[ben6]

USER INFO
Country: United States

3DMARK2001 PROJECT FILE


PROJECT
Name My Benchmark
Description
Registration Name
Registration Key
3DMark Version 300

DISPLAY
Platform Xabre
CPU Opimization D3D Hardware T&L
Width 1024
Height 768
Depth 32 bit
Z-Buffering 32bit
Texture Format Compressed
Buffering Double
Refresh Rate Default
FSAA Mode None

OPTIONS
Show Title Screens Yes
Continuous Benchmark No
Benchmark Run Count 1
Demo Sounds Enabled Yes
Continuous Demo No
Game Sound Effects Enabled Yes
Game Music Enabled Yes
Game Detail Level Low

RESULTS
3DMark Score 6796
Game 1 Low Detail 124.5
Game 1 High Detail 42.7
Game 2 Low Detail 104.2
Game 2 High Detail 55.8
Game 3 Low Detail 104.2
Game 3 High Detail 50.9
Game 4 Nature 24.0
Fill Rate Single 440.3 MTexels/s
Fill Rate Multi 1622.3 MTexels/s
High Polygon Count 1 25.2 MTriangles/s
High Polygon Count 8 7.8 MTriangles/s
EMBM 43.7 fps
DOT3 91.5 fps
Vertex Shader 49.4 fps
Pixel Shader 28.9
Advanced Pixel Shader DNR
Point Sprites (crashed after APS haven't had a chance to run it)

SYSTEM INFO

System Info Version 2.2
Installation ID 0x00000000
OEM ID

CPU INFO
CENTRAL PROCESSING UNITS
Manufacturer AMD
Family AMD Athlon(tm) XP 1800+
Internal Clock 1.50Ghz
Internal CLock Maximum 1.50 Ghz
External Clock 0Hz
Socket Designation
Type <unknown>
Upgrade <unknown>
Capabilities MMX, CMov, RDTSC, 3DNow!, Extended 3DNow!, SSE
Version N/A
CPUID 0x00000662

CACHES
Level 1
Capacity 128KB
Type <unknown>
Type Details <unknown>
Error Correction Type <unknown>
Associativity <unknown>

Level 2
Capacity 256KB
Type <unknown>
Type Details <unknown>
Error Correction Type <unknown>
Associativity <unknown>

DIRECTX INFO
DirectX Version 8.1

DIRECTDRAWINFO
DirectDraw Version 4.08.01.0881

DISPLAY DEVICES
Description Primary Display Adapter
Manufacturer SIS
Name Xabre
Total Local Video Memory 64MB
Total Local Texture Memory 64MB
Total AGP Memory 8MB
Display Driver sisgr.drv
Display Driver Version 4.13.01.3000
Driver WHQL Certified No
Max Texture Width 2048
Max Texture Height 2048
Max User Clipping Planes 0
Max Active Hardware Lights 10
Max Texture Blending Stages 8
Textures in a Single Pass 4
Vertex Shader Version N/A
Pixel Shader Version 1.3
Max Vertex Blend Matrices 0
Max Texture Coordinates 8
Vendor ID 0x1039
Device ID 0x0330
Sub-System ID 0x03301039
Revision 0x01

Texture Formats
Pixel Format 32bit ARGB [8888]
Pixel Format 32bit RGB [888]
Pixel Format 16-bit RGB [565]
Pixel Format 16bit RGB [555]
Pixel Format ARGB [1555]
Pixel Format ARGB [4444]
Pixel Format FourCC [DXT1]
Pixel Format FourCC {DXT2]
Pixel Format FourCC [DXT3]
Pixel Format FourCC [DXT4]
Pixel Format FourCC [DXT5]
3D Capabilities: Hardware Transformation and Lighting, Positional Lights, Subpixel Accurate Rasterizing, Stencil Buffers, Table Fog, Vertex Fog, W-Fog, Specular Gouraud Shading, Bilinear Filtering, Point Sampling, Trilinear Filtering, Additive Texture Blending, Dot3 Texture Blending, Multiplicative Texture Blending, Subtractive Texture Blending, Environmental Mapped Bump Mapping, Environmental Mapped Bump Mapping with Luminance, Factor Alpha Blending, Vertex Alpha Blending, Texture Alpha Blending, TExture Clamping, Texture Mirroring, Texture Wrapping, Guard Band Support, Rendering to a Window, Mipmap LOD Bias Adjustment, Projected Textures, Volume Textures, Point Primitive Support, Full Screen Anti-Aliasing, DXT Compressed Textures

[Kristof]

What is their clockspeed and how many TMUs do they claim to have ?

Fill Rate Single 440.3 MTexels/s
Fill Rate Multi 1622.3 MTexels/s

This seems to indicate about 4TMUs per pipe.

