hmm interesting, in particular the bit about the FX stuff, I still should have one of the big boys of the FX line in storage and at the time had a bunch of discussions with people who were "pooping" on it due to articles that came out dissing the architecture. Sure, it was imperfect, but it was far from the absolute dog it was made out to be, at least the higher end models were quite OK - expecting big things from the fx5200 or how it was called was a strange idea anyway.

3dlabs FANBOY in denial.

The 3Dlabs split of higher precision texture coordinate processing and lower precision fragment processing seems most similar to the Geforce3 and Geforce4. NVIDIA had a very similar split there, and this can be seen in the NV_texture_shader and NV_register_combiners split. At least some modern architectures still have vestiges of this. On Intel GPUs, the texture instructions conceptually send a message to an asynchronous unit that performs the texture look up. Once the data has been fetched and processed, it is delivered back to the shader execution unit. Depending on numerous factors (e.g., caches) the texture access can take a wildly variable amount of time. The texture sampler hardware is on a separate piece of the die from the execution unit. The huge difference between current and vintage architectures is that current architectures are asynchronous (other things can happen in the shader while waiting for texture results) and vintage architectures were synchronous (the next phase of ALU operations would block until the texture unit was done). As I mentioned in the video, Geforce FX only had flow control in the vertex stages. Fragment processing had predication that could be used to emulate if-then-else structures (similar to conditional move instructions on CPUs), but loops were impossible. You are correct that a lot of hardware from this era has a concept of alternating "texture phases" and "ALU phases." These architectures have a limited number of texture indirections, so the compiler / driver has to try to group texture accesses into bundles to reduce the number of phases. Each Shader Model has a limit for the number of texture indirections that are guaranteed to be supported. On R200 it's really small... 2 or 3. I know ATI_fragment_shader only allowed 2 passes. I think I'm going to have to do a deep-dive video on the Radeon 8500 architecture. On Intel i915, it's also pretty small. I want to say 4, but it has been years since I did any serious work on that driver. The general idea is that a texture phase receives data from a previous stage. That is either the vertex processing / interpolation or a previous ALU phase. Texture look ups occur, and that data is fed to either the next ALU stage or the color / depth outputs. ALU stages have some number of values that can be persistent by passing through the texture stages (this is related to the "pass any raw value they calculate" bit that you mention) and some temporary values that are lost at each texture stage. Fragment shader color inputs (via gl_Color and gl_SecondaryColor variables), color outputs (via gl_FragColor), and depth output (via gl_FragDepth) are implicitly clamped to [0, 1] and may have reduced precision. Intermediate calculations are expected to be "full" precision. The spec is a little vague about what full precision would be. Issue 30 of the GLSL 1.10 spec defers to the OpenGL 1.4 spec, "It is already implicit that floating point requirements must adhere to section 2.1.1 of version 1.4 of the OpenGL Specification." Section 2.1.1 is where I assume you got the 2^32 and 2^10 values. That section also says, "The maximum representable magnitude for all other floating-point values must be at least 2^32 ." "All other floating-point values" would include intermediate calculations in a shader. NV30 and R300 had different representations. NV30 could use either 32-bit single precision or 16-bit half precision. The latter was not directly exposed in GLSL. R300 used some sort of 24-bit precision internally. It met the requirements of all the specifications at the time, but I have some memory of it causing problems for developers that expected / needed full single precision. Given that the P10 shaders can supposedly be quite large, there are a few ways they could have gotten around this. A clever compiler could analyze the shaders to determine that some calculations would be fine at lower precision. I implemented a similar pass in my compiler. For many values in simple shaders, it is very easy to prove that the range will be limited. That would cover a lot of things. The GLSL spec and OpenGL 1.4 are pretty loose in the definition of precision and the processing of exceptional values (infinity, NaN, etc.). Perhaps this enables the compiler to use a different representation for floating point values... like storing an integer numerator and integer denominator. This is just speculation. There is one thing that still bugs me. GLSL is a superset of ARB_fragment_program (the assembly language shader). If the driver and hardware can do the former, it can, by definition, do the latter. Why not support it? Why not support Shader Model 2.0?

Those later cards seem like they might be more interesting. Did you have any of the 3Dlabs cards back in the day? I am curious to hear about people's real world experiences with them.

