DeltaChrome preview at ExtremTech

[FUDie]

http://www.extremetech.com/article2/0,3973,1417221,00.asp

What a crock...

S3 didn't make mention of any compiler optimizations it is (or isn't) doing on the fly, but S3 has implemented what its driver calls "fast trilinear filtering." According S3, this technique has the memory bandwidth impact as bilinear filtering. S3 apparently uses only four texel samples (rather than the traditional eight), and does some kind of interpolation between them on-chip, rather than fetching eight texels to do traditional trilinear filtering. This could potentially cause textures to be blurred, although we didn't see any noticeable blurring during testing.

ExtremeTech doesn't provide a single screenshot showing image quality, yet they go on about how DeltaChrome wins in UT2003. Let me give you a hint ExtremeTech: If you don't do trilinear, you can go faster in UT2003. My guess is that this "fast trilinear" is nothing more than taking four samples required for bilienar, interpolating one texel from the next lower mipmap and using those 5 texels to compute "trilinear". Good show S3! Also, where's an AF quality comparision?

Lastly, ExtremeTech's UT2003 numbers don't jibe with TechReport's. TechReport shows DeltaChrome at 36.6 at 12x10 without AF yet ExtremeTech shows DeltaChrome at 40.5. The Radeon 9600 Pro results on ExtremeTech seem consistent with AF being enabled, however.

And talk about bias:

S3 is once again sandwiched between nVidia and ATI, and while some of the deltas here are large in terms of percentage, they're small in terms of actual frame.

Uh, percentage is all you can go by. If in one test you're 50% slower, it doesn't matter if it's the difference between 10 and 15 fps or 100 and 150 fps, 50% is 50%.

Get a clue...

-FUDie

 

[Mendel]

firingsquad has a review up too.
http://firingsquad.com/hardware/s3_deltachrome_s8_preview/

UT2003 performance is ok (due to deferred rendering?). Otherwise not very good. Opengl seems like a disaster. Let's hope F1 cards will be better than these S8 models.



Also note extremetech has reviewed with intel equipment while others use AMD. That might make difference with beta drivers too.

 

[Ailuros]

What "deferred rendering"?

 

[nelg]

Now we have trilinear, bilinear, brilinear and introducing halflinear.

 

[FUDie]

Now we have trilinear, bilinear, brilinear and introducing halflinear.

Scratch my idea about taking the four samples for bilinear and generating a texel from the level below... that doesn't work in a trivial fashion because those same four samples are what give you the bilinear texel. Maybe you could do some sort of weighting on each of the samples based on the slope of the pixel, I dunno.

-FUDie

 

[Brandon]

I don't think this setting will come into play until S3 gets AF working in their driver. I played with this and the texture quality setting (I believe they were both set to "auto" by default) and noticed no performance or IQ differences.

 

[FUDie]

I don't think this setting will come into play until S3 gets AF working in their driver. I played with this and the texture quality setting (I believe they were both set to "auto" by default) and noticed no performance or IQ differences.

Well that certainly explains some of the performance anomalies. Now why didn't ExtremeTech notice the lack of aniso?

-FUDie

 

[tEd]

http://www.extremetech.com/article2/0,3973,1417221,00.asp

Let me give you a hint ExtremeTech: If you don't do trilinear, you can go faster in UT2003. My guess is that this "fast trilinear" is nothing more than taking four samples required for bilienar, interpolating one texel from the next lower mipmap and using those 5 texels to compute "trilinear". Good show S3! Also, where's an AF quality comparision?



-FUDie

opengl guy can definitely tell how it's done or at least how it was done on previous s3 graphic hardware

I think they actually take the 8 samples from the same mipmap and using those to do the trilinar effect

 

[bloodbob]

Now we have trilinear, bilinear, brilinear and introducing halflinear.

They are all fine in my book as long as there is a user option to actually use trilinear when requested.

Nice to see that the firing squad regularly checks the future driver approved pages since they are using the 52.16 drivers to compare nvidia cards against ATI and S3 cards.

 

[Nappe1]

hkultala: as I said about 12 hours ago, wait until the shit hits the fan at Beyond3D...

in any case, I am so tired about two big ones using all kind of software specific optimizations and dirty marketing tactics that I am ready to leave 3D features as minority features when deciding next card. PLUS, I haven't played any very 3D intense games during last 6 months( it looks like I am turning to a casual player, I guess.), so for me DeltaChrome S8 performance looks pretty viable option. (plus the price is expected to be near 100 Euros in europe.) on 1024x768 resolution it still seems to be capable keeping framerates on playable level on most games.

and as photographer, I just love the Delta Chrome's ability to set gamma correction values on per monitor basis. (plus their multimonitor features seems a better fit for my needs than ATIs.)

 

[Tridam]

The goal of the trilinear filtering is to hide the boundary between two mipmaps and to reduce texture aliasing (that is also a goal of mipmapping) by sampling from more texels.

If S3 can do the same things internally without using mipmapping and with another filtering technique what's the problem ?

 

[parhelia]

a decrease in image quality perhaps?

 

[Tridam]

There is a little decrease in quality but it's difficult to see it (you should take a look at the firingsquad's screenschots).