K-

[mboeller]

What sort of OS did You use? If You use XP, did you find the same problems like tweakup reported, or was the performance the same as with Win2000?


Kristof,

the Xabre400 has 4pipelines with 2TMU's each @ 250 MHz

[CHHAS]

It also seem to indicate 2 pipes at 220 MHz to me.

[ben6]

The OS used for this run was ME. I'll run XP tests after I play a bit of Morrowind with Pixel Shaders on (try it anyway).

Kristof, 4 pipes 2 TMUs a pipe. Not sure on why I get these results for fillrate

[Pete]

How much is this expected to cost? With Radeon 8500 64MB (250/275) going for $100, this has some tough competition.

[Grall]

Total score is quite high for a budget chip and with what one must assume to be pretty immature drivers. My P4@1.7 and GF3 overclocked to 250/550 gives "only" a bit over 7100 with the latest reference drivers.

Nature score wasn't that hot maybe though, but does this thing have true hardware vertex shaders or is it done in software "hardware emulation" a bit like Kyro SE?

No matter, this chip looks to be a winner if released to the right price and without too much in the way of software glitches and driver quirks. If not for any other reason than to light a fire under Nvidia's butt!


*G*

[Hellbinder]

Their pixel shaders are not correctly supported in drivers yet. Thats why the seriouslly low scores in nature....

Considering these are alpha drivers... not to bad.. Id wager that withing a coupple months their card will be on par with an 8500 LE and GF3.

Makes you wonder what they will release next.. Its looking more like they would be totally capable of releasing an ultra high-end part.

Xabre 1000 Anyone

[Kristof]

OK, I am confused here... the numbers and all the different specs for pipes and TMUs I have so far seen do not match up at all.

If the clockspeed is indeed 250MHz for this part then the **numbers** seem to indicate a part with 2 pipelines with 4 texture units.

250 MHZ equals a theoretical fillrate of 500, they achieve 440 which is 88% efficiency which is pretty much an OK ratio. The Multitexture rate is a whopping 1622 out of a theoretical 2000 or an efficiency rate of 81% which is very impressive. Thats the logical 2 pipes with 4 TMUs.

If, however, we assume, the claimed, 4 pipes with 2 texture units then the pixel rate should be 1000 of which they achieve only 440 or a rate of only 44%, which is **very very poor** given that this 3D Mark test is very cache and fb write friendly. The texel rate number still makes sense since assuming 4 pipes with 2 units or 2 pipes with 4 units boils down to the same number.

So we have 2 options :

1. They mixed up the concept of pipe and TMU in all their marketing talk. Meaning they say 4 pipes with 2 TMUs but they mean 2 Pipes with 4 TMUs.
2. Somehow only half their pipelines work when doing single texturing but they do work when multitexturing.
3. Something else.

I personally think that having non-functional pipelines when sending out drivers for review or even preview is a big no-no... so IMHO its much more likely that they mixed up pipes and TMUs... Obviously there is still option 3.

Calls for more investigation IMHO...

K-

[Geeforcer]

Ben can test Kristof's theory by firing up SS and checking whether is allows for more then 2 textures.

[CoolAsAMoose]

Nature score wasn't that hot maybe though, but does this thing have true hardware vertex shaders or is it done in software "hardware emulation" a bit like Kyro SE?

This presentation says:

Vertex Shader
- Programmable to vertex data
- Can be emulated by CPU, P4 especially


I guess it means they run the vertex shaders on the CPU, which makes sense for a low-cost chip, as they in most cases sit inside new computers with fast CPUs.

Makes you wonder what they will release next.. Its looking more like they would be totally capable of releasing an ultra high-end part.

Their official roadmap show a Xabre600 with 275 MHz core, 300 MHz memory in Q3 this year. I have already seen tests of early Xabre400-based cards where they run it @ 270/270 so I guess the Xabre600 might go even higher than 275/300. More interesting though, is that they show the XabreII - with DirectX9 support - as entering mass production in Q4. I wonder what to expect from that puppy? Vertex shaders still on CPU: I would guess so! But with fully PS2.0-compatible pixel pipelines.

If the clockspeed is indeed 250MHz for this part then the **numbers** seem to indicate a part with 2 pipelines with 4 texture units.

Agree. And if you check the available info on xabre.com there is no info given on the number of pipelines present. However, this page includes a page missing from the presentation I linked to earlier. That page is a comparison to the GF4MX and it states Xabre400 has got "4P8T" compared to "2P4T" for the GF4MX.

[arjan de lumens]

Originally Posted by Kristof

So we have 2 options :

1. They mixed up the concept of pipe and TMU in all their marketing talk. Meaning they say 4 pipes with 2 TMUs but they mean 2 Pipes with 4 TMUs.
2. Somehow only half their pipelines work when doing single texturing but they do work when multitexturing.
3. Something else.