@@TalesofWeirdStuff I was born far too late for that. I got my first 3dlabs card (Wildcat Realizm 200) in 2011 at a flea market, without a clue on the significance it had. It was already outdated by then but ran TrackMania and Minecraft okay-ish on the terrible hands-me-down computer I had. I have collected a few other cards since, and I even managed to score some ZiiLabs ZMS stuff (the embedded ARM SoCs with StemCell GPUs, derived from their existing IP).

​@@TalesofWeirdStuff I remember the realizm cards whilst impressive for some professional applications had real issues if you attempted to game on them.

Damn you're lucky, mine died in 2004 :( I had the Powercolor 9700Pro AIW.

@@supabass4003 Why did it die? That's a mystery, because the high-end, especially that high-end, aren't even likely to have bad caps. Except for motherboards, I usually would see bad caps more in 2004-2005.

FX5600XT if I recall correctly was a sysintegrator only SKU, I it came as standard in some branded PC package deal? Pre-built can mean a local shop built PC package, or something totally different like an Acer, HP or similar. I recall at least Medion had FX5x00XT sysintegrator only cards that were strange ones, they seemed to be binning rejects or bus width gimped GPU's. I had either 5500xt or 5600xt that was a "faulty/disabled" 5800 where the memory controller was limited to 64mb at half speed or something like that. Very strange card, but not very useful.

Yes! I'd love to hear more about.

I expect that it will come out more as I continue to talk about graphics hardware and OpenGL. :)

Ha! The worst part is... I just got it in October or November of 2002 to upgrade my original Radeon DDR. A little late to that particular party.

@@TalesofWeirdStuff Not really. Due to the PS2 and original Xbox, games ran on DirectX8.1 until around 2006, with some games even in 2007+ still running on that API.

Uwe Scheffel's other articles are worth a read too, in this case ... " British Understatement: 3Dlabs' Wildcat VP Put To The Performance Test"

Thanks for finding that. $1,200 is still quite a bit more, but it's not as absurd.

@@TalesofWeirdStuff Correlates with what i found. (mostly german Hardware/IT Sites who had reports from 2002-2003 for these cards.) :D A german Hardware Website had a report with benchmarks who stated that the ATI FireGL X1 256MB was listed and sold for 995,-€. The 3Dlabs Wildcat VP990 Pro was MSRP 999,-€ before taxes, 1160,- with VAT applied. The VP760 was listed for 449,-$ The VP870 was listed for 599,-$ The VP970 was listed for 1199,-$ As stated in a report from June 24th 2002 on a german IT website.

I think there's plenty of room for lots of different perspectives. I equally enjoy LGR videos and Modern Vintage Gamer videos, but they tend to be very different levels of technical depth. I've been doing videos for almost 4 years, and I'm still trying to find where "my place" is.

Every little bit helps. :)

The bummer is that the scenery looks pretty good... too bad it's so dark you can't see any of it.

@@TalesofWeirdStuff The games ware also so scary before people had ratherd to turn it off.

I realized afterwards that since these cards have DVI, I could probably use my regular HDMI capture device. 🤦‍♂️ I'll try that for the next video.

That is a fabulous question! I will for sure check the noise function. For awhile Mesa had a real noise function, but we took it out because lower end GPUs couldn't do it. It generated way too many instructions.

Somewhere in my collection I have a 6600GT with Doom 3 on the fan shroud. I think eVGA made it?

Hatte auch die 6600GT. eine super Karte

I rarely "correct people," but it does annoy me when people call even something like a Voodoo 2 a GPU. But... I always tell people the correct way to say GIF is the *opposite* of however they just said it. 😂

I'm going to delve into the rendering modes in the next video. I finally looked it up... "gfxinfo" will show which modes are available, and "r_renderer" can be used to specify a mode. Based on the available functionality, it should be able to do the NV20 mode, but it might be falling back to ARB. The card doesn't have occlusion query, but I don't think Doom 3 uses that. I'm not 100% sure, though. It's possible that's a factor.

