'8 rendering pipelines' + '4 Pixel Shaders' = Volari V8?

[Pete]

Saw this at AT: Club 3D is practically stealth launching Volari V8 Duos, apparently.

Now, remind me again how a GPU with "8 rendering pipelines" can have only "4 Pixel Shaders?" :? Is this Xabre redux, with marketing taking liberties the hardware can't afford?

As I said in the AT comments, I'm all for more competitors in the consumer 3D arena. Perhaps I'm being unrealistic by expecting another nV/ATi to spring forth fully formed, even on the second try. I wish the Volari series good luck, provided they offer a good card (marketing aside).

 

[demalion]

Nestled in some threads with further discussion of the Volari, I offer these posts related to your question: my initial impression and my eventual conclusion, as well as a collection of performance characteristic comparisons.

 

[Ailuros]

I've no idea about the real unit count and arrangements, yet early single texturing tests point rather to a 2*4 setup per chip than anything else. Beyond that is anyone's guess.

 

[Nappe1]

is it possible that they have same kind of arrangements as PS2 had?

Like PS2 had 11 pipelines, but only 4 of them have texture samplers.... in this particular case, if there's PS stuff running, only 4 pipelines are used, but for older applications and for stencil fill, there's 8 pipelines per chip. (in dual chip mode, that would mean 16 DX7 pixel pipelines and 8 DX8/DX9 pixel pipelines.)

 

[compres]

If that's really the case then looks like the Volari is going to be DOA.

 

[Robbitop]

AFAIK it has been proven that Volari V8 is an 4x2 Architecture.
But not all Pipelines are "Fullprescision". Half of them are Fixed Function (w/out PS FPUs)...so they save a few million of transistors.
A fixed function pipeline costs about 3 Million Transistors; a Full Prescision Pipe costs about 10Mio Transistors.

A Vertexshader costs about 7-10Mio Transistors. Crossbar and Early Z costs about 20 Mio Transistors.
V8 got about 80Mio Transistors.
So we got 4x8Mio (2D Core dunno) Transistors for VS +20Mio Transistors for LMA Pendant ...so there are about 30-35Mio Transistors left for Pipelines and 2D Core. And when I look at PS Performance I would say we have 2 Fullprescision Pipes and 2 Fixed function.

just my 2 cents
(sorry its a bit confusing and my english is not very well)

 

[parhelia]

Interesting speculation but... where did you hear it has 80 millions transistors???? I found no such info anywhere.

 

[Arun]

But not all Pipelines are "Fullprescision". Half of them are Fixed Function (w/out PS FPUs)...so they save a few million of transistors.
A fixed function pipeline costs about 3 Million Transistors; a Full Prescision Pipe costs about 10Mio Transistors.

Hmm, so 26M for the 4 pipelines? That sounds pretty darn icky; that's less than on a fricking GF3 IIRC!
But then again, at identical clocks, a single V8 would get beaten by a GF3 in everything pixel-related I guess

A Vertexshader costs about 7-10Mio Transistors [...] So we got 4x8Mio [...] Transistors for VS[/quote]

I don't know how good the V8's VS units are, but on a NV3x, the VS units cost 9M transistors each. And there are 3 units on the NV30, NV31, NV35 and NV36. And the NV34 has only one unit.
Yes, I know what you're saying: The NV31 only has 1 VS unit.
Well, AFAIK, it has 3, but 2 are bugged. Yes, that's approximatively 15M transistors wasted if they're as good as on the NV30.
Yes, the NV31 is one of the most buggy chips in history I guess: LMA stuff is broken, VS units are broken, PS units are as icky as on the NV30, and so on...

With possibly as much as 20M transistors bugged or highly innefficient, it's no surprise the RV350 with 10-15M less transistors still beats the hell out of it in most cases...


Uttar

 

[arjan de lumens]

Interesting speculation but... where did you hear it has 80 millions transistors???? I found no such info anywhere.

FWIW, http://www.tomshardware.com/graphic/20031107/xgi-volari-01.html states 80 million transistors per chip; apparently the chip is physically very large for the process/transistor count too, whatever that portends.

 

[Pete]

Thanks much for the enlightening posts, all. Guess I skimmed over or just missed those previous Volari threads. I assumed the card wouldn't be 8x1 per chip all the time, but I didn't realize there was some proof already.

dem, thanks for the links. Robbitop, thanks for the analysis. Uttar, thanks for the NV31 "bonus info."

I'm really interested in seeing full benchmarks with current drivers. Surely some sites must have newer review cards by now, considering Club3D is already shipping cards?

 

[Paul]

Like PS2 had 11 pipelines, but only 4 of them have texture samplers....

16 pipes * 150mhz = 2400M fillrate.

