sliderider wrote:

How about i740. I heard that card doesn't use any AGP features, either. 😜

Funny thing too, the PCI versions of that card actually ended up being faster.

For a Socket 7 I probably wouldn't shoot for anything higher than a Voodoo3 or TNT2. The CPU alone is going to be quite a bottleneck.

Pippy P. Poopypants wrote:

sliderider wrote:

How about i740. I heard that card doesn't use any AGP features, either. 😜

Funny thing too, the PCI versions of that card actually ended up being faster.

Yup, that's because, to promote AGP and it's features, Intel decided to pull one of it's typical stunts again. This time by having the i740 use the AGP memory bus for keeping the textures in regular memory. Even though the i740 has 8MB of onboard RAM, it's not even using all of it! It will always use the system RAM instead -_-

As the PCI card couldn't make use of the system RAM, it used it's own onboard RAM which of course is faster.

I never came across the PCI version of that card though

Pippy P. Poopypants wrote:

For a Socket 7 I probably wouldn't shoot for anything higher than a Voodoo3 or TNT2. The CPU alone is going to be quite a bottleneck.

Yup, definitely!

I'd install a V3 only so I don't have to use a separate 2D card and save a PCI slot that way 😜

PCI cards had 12 or 16 MB to prevent texture swapping. Yes those seems rare.
http://cgi.ebay.com/REAL-3D-PCI-Video-Card-SF … =item51923040d1

The AGP i740 board stores all textures in "AGP memory" (that's system RAM) and the framebuffer / z-buffer in its local RAM I think.

The PCI version of the card actually implements an AGP bus on the card so that the chip still thinks it has AGP texture memory. There's a bridge chip to convert PCI to AGP for the 740 chip. It's pretty obvious that this card cost a lot more for them to build.

I would think that trying to use an i740 AGP card on Super 7 would be a bad idea because AGP texturing usually doesn't work correctly on those chipsets. VIA actually disabled it entirely in some of their AGP drivers.

In some ways 740 is an IGP on a card. 810 works very similarly. Some 810 mobos have a small amount of separate RAM used as a local z-buffer/framebuffer. Textures are always stored in system RAM AFAIK. And of course later IGPs again always used system RAM.

I would go for a Riva128 based card,
Elsa Victory Erazor,
Diamond Viper 330
or something similar for a socket 7 system.

i740 really uses on board texture memory, frame buffer is limited to 4MB and whatever more the boards have is for textures. As you can see they even use different memory type for each. Just found out there was even 24MB version which had 8 MB framebuffer, they were not adding memory just for kicks.

Putas wrote:

i740 really uses on board texture memory, frame buffer is limited to 4MB and whatever more the boards have is for textures. As you can see they even use different memory type for each. Just found out there was even 24MB version which had 8 MB framebuffer, they were not adding memory just for kicks.

i740 AGP usually has only 8mb memory on board and does use sideband memory over the AGP bus so it's very slow when fetching textures that are outside of the onboard memory. If you get the PCI version, (very hard to find), then it needs to have more memory onboard because it no longer has access to system memory through AGP. The PCI version is allegedly much faster than the AGP version because it never has to fetch textures from main memory. It's actually faster to swap textures in and out of a smaller amount of onboard memory than it is to have all your unused textures stored in sideband memory. That's the main reason why the i740 AGP was so much slower than other accelerators, even ones that were on the PCI bus because texture swapping over AGP extracts a bigger performance penalty than swapping with onboard memory. Once video cards started selling with enough memory, storing textures in main memory became less and less practical. Video cards that have little or no dedicated memory of their own and rely heavily on system RAM aren't called "vampire video" for nothing.

Duh, vampires. 8 MB was fine in 1998. If some people found the card slow it was for other reasons.

what sliderider says is very interesting from a technical point of view; I never really understood why an Intel tagged 3D accelerator performs so "modestly" amidst the competition... That AGP memory issue must have been one of the main reasons.

On the topic question:

I am a bit puzzled why everybody suggests avoiding semi-modern graphics cards with a K6-2 system. In my exprience, a decent k6-2+ or k6-3+ on a good board can drive a Geforce 4 or Geforce FX 5500 or 5700 pretty well. Of course these cards are a serious overkill performance-wise but if you choose a low-power version which does not need its connectors to be soldered - wired to your PSU's molexes then it is a good solution. A good PSU is still a must, of course.

i always found a gf2 ti-whatever worked great in socket 7. but it tempts you to try games the cpu can't handle brilliantly like quake3.

I would also suggest to use a GF2 based card (not MX). Preferrably a GF2 Pro with 64 MB. The point there is, that you get the TnL advantage, a card that is more than fast enough (GF2 scales very well upto about 1 GHz+ CPU) and GF2 takes only about 18W power. And power draw can be the cause of stability problems on those boards. It also works with AGP 1x,2x.

or go for a Voodoo3. (but check mainboard compatibility with certain graphics chipsets before).

Jolaes76 wrote:

what sliderider says is very interesting from a technical point of view; I never really understood why an Intel tagged 3D accelerator performs so "modestly" amidst the competition... That AGP memory issue must have been one of the main reasons.

