to be pretty interesting. They are still around big time too, just look at most of the phones and tablets out there. These little handheld devices that put off very little heat can play Quake 3 well and have OpenGL support similar to DirectX 9.

BTW, Intel's 845G to GMA 950 (and some of the newer non "HD" GPUs) use a form of deferred rendering called zone rendering that has similar overdraw benefits. It makes sense to do something like this when you have absolutely pathetic memory performance (IGP world).

Probably doesn't work too well in the real-world when combined with Intel's usual crappy graphics drivers. Also the GMA 500 and 600 is basically a PowerVR SGX535 core licensed to Intel, found in netbooks and other low-power devices. Too bad its gaming performance sucks though (well for a low-power chipset, what do you expect anyway), but it seems decent enough for decade-old games such as Q3A/UT99.

Luckily PowerVR managed to find a market where they could succeed, otherwise they could have gone the way of 3dfx. We could certainly use some more competition for desktop graphics nowadays though............

Pippy P. Poopypants wrote:

Probably doesn't work too well in the real-world when combined with Intel's usual crappy graphics drivers.

I have an EeePC 900 that has GMA 900 inside. I use it for a lot of retro games. It seems to have the performance of a GeForce 2 GTS or so. It really does work pretty well for games from 2001 and earlier. Of course it is advertised as a DX9 GPU and it is nowhere near fast enough for that.

Glad you found great use out of it though. 😀 The DX9 part was pushed to garner consumers and maybe to meet the then-upcoming Aero spec. for Vista. Still beats the GMA500 in performance, even though the GMA500 is more feature-rich.

I was very curious about that GMA 500. It's too bad that it's such a gimpy PoweeVR design, but it was designed for phones.

I think PowerVR just has never had the budget to really push ahead. But this is probably a result of their inability to even remotely dominate during the '90s.