 

[Grall]

Like PS2 had 11 pipelines

11? Where did you get that weird number from?

PS2 has 16 pixel pipes, half of them capable of texturemapping and the other half not.

It's not 11 pipes total (which is the most weird-ass figure I ever heard mentioned about PS2 ), and not 4 pipes with texturemapping. 16/8 respectively is the real number.


*G*

 

[Nappe1]

Like PS2 had 11 pipelines

11? Where did you get that weird number from?

from my ass. (yeah, it's pretty wierd that guy like me have donkey... it even doesn't remember PS2 pipeline count right.)

I am just so f*cked up that I even don't want to start arguing if it was the point of my message if PS2 has 11 pipes or more...

oh god, let this unfortunate year end soon. I just can't stand all of this much longer. (not that this would be anybody elses fault. it is just me me me. just like always.)

 

[Ailuros]

AFAIK it has been proven that Volari V8 is an 4x2 Architecture.

Feel free to show me where it's been proven considering the following (yes those are duo V8 scores from PCGH-Thilo):

Fillrate Tester
--------------------------
Display adapter: XGI Volari Family
Driver version: 6.14.10.1000
Display mode: 1024x768 A8R8G8B8 85Hz
Z-Buffer format: D32
--------------------------

FFP - Pure fillrate - 2822.892822M pixels/sec
FFP - Z pixel rate - 1713.361938M pixels/sec
FFP - Single texture - 495.003021M pixels/sec
FFP - Dual texture - 541.937439M pixels/sec
FFP - Triple texture - 521.537720M pixels/sec
FFP - Quad texture - 269.535889M pixels/sec
PS 1.1 - Simple - 1349.516235M pixels/sec
PS 1.4 - Simple - 361.774017M pixels/sec
PS 2.0 - Simple - 835.204102M pixels/sec
PS 2.0 PP - Simple - 835.240173M pixels/sec
PS 2.0 - Longer - 483.130280M pixels/sec
PS 2.0 PP - Longer - 483.130371M pixels/sec
PS 2.0 - Longer 4 Registers - 483.124023M pixels/sec
PS 2.0 PP - Longer 4 Registers - 483.122162M pixels/sec
PS 2.0 - Per Pixel Lighting - 38.662209M pixels/sec
PS 2.0 PP - Per Pixel Lighting - 38.663109M pixels/sec

edit: one more for kicks

GL EXTREME V8 Duo

Results:
--------

Overdraw/HSR:
-------------
Overdraw factor 3, back to front: 404.39 fps
Overdraw factor 3, front to back: 787.08 fps
Overdraw factor 3, random order: 548.29 fps

Overdraw factor 8, back to front: 170.49 fps
Overdraw factor 8, front to back: 599.79 fps
Overdraw factor 8, random order: 360.74 fps

Fillrate:
---------
Pixel fillrate: 1656.39 MegaPixels / s
Texel fillrate: 1087.80 MegaTexels / s

T&L/High polygon count static display list:
-------------------------------------------
Pure transform: 20252723 vertices / s
2 point lights: 16240694 vertices / s
8 point lights: 16228193 vertices / s
2 directional lights: 16248049 vertices / s
8 directional lights: 16230424 vertices / s

High memory bandwidth load/texture cache efficiency:
----------------------------------------------------
One 1024x1024x32 texture: 132.79 fps
Four 1024x1024x32 textures: 131.29 fps

 

[StealthHawk]

AFAIK it has been proven that Volari V8 is an 4x2 Architecture.

Isn't the Volari V8 a 8X2 architecture?

 

[Ailuros]

16 TMUs is the only thing that's "certain" at this point.

 

[vb]

16 TMUs between two V8 chips; 8 TMUs for single.

 

[Simon F]

FWIW, http://www.tomshardware.com/graphic/20031107/xgi-volari-01.html states 80 million transistors per chip; apparently the chip is physically very large for the process/transistor count too, whatever that portends.

It could mean that the chip has a lot of pins as these generally need considerably more area to support. Given that they can connect two chips together (and hence may need a communication bus), this may be the case.

 

[Dio]

When designing a chip a transistor or gate count isn't an accurate measure of size. A particular pattern of transistors or gates may be more or less dense than a different pattern.

This isn't necessarily an indication of 'bad design' - there can be good reasons for using less dense patterns (e.g. noise immunity, heat).

 

[arjan de lumens]

The interconnect between the V8 chips has a bandwidth of only 2.13 GB/s, which shouldn't add more than ~40-60 pins to the total, hardly enough to give a chip nearly twice the size of the RV360 at only marginally larger transistor counts, especially since the V8 comes in a flip-chip package. Noise and heat may be more valid concerns though, given how large the power supplies and heatsinks of the Volari Duo are.