Radeon 8500 faster than 9500!?!

[Tokelil]

Okay the headline might not be all true, but it is in a few benchmarks made by FiringSquad.(http://firingsquad.gamers.com/hardware/catalyst3/)
The question is why?

What I know:
8500 is a 4x2, 9500 4x1. Both at 275 MHz. (If I remember correctly for the 8500)
Both with 128 Bit memory interface.
Then we have more advanced/faster PS/VS, better memory interface, Hyper Z III in the 9500.
When using FSAA the 8500 is using super sampling and 9500 is using multisampling, so the 8500 should have a hard time keeping up.

So looking at the above I would think that they were evenly matched depending on the application. When running an older game with no PS/VS, and 2 textures in most of the scene I would guess the 8500 should win. And when running applications with either advanced PS/VS (if any games has that yet…), with FSAA or a game with large overdraw the 9500 would win.

So going back to the FS article here is a list of who wins what:
Approx. equal:
Serious Sam 2 (With a small win for 8500)
Unreal Tournament Flyby (44 fps for 8500 - 45.6 fps for 9500 at 1600x1200. 9500 wins a bit more at lower resolutions.)

8500:
Quake3 (93.3 fps against 106.6)
Jedi Knight 2 (70.5 vs. 63.4 @ 1600x1200.)

Quake3 AF
UT Botmatch AA/AF

9500:
Comanche 4 (40.4 vs. 20.5 at 1600x1200)
Unreal Tournament Botmatch (28.4 vs. 35.8 at 1600x1200)

Quake3 AA
Quake3 AA/AF
UT Flyby AA/AF

So 8500 is better in the Quake3 engine. Fair enough I guess since it uses multitexturing (as far as I know) and has no PS/VS. (Actually if someone asked me before I read the article I would have said the 9500 would win, hands down!)

What I find odd though is that the 9500 is noticeable faster in UT Botmatch when they are almost equal in Flyby. When AA/AF is on though it is the opposite. Why does the 8500 even stand a chance in this game with AA/AF?

You could ask why I want to compare 9500 against the 8500 and not the 9000 since those are both 4x1 designs. I guess that if we want to see the effects of the new generation of Hyper Z, this would be better…

At first I was a bit surprised to see the 8500 beat the 9500 while loosing so much in other benchmarks. After a bit more inspection it makes sense though. So I was thinking: Can I make any predictions to what an 8x2 design would bring to the table.
Humus is claiming that the second (Please correct me Humus, if that is a wrong “quote”!) TMU would be a waste when we get longer pixelshaders. But what we want is to run games in highres. with FSAA and AF, and would the second TMU not help a lot in FSAA? Even with quite complex pixelshaders?

[BRiT]

I cant seem to find anywhere in the article if the ATI-8500 is 128M or 64M setup, and if the ATI-9500 is a 64M or 128M setup. From a quick look online, most of the 9500s seem to be 64M setups, while ATI moved most of the 8500-non-LE over to 128M. Perhaps that could be the reason.

IMO, I think the drivers just arn't as tuned to the newer architecture of the 9500 as they are already for the aged architecture of the 8500. Notice how the Cat-3 drivers managed to increase some scores by 100% for the 9500! In time, I think the 9500 will beat out the 8500 with better tuned drivers.

[Sharkfood]

I believe the reason can be found on page 2 of the article:

>>"Intel Pentium 4 2.8GHz"<<
...
>>"256MB Mushkin PC3200 (operating at DDR333) SDRAM"<<
...
>>"Windows XP Professional"<<

When you bandwidth starve a P4 and run it with a seriously low amount of memory for Win XP Pro + modern games, the two together are going to have adverse effects on the 9700 Pro/9500 Pro.

I'd be willing to bet the same tests performed on a P4T533-C with 512MB would reflect a totally different set of results.