Anyway this little decrease is nothing when you look at the decrease brought by Volari drivers (no trilinear, positive LOD, reduced number of mipmaps...).

 

[OpenGL guy]

http://www.extremetech.com/article2/0,3973,1417221,00.asp

Let me give you a hint ExtremeTech: If you don't do trilinear, you can go faster in UT2003. My guess is that this "fast trilinear" is nothing more than taking four samples required for bilienar, interpolating one texel from the next lower mipmap and using those 5 texels to compute "trilinear". Good show S3! Also, where's an AF quality comparision?

opengl guy can definitely tell how it's done or at least how it was done on previous s3 graphic hardware

I think they actually take the 8 samples from the same mipmap and using those to do the trilinar effect

No, eight samples from a single mipmap is not enough for trilinear, you need sixteen in order to generate the four samples from the next smaller mipmap. I have no idea what S3 is doing with the newer hardware, but if the claims that they are only taking four texture samples are true, I find it hard to believe there isn't a large image quality tradeoff.

 

[Hellbinder]

hkultala: as I said about 12 hours ago, wait until the shit hits the fan at Beyond3D...

in any case, I am so tired about two big ones using all kind of software specific optimizations and dirty marketing tactics that I am ready to leave 3D features as minority features when deciding next card. PLUS, I haven't played any very 3D intense games during last 6 months( it looks like I am turning to a casual player, I guess.), so for me DeltaChrome S8 performance looks pretty viable option. (plus the price is expected to be near 100 Euros in europe.) on 1024x768 resolution it still seems to be capable keeping framerates on playable level on most games.

and as photographer, I just love the Delta Chrome's ability to set gamma correction values on per monitor basis. (plus their multimonitor features seems a better fit for my needs than ATIs.)

Two big ones "specific optimizations" and "Marketing tactics".

I would really like to know how you justify rolling ATi up into the same package as Nvidia and all the Bull**** they have and still are pulling.

 

[YeuEmMaiMai]

S3 always had free trilinear with Savage3d and 4 but not with S2K (POS card if you ask me)

 

[Quitch]

hkultala: as I said about 12 hours ago, wait until the shit hits the fan at Beyond3D...

in any case, I am so tired about two big ones using all kind of software specific optimizations and dirty marketing tactics that I am ready to leave 3D features as minority features when deciding next card. PLUS, I haven't played any very 3D intense games during last 6 months( it looks like I am turning to a casual player, I guess.), so for me DeltaChrome S8 performance looks pretty viable option. (plus the price is expected to be near 100 Euros in europe.) on 1024x768 resolution it still seems to be capable keeping framerates on playable level on most games.

and as photographer, I just love the Delta Chrome's ability to set gamma correction values on per monitor basis. (plus their multimonitor features seems a better fit for my needs than ATIs.)

Two big ones "specific optimizations" and "Marketing tactics".

I would really like to know how you justify rolling ATi up into the same package as Nvidia and all the Bull**** they have and still are pulling.

I agree. I am getting tired of seeing this sentiment time and time again, without any justification.

 

[YeuEmMaiMai]

One thing that suprised me about the S8 was it's DX9 performance.....It is extremely good for being alpha drivers and pre-production (engineering sample) board. They even have performance that is superior to Nvidia's in most of the DX9 benchmarks....

 

[Joe DeFuria]

No, eight samples from a single mipmap is not enough for trilinear, you need sixteen in order to generate the four samples from the next smaller mipmap. I have no idea what S3 is doing with the newer hardware, but if the claims that they are only taking four texture samples are true, I find it hard to believe there isn't a large image quality tradeoff.

I believe (someone correct me if I'm wrong), S3 used to do "free trilinear" by taking 4 texels from one mip-map, and autogenerating the texels for the next mip-map, rather than actually reading that map's texels. They still did a full blend.

While not true trilinear, this seems to be one of the better performance / quality trade-offs. Though I suppose there could be some "interesting" artifacts depending on how close the actual mip-map is from the auto-generated one.

And to be clear, I have no idea if this is how DeltaChrome is doing things, so this may be completely irrelevant anyway.

 

[OpenGL guy]

No, eight samples from a single mipmap is not enough for trilinear, you need sixteen in order to generate the four samples from the next smaller mipmap. I have no idea what S3 is doing with the newer hardware, but if the claims that they are only taking four texture samples are true, I find it hard to believe there isn't a large image quality tradeoff.

I believe (someone correct me if I'm wrong), S3 used to do "free trilinear" by taking 4 texels from one mip-map, and autogenerating the texels for the next mip-map, rather than actually reading that map's texels. They still did a full blend.

Ok, I'll correct you You need 16 texels to generate the 4 you need from the next lower mipmap.

While not true trilinear, this seems to be one of the better performance / quality trade-offs. Though I suppose there could be some "interesting" artifacts depending on how close the actual mip-map is from the auto-generated one.

It's true trilinear as long as the mipmaps are generated using box filtering, not an unreasonable assumption.

And to be clear, I have no idea if this is how DeltaChrome is doing things, so this may be completely irrelevant anyway.

Someone said that S3 made some comment about 4 texture samples... that's not going to give you trilinear. I'd really like to see some image quality comparisons.