Yeah... maybe in 20 years I'll do a video about that card too. 🤣

That comment made me laugh so hard. Thank you! 😆😂🤣

The shirt is from the OpenGL "birds of a feather" session at SIGGRAPH 2021. It was celebrating the 20th anniversary of OpenGL. I'll show it a little better in the next video. :)

iD Software were known for pushing the boundaries. Back then, the games developers pushed the hardware guys every few months. And iD had been groundbreaking since Wolfenstein and Doom. Quake 1 required a very beefy computer. And Doom III was probably their most ambitious project of all. They worked directly with ATI (at first), gaining access to ATI's latest unreleased hardware, to code the game to stress the most capable cards out there from day one of the games release. iD made games ahead of their time. So a 2 year old GPU had no chance! Only the top end cards at the time of release were gonna "breeze" through Doom, same way only a top end Pentium 200 could "breeze" through Quake 1, 7 years earlier. ATI ripped iD Software off, a member of ATI Technologies "leaked" the code to the E3 Doom 3 demo onto the web, betraying iD's trust and releasing Carmacks code to ALL his competitors, for free! iD abandoned working with ATI on Doom 3 and switched instead to their main rival, Nvidia. This hiccup didn't do performance any favours for the gamers, as they had to essentially recode a lot of enhancements from scratch. In short, Doom 3 needed a MODERN FAST GPU to run smoothly. 2 year old hardware need not apply. The Radeon 9700, a powerhouse of a card, struggled to maintain 30fps in decent settings. Same with the top end Geforce FX5800 and FX5900. It wasn't until Nvidia released their Geforce 6XXX series, a week after the launch of Doom 3, that gamers could enjoy the engine cap of 60fps implemented in Doom 3 by using the latest greatest 6800 Ultra / 6800 GT. The 6600GT did a commendable job if you sacrificed a few settings. And ATI's X800 series was fine too, although it was clear that iD had truely abandoned optimizing for ATI (thanks to that aforementioned leak), as the 6800 series was at least a tier ahead of the X800's.

I had a 9800 Pro and it ran pretty well. These cards aren't really for gaming. They are for stuff like CAD, CAM etc.

@@dave7244 9800Pro ran at 30-40fps at decent settings. Definitely playable! But no chance of maxing the game out, even at 800x600. The 6600GT did much better, even beating the 9800XT by some margin, despite the mid range Nvidia card being less expensive. 9800Pro beat the 6600GT in HalfLife 2 though! Didn't play well with OpenGL games but it was a DX9 powerhouse.

There is a slider somewhere in the configuration panel for geometry vs texture. I think that just adjust how memory allocation on the card is prioritized, but I'm not sure. I would hope selecting "OpenGL gaming" would set the tweaks for gaming performance. There may yet be registry keys or something that could be adjust to improve performance.

Are you sure it needed 1gb? I played it with a 8800gtx (and a 4200ti and 5900gtx before that) with 768mb and I don't recall that, maybe I'm wrong, It was a long time ago.

I had a kyro 2, love to see a video on that chip.

We are not related that I'm aware of, but... it looks like an interesting channel!

Yup,I remember the GeForce 256 SDR (wich I bought at the time for my Athlon 500 rig),being addressed as a geometry processing unit(gpu). Before that time, my voodoo 2 12 Meg was defined as a 3d add in card ,just like my ATi rage pro 4meg,wich also can be defined as a display adapter. Although my S3 virge 2 meg before that was my real first 3d card. I know it's bad,but seeing dark forces 2 run in 320x200 w/texture filtering hooked me for life..

Gpu come after t&l support . But really a full gpu is a directx 10 after the uni shaders introduction.

Video card was the term I was using back then.

@@seebarry4068 Today, the term "graphic card" is also used, I believe. I think the term "GPU" is more oriented towards referring to the chip itself in a strict sense and, in general, the video card

I remember in the 90s the term graphics accelerator was used a lot.

I started by coding Amiga demos in the early 90s. More recently, I've been working at Intel on the open source (Linux) drivers for their GPUs.

That is a SCSI Zip drive. I used it in my previous video about NEC PC, and I hadn't put it away yet.

@@TalesofWeirdStuff Awesome. Hadn't seen one in ages.

Priorities 😁

800per month or per year? not trying to be condescending, but sadly there are many many countrys where 800per year is more than double/triple(/or more than that even) of their salary per year

I have a suspicion that a strong CPU would help the performance of this card. Doom 3 was demanding on the GPU and the CPU too. The Pentium 4 2.4GHz seems a little low. The minimum system spec was 1.5GHz. We'll find out soon enough.