I've noticed the 9700 Pro/9500 Pro's and the newer Catalyst drivers seem to be much more reliant on system memory bandwidth/bus bandwidth (especially on P4's) than the joe-generic 8500's.

There is no doubt in my mind that 256MB + Win XP PRO + the games listed that bus saturation from swapping and having the CPU memory bandwidth starved are totally pegged throughout the entire runs.

[RussSchultz]

I'm willing to bet the extra system memory won't make a bit of difference.

If it was that CPU limited, the 9500 pro wouldn't have made a difference.

It all boils down to texel rate, of which the 8500 has more (than the plain 9500)

[Dave Baumann]

Well, direct comparisons between 8500 and 9500 cannot really be made since they have fundamental differences, and you don't always know the setup the reviewer used.

For instance, if the 9500 as faster in UT2003 under normal rendering that could be due to the architectural differences - 8500 only has a 2 level heirarcical Z-Buffer, with no early Z reject on the full resolution Z buffer, whereas R300 (hence 9500) has a 3 level heirarcical Z buffer with early Z reject on the full buffer. This could quite easily account for why 9500 is faster under nrmal rendering.

Alternatively with Aniso, do we know the settings? 8500 always used Bilinear Aniso, whereas 9500 does both Tri or Bi dependant on what is selected in the drivers. Its quite possible the reviewer was comparing Bilinear Aniso on the 8500 to Tri on the 9500.

[martrox]

I think we can all say the performance differences between the 8500 and 9500 are pretty small.... in fact the biggest differences are MSAA, FSAA & DX9. It has been stated that, of all the r300 based products, that the plain 9500 is the least likely to be around very long. In fact, the plain 9500 is just a poor product, period. Do yourself a big favor & spend the small tuppence difference and get the 9500 Pro.... which is (at retail prices) the best deal among the r300 family......

[Bjorn]

I think we can all say the performance differences between the 8500 and 9500 are pretty small.... in fact the biggest differences are MSAA, FSAA & DX9. It has been stated that, of all the r300 based products, that the plain 9500 is the least likely to be around very long. In fact, the plain 9500 is just a poor product, period. Do yourself a big favor & spend the small tuppence difference and get the 9500 Pro.... which is (at retail prices) the best deal among the r300 family......

I agree with you about getting the 9500 PRO instead. But i wouldn't say that the plain 9500 is a poor product.
As you say, the biggest difference are in MSAA/FSAA and DX9. But aren't those the reasons to buy a new card these days ?

From firingsquad's test:

Quake 3 High quality, 1024*768, 4X FSAA, 8X Aniso:

9500 108.9 fps
8500 36.5 fps

That's more then a big difference imo.

[martrox]

When there's a $30.00 difference between a 64meg 9500 and a 128 meg 9500Pro - at the retail level, $179.00 vs. $209.00 - you have to be totally unaware of the difference to even consider the 9500. In fact, I would probably recommend a nVidia :shock: GF4 TI4200 or Radeon 8500 over the 9500. They are both just much better buys, IMO.

[Bjorn]

When there's a $30.00 difference between a 64meg 9500 and a 128 meg 9500Pro - at the retail level, $179.00 vs. $209.00 - you have to be totally unaware of the difference to even consider the 9500. In fact, I would probably recommend a nVidia :shock: GF4 TI4200 or Radeon 8500 over the 9500. They are both just much better buys, IMO.

As i said, i agree with you about getting the PRO instead.
It's not a good buy because of the small price difference up to the 9500 PRO. But it's still more then twice as fast as the 8500, not only in a few benchmarks but in the most important benchmarks imo, FSAA + aniso.

And for a comparision with the ti4200 (firingsquad again):

Quake3, 1600*1200, 4X FSAA:

Ti 4200: 34.3 FPS
R9500: 54.3 fps (26.1 with Cat 2.4 drivers)

Should be even more of a difference with with Aniso added.
And of course, you'll get a DX9 card also.

