Dot products per second

[MistaPi]

What is dot products per second and how relevant is it to measure performance?

RSX: 33 billion dot products per second (51 billion dot products with CPU)

Xenos: Dot product operations: 24 billion per second or 33.6 billion per second theoretical maximum when summed with CPU operations

Wikipedia.
http://en.wikipedia.org/wiki/Xbox_360#Grap...processing_unit
http://en.wikipedia.org/wiki/RSX_'Reality_Synthesizer'

 

[Graham]

A dot product is a pretty common maths operation. It's basically:

given two vectors, A and B, they each have upto 4 components. {x,y,z,w}.

Dot product is:

dot(A,B) = A.x * B.x + A.y * B.y + A.z * B.z + A.w * B.w;
It's a very useful operation. But is not as common as others.

The problem is you cannot judge overall performance because of this metric. Saying 'oo wow, RSX is 37.5% faster than Xenos' would be ignoring the fact that this is a single metric of *many*. Natrually each chip has strengths and weaknesses. This happens to be a strength of both RSX and Cell. Branching is probably a strength of Xenos/Xenon for example, it all depends on the appication and how it is tailored to the hardware.

Also consider that getting 100% of the theoretical performance out of any of these chips will be pretty much impossible.

 

[inefficient]

The Xenos peak performance is calculated as
48 * 500mhz = 24 billion ops

Not exactly sure how they got the RSX peak. But they must be counting for both pixel and vertex shader pipelines.

(24*2 + 8*2) * 500Mhz = 32 bilion ops.


In any case these numbers are no more useful for comparison purposes than comparing the number of pipelines or clock speed themselves. As this seemingly new number is just derived from them anyway.

 

[nAo]

A dot product is a pretty common maths operation. It's basically:

given two vectors, A and B, they each have upto 4 components. {x,y,z,w}.

Dot product is:

dot(A,B) = A.x * B.x + A.y * B.y + A.z * B.z + A.w * B.w;
It's a very useful operation. But is not as common as others.

Nothing wrong here but we must also say that this is not THE dot product operation, it's just A dot product operation, surely the most common used in our field (the euclidean/canonical dot product)

The problem is you cannot judge overall performance because of this metric. Saying 'oo wow, RSX is 37.5% faster than Xenos' would be ignoring the fact that this is a single metric of *many*. Natrually each chip has strengths and weaknesses. This happens to be a strength of both RSX and Cell. Branching is probably a strength of Xenos/Xenon for example, it all depends on the appication and how it is tailored to the hardware.

agreed, we can't judge overall performance taking in account just the number of dot products per second.

 

[assen]

To quote Dan's Data from an article I'm too lazy to dig up a link for, these kind of comparisons are like comparing two tanks by max rpm of their axles when they are upside down and the treads are removed

 

[MistaPi]

Thanks for the input.