@@TalesofWeirdStuff I had the 256 bit bus 128mb ram on the card I believe. It was fairly smooth at 1024x768. I had a good air cooler on the chip with heat pipes and a medium overclock. I think it was running in the mid to high 50's FPS from what I remember and we thought that was great at the time. I think the Ram I had in the system was running at 400mhz. That was the first game I played other people online with, I had played games on a network or 1v1 with a modem but not online before this. I was working at a grocery store and saved my money for the parts. The Pentium 4 was real close in actual games from what I remember.

@@TalesofWeirdStuff I had the card with the 256bit bus and 400mhz Ram medium overclock with heatpipe air-cooler. Ran smooth.

Mobility Radeon 9200. Turns out it was 16mb as the macpro which my friend had I remember it having double the ram and those ran 32 mb. Geez.

Wow... you *really* wanted to play that game. Lol.

3dlabs bought Intergraph IP in 2000, and they continued to use the Wildcat branding after that. I don't think it's a continuation of the architecture, but I also don't think it's a continuation of the Gamma / MX architecture from the 3dlabs in-house chips.

@@TalesofWeirdStuff Your knowledge left me speechless. Thanks for your answer ;-)

In fairness... I had to look it up on Wikipedia. :D

@@TalesofWeirdStuff Shhhhhhhh !!!!! 😀

Things were different back then. Rapidly accelerating technology, hardware guys pushing software developers and vice versa, it was more than a 2 horse race still (barely) so more competition. Generational gains in GPU speed was insane. Doom III was groundbreaking, as were all iD games of that era. 2 year old GPU's didn't stand a chance, more than 10 years in todays industry. They were aiming for the GF3 but that boat sailed, as the Ti4600 provided almost 2x the horsepower of the GF3 and even the flagship GF4 didn't stand a chance in Doom III. Geforce FX5800/5900 did an okay job ( playable at least ) as did the Radeon 9700/9800, but neither could come close to running Doom III in max settings at the engine cap of 60fps! More like 25-35fps, until you sacrificed details in settings or dropped to lower resolutions. Doom III didn't run "smooth" on anything until a week after it's release, when Nvidia released the Geforce 6800GT/Ultra. ATI's X800 series also did the game justice, to a lesser extent.

Unfortunately, in 2007, I got a used one, and it was a bad video card! Even the BIOS screen would get corrupted!

I had a fx 5900 at the time and it wasn't too good there either.

@@harrylarkins1310 yeah doom 3 was a late OG xbox /PS2 era game (but on PC) , and the radeon 9000 series and nvidia 5000 series were really kinda prior to that gen , so they struggled with games like doom 3 or Far cry. that were games of the latter half of the PS2 /XBOX era. now the cards that ran them like champs for the time were the nvidia 6000 series and the ATI Radeon 800XT series (not to be confused with AMD's modern XT' line). a year after it came out i actually had got a radeon 850XT that ran doom 3 at 1024x768 setting maxed with frame rates of 50+ which at the time was amazing to me . prior to that card. my frame rate target had always been 35-40 fps 60 fps was a rarity becaue hardware and software was developing so fast from 1996-2005

Yeah man you were lucky to have yours last until the 8000 series, my Powercolor 9700Pro AIW died in 2004.

@@supabass4003 I simply think it was because I had that extra cooling blowing on it.

ATI not AMD. AMD didn't acquire ATI until 2006 so 4 years after this 3DLabs card came out. ATI had been in the graphics business since the late 1980's if I recall correctly.

@@vigilant_1934 who gives a shit, you didnt contribute anything to the discussion what so ever, just came in with your dumb ass superiority complex.

My 9700pro couldn't do that, I think your memory might be a bit fuzzy!

@@supabass4003 Nop for almost 2 years most games run on high I just usually reduced shadows and no AA. I might be wrong is in the resolution, my display maybe was 1024x768. Ps: and off course wasn't 100% 60FPS all the time but was always close to it most of the game... I had more trouble on Far Cry 1 and even when I updated to a ATI X1800GTO they updated Far Cry and could never run it maxed out LOL, was the benchmark before Crysis 😁

It's actually been a part of 3D computer graphics from the start

Sure, but not typically in graphics hardware. Even Pixar's renderman cheated with rasterization tricks. Some sfx houses were doing ray tracing by that point but it was almost entirely on the CPU and usually distributed over a render farm.