[Magnum PI]

knowing the 9500 pro is only 30 dollars more than the 9500, and having double fillrate (8 pipes instead of 4), who would buy a 9500 anyway ?

so who care about how perform a 9500 compared to a 8500..

[Bjorn]

knowing the 9500 pro is only 30 dollars more than the 9500, and having double fillrate (8 pipes instead of 4), who would buy a 9500 anyway ?

so who care about how perform a 9500 compared to a 8500..

I basically agree. They need to cut 20-30$ or so from the price to make it a bit more attractive with regards to the 9500 PRO. But i just don't want to paint the picture as black as some people here does. Besides, the 9500 PRO might have double the fillrate but it's far from twice as fast. Especially under more bandwidth limited situations.

[Tokelil]

knowing the 9500 pro is only 30 dollars more than the 9500, and having double fillrate (8 pipes instead of 4), who would buy a 9500 anyway ?

so who care about how perform a 9500 compared to a 8500..
I didn't ask because I was going to buy either of them! (Though 9500 Pro isn't available in DK yet!! :x )

I was just wondering why the scores between them are as posted at FS, and if we would still and in the future gain anything from the second TMU. Especially in FSAA or AF.

[Joe DeFuria]

I basically agree. They need to cut 20-30$ or so from the price to make it a bit more attractive with regards to the 9500 PRO.

I agree, but I'm betting the 9500 non-pro is priced where it is because they can't make any money on it at lower prices. The current 9500 has both an expensive chip, and an expensive PCB.

This is why I think the 9500 (as we know it) will soon disappear. Either the product itself will go away, or maybe we'll see the "9500 Version 2"...which might be same exact specs, but using an RV-350 chip (speculation: 4 pipe R-300), on the 9500 Pro PCB.

[Bjorn]

I agree, but I'm betting the 9500 non-pro is priced where it is because they can't make any money on it at lower prices. The current 9500 has both an expensive chip, and an expensive PCB.

This is why I think the 9500 (as we know it) will soon disappear. Either the product itself will go away, or maybe we'll see the "9500 Version 2"...which might be same exact specs, but using an RV-350 chip (speculation: 4 pipe R-300), on the 9500 Pro PCB.

I agree. As you said, might be the exact same specs but with the 4 pipes physically removed also.

[Dave Baumann]

I agree, but I'm betting the 9500 non-pro is priced where it is because they can't make any money on it at lower prices. The current 9500 has both an expensive chip, and an expensive PCB.

Well, dependant on the pipeline failure rate they are at least making some money as opposed to none.

[Randell]

Anybody considered either mistakes in the FS numbers e.g. swopped around mistakenly? Or driver bugs, I'm sure because AA doenst work in UT2003 on the 8500 its probably turned off by thedrivers so those UT2003 Botmatch scores arent true.

People may buy a 9500 becasue they cant get a 9500Pro at the moment.

[Nagorak]

The 9500 sucks, I agree, but don't disregard it out of hand. Remember soon you'll be able to change your 9500 128 MB into a 9700. :twisted:

Actually, I take it back, I think it's impressive that then 9500 can be just as fast as the 8500 even though it's a 4*1.

[cellarboy]

I'm willing to bet the extra system memory won't make a bit of difference.

If it was that CPU limited, the 9500 pro wouldn't have made a difference.

It all boils down to texel rate, of which the 8500 has more (than the plain 9500)

I can tell you from personal experience that going from 256 to 512MB of RAM under WinXP had a MASSIVE effect on framerates with my 9500pro. Completely subjective, but UT2K3 and Battlefield 1942 doubled in performance.

[Chalnoth]

I'll second that. I was getting pretty significant virtual memory thrashing with only 256MB. Of course, it does depend on how much you have running in the background, but I'd rather not turn things off just for games.

[YeuEmMaiMai]

the 9500 Pro is a worthy upgrade from a 8500 when you run newer games such as UT 2K 3