On the topic question:

I am a bit puzzled why everybody suggests avoiding semi-modern graphics cards with a K6-2 system. In my exprience, a decent k6-2+ or k6-3+ on a good board can drive a Geforce 4 or Geforce FX 5500 or 5700 pretty well. Of course these cards are a serious overkill performance-wise but if you choose a low-power version which does not need its connectors to be soldered - wired to your PSU's molexes then it is a good solution. A good PSU is still a must, of course.

The thing with sidebanding is it's like the difference between a CPU with on chip cache and a CPU with off chip cache. Assuming two CPU's are identical except the location of the cache, the one with off chip cache will be slower than the one with on chip cache. That's what sidebanding does, it places texture information in main memory, the equivalent of off chip cache. There is also the possibility of a collision with the CPU when both attempt to access main memory at the same time. Both requests can't be processed simultaneously so one or the other has to wait. It doesn't matter which one has to wait because they both will affect your gaming performance. When you have the CPU and GPU fighting with other over memory access that can cause substantial lag because these collisions can happen hundreds of times per second and the faster the CPU and video card are, the more frequently collisions will occur because both will be sending more requests to main memory in the same space of time. That's one of the reasons why onboard video has been horrendously slow in the past. It typically has had little or no physical memory of it's own to keep costs down, so everything goes into main memory. RAM prices have dropped so much in more recent times so that many integrated video solutions come with their own memory now so sideband memory doesn't get used as often but there are still some memory leeches out there, mostly in cheap lowend systems,that draw down system RAM for video.

sliderider wrote:

The thing with sidebanding is it's like the difference between a CPU with on chip cache and a CPU with off chip cache. Assuming two CPU's are identical except the location of the cache, the one with off chip cache will be slower than the one with on chip cache. That's what sidebanding does, it places texture information in main memory, the equivalent of off chip cache. There is also the possibility of a collision with the CPU when both attempt to access main memory at the same time. Both requests can't be processed simultaneously so one or the other has to wait. It doesn't matter which one has to wait because they both will affect your gaming performance. When you have the CPU and GPU fighting with other over memory access that can cause substantial lag because these collisions can happen hundreds of times per second and the faster the CPU and video card are, the more frequently collisions will occur because both will be sending more requests to main memory in the same space of time. That's one of the reasons why onboard video has been horrendously slow in the past. It typically has had little or no physical memory of it's own to keep costs down, so everything goes into main memory. RAM prices have dropped so much in more recent times so that many integrated video solutions come with their own memory now so sideband memory doesn't get used as often but there are still some memory leeches out there, mostly in cheap lowend systems,that draw down system RAM for video.

I kinda knew this already, but your explanation is pretty darn good, it's more detailed 😉

But it is wrong. Sideband is extra adressing channel reducing latency of agp texturing. It is a performance enhancer, just like whole agp texturing. i740 like all the other cards with agp texturing prioritize onboard memory over main memory. You could argue supporting it is a waste of silicon, but you cannot say it hurts the performance.

Putas are you saying that i740 behaves like every other card and uses its local RAM for everything before using AGP texturing?

I recall Matrox G200 being advertised as designed for AGP texturing as well, having AGP DiME and sideband before the competition.

But if you look up i740 AGP it is hard to judge if people are saying that i740 uses only AGP texture memory of it if just uses it better than the rest.

the whole time

Putas wrote:

i740 like all the other cards with agp texturing prioritize onboard memory over main memory.

But I thought i740 didn't even use on-board texturing and -always- used AGP texturing, even if not all on-board memory was even used? This is what makes the PCI version faster then the AGP version...I'm starting to get puzzled 🤣!

That would be true with the early 4MB board with unused memory expansion slot. Anything more means texture memory and PCI versions have lots of it.

I was never impressed by intels card. I had one or two, and I used them to compare with my other cards at the time before putting them in my bits drawer. The only game which looked nice(ish) was monster truck madness. They were ok, but couldn't compete with the cards out at the time. I remember thinking ... if only these were available LAST year. But I got them for free so, I didn't bitch.

Tetrium wrote:

Putas wrote:

i740 like all the other cards with agp texturing prioritize onboard memory over main memory.

But I thought i740 didn't even use on-board texturing and -always- used AGP texturing, even if not all on-board memory was even used? This is what makes the PCI version faster then the AGP version...I'm starting to get puzzled 🤣!

How can a PCI card use AGP texturing? It would have to be plugged into an AGP slot to do that. PCI slots can't do it. The PCI i740 uses a bridge chip to make it electrically compatible to a PCI slot but it still can't implement AGP sidebanding through the PCI slot so the PCI version has to use onboard memory and have more of it.

Here's a quote from the Wiki entry on AGP

"Intel's i740 was explicitly designed to exploit the new AGP feature set. In fact it was designed to texture only from AGP memory, making PCI versions of the board difficult to implement (local board RAM had to emulate AGP memory.)"

So you had to have more memory on board the PCI version in order to emulate AGP sidebanding because the card can't access system memory from a PCI slot. The bridge chip has to use onboard memory for sidebanding, which means more of it is needed than the AGP version.