@@kungfujesus06 I thought you meant the rendering technique in general.

Erm. Nvidia never said that. It was even Nvidia that teached us of science that happened during the middle ages with strings if I recall. What Nvidia said and did was to give us hardware RT that is also not to be confused with compute RT which is not actual hardware RT. Also. Before you get the "the more you buy the more you save" wrong. No. It is not for gamers. That is the reason why you fail to understand it in the first place. It is for businesses. The more productive the tools are the more profit there is. So you buy more tools to save even more. Duh...

Of course it was. I was waiting on Real time raytracing in games from 2002ish. Was sure sooner or later it will come and I don't have to Puke anymore everytime i see a screenspace reflection disappear when i Look down.

No. They were 3d accelerators.

This is a bit pedantic, and that's why I rarely give people grief about it. But since you brought it up... ;) Back in the day... that would have been called "3D accelerator" as @tenchi71 says. This was to differentiate it from a card like the ET4000 that was just a "video card." At one time even cards like ET4000 and S3 Trio were called out as "Windows accelerator" or "2D accelerator" to differentiate them from more simple VGA cards like Trident 8900. GPU is supposed to be a riff on CPU. The feature that sets CPUs apart from the more primitive logic that preceded it was the ability to run programs. That's why I use that as the difference between a GPU and a 3D accelerator. In all of these cases, the cards in the old category eventually cease to be made, so the old words disappear. When the old words disappear, people start calling everything by the new words... even when it doesn't really apply.

What? Doom 3 ran fine on literally all reasonable agp cards, even the gf4mx which was a rebranded gf2. People even hacked the voodoo 2 driver for it. I know it ran fine on dx8 and dx9 cards. If you can't run it, you're doing something wrong, bad hardware/driver IDK. Also, afaik Radeon had the better agp cards, and switched to bridge chips later than Nvidia. Runs great on the x800 through x1900. X800 doesn't do sm3, but most sm3 games had community sm3 patches.

@@JohnDoe-ip3oq For whatever reason it ran like shit for me. My sister had this cheap ass computer and put in an GF 8100 and got better general performance than my high end 7000 series. I want to say it was a 7950. It sounded like it should have had 2 GPUs, but it didn't. Once I moved over to core 2 with PCI-E I never had anymore issues.

@@awilliams1701 those dual GPU cards were super broken, especially because they were not intended for agp, optimized PCIe for the independent lanes, putting that on a bridge chip and wonky drivers is a guaranteed bad time. SLI was never perfect, but definitely trash until dx10 bare minimum. I would not recommend Nvidia for retro gaming anyway, because Nvidia had garbage hardware until dx10. Dx8 was kinda acceptable if you didn't use anisotropic filtering, but nobody wrote games for it on PC. Xbox did, but even Halo was ported to dx9. Devs did not care to write assembly for dx8 both vendors, and ended up lowest common denominator. Then only ATI had good dx9, and Nvidia yet again ruined games with poor performance and optimization was lowest common denominator until sm3. Which Nvidia had poor performance on again, but developers at that point didn't care and maxed out sm3, best on Radeon. The dx8 Radeon did have driver issues on Windows 98, no problem on XP/dx9. OpenGL was better on Nvidia, but you are basically talking quake 3 and doom 3, which were fine on Radeon. So why Nvidia at all? Maybe some games like splinter cell using Nvidia specific shadow effects. That's about it, and definitely not a good reason for one game, and overall worse performance. Also, anything not Windows 98 doesn't need period accurate hardware, and runs better on modern hardware and windows. Especially if you're using steam, doesn't work on XP anymore. At least a dual GPU agp card should have collectors value on eBay. That's one positive I guess.

@@JohnDoe-ip3oq It had the number indicating dual GPU, but it wasn't. It was single. It was like 7950 or something. But.......I also had a 6800.....or was it a 6600? I don't recall. Neither card was playable on my Athlon XP in doom 3. They were perfectly fine in unreal 3 which looked just as good. So I don't know. lol

@@JohnDoe-ip3oq I do remember when I moved from a geforce 2 gts pro to a 5200 (still in the mindset that is only normal and TI models and having no idea the 5200 was trash). I found the performance was mostly unchanged........except......particle effects. They were a LOT faster. I remember seeing smoke in most games would tank the frame rate.......but not anymore.