My guess: framebuffer bandwidth. When doing single-texturing, the renderer must read framebuffer contents for every pass but the first; when doing multitexturing, there is presumably no need to read the framebuffer. So, assuming 4 textures multitexture chain, the framebuffer bandwidth usage will increase by a factor of more than 4 when going from multitexturing to single-texturing. Texture bandwidth doesn't really have a big effect here, as compressed textures are used.

If this is correct, the card should get a ~1.5x speedup from simply reducing the frame/Z buffers from 32 to 16 bits. Also, I'd expect a substantial performance hit from using 32-bit rather than compressed textures as well.

Another possibility is that the Xabre architecture is very poorly optimized for framebuffer reads, suffering lots of bus turnarounds all the time. This condition may require more specialized tests (beyond those of 3dmark2001) to detect.

[ben6]

running some more tests . Quite interesting performance for their targeted segment really. Unfortunately, I have the feeling that with Ti4200s running $159 and R8500LEs starting to get to the $100 online prices they might have a hard time . But we'll see. There's apparently a lot of interest from people like Gigabyte, PowerColor, Triplex in this chip...

[Kristof]

I noticed this in 3DMark output :

Textures in a Single Pass 4

This either means 4TMUs or 2TMUs with a loopback... so both options are still open...

We need a benchmark fillrate test with little pixel bw usage, say something drawing opaque quads with no Z write or compare, smallest bw usage per pixel (just an always write for colour) and then see if the score goes above 500Mpixels for single texture case.

K-

[Humus]

The fillrate test in GL_EXT_reme would suffice for that.

[Chalnoth]

Originally Posted by Kristof

This either means 4TMUs or 2TMUs with a loopback... so both options are still open...

There would have to be huge memory bandwidth limitations for it to have 2TMU's. I truly doubt this is the case, but in the end, it doesn't really matter. The end result is still the same: quad-texturing performance is 4x that of single-texturing performance.

[Hyp-X]

Originally Posted by Grall

Nature score wasn't that hot maybe though, but does this thing have true hardware vertex shaders or is it done in software "hardware emulation" a bit like Kyro SE?

Well on the said CPU (XP 1800), on a GF3 running the Nature test with software T&L is faster than with hardware.
So this might not be a disaadvantage after all.

[Mintmaster]

I think its very possible that its 4 pipes with 2 TMU's on each. Remember the Geforce2 series? They had worse fill rate than the Radeon, with double the pipes and a higher clock. The Geforce2 Ultra had a theoretical 1000 Mpix/s fillrate, but only achieved about 350 in 3DMark2001. However, if SiS can do many texture per pass using loopback, they can greatly reduce the memory load because the texture bandwidth is very little for these tests, and the framebuffer bandwidth gets distributed over the number of loopback cycles.

SIS is new to high speed of 3D graphics, so they're bound to run into huge memory problems. Since its a value card, it probably wouldn't have much of the fancy texture caches or memory optimizations.

4 pipes with 2 TMU's each makes perfect sense.

[Geeforcer]

Originally Posted by Mintmaster

The Geforce2 Ultra had a theoretical 1000 Mpix/s fillrate, but only achieved about 350 in 3DMark2001.

I think you are confusing Serious Sam and 3Dmark2k1 fillrate tests.

[Mintmaster]

Originally Posted by Geeforcer

I think you are confusing Serious Sam and 3Dmark2k1 fillrate tests.

Nope. 32-bit single texturing fillrate in 3DMark2001, just like what we're discussing with the Xabre:
http://www.digit-life.com/articles/g.../fr-single.gif

The Geforce2 doesn't have loopback, however, so it can't get more than a multiple of 2 going to multitexturing.

[Geeforcer]

Isn't standard test resolution 1024x768?

[Chalnoth]

Originally Posted by Geeforcer

Isn't standard test resolution 1024x768?

Shouldn't matter...the results would still be roughly the same.

[Mintmaster]

Another indication of a rather crappy memory controller is how bad the Xabre is at dependent texture lookups. Look at the EMBM score. The current generation of video cards gets 3 times that score. The same goes for the pixel shader score. I know there is some math in the pixel shader (its a matrix multiply for a cube lookup), but the DOT3 score is fine, indicating a decent math engine.

Again, the argument I gave previously about the fillrate also points to a memory problem. I suppose things can change with another spin of silicon, but this usually just gets rid of bugs. This may be a DX8 video card, but it doesn't look like SiS really cared about a good implementation. The dependant texture lookups are pretty much the best part of pixel shaders. I suppose that'll be good news for RV250 when it comes out, assuming they both attack the same market (as SiS said).