Cell benchmarked

[Panajev2001a]

Just about all the tests in that chart are best case scenarios for CELL, especially the encryption, which is very streamable.

Considering single precision is not IEEE compliant, and double precision is only completely compliant if you do some icky hacking (see the IBM SPE ISA manual, p192.), I'm not sure this particular version of CELL has much future in scientific computing.

You are right, but IBM is already at work modifying the SPE FP hardware to replace the crippled DP unit with full speed hardware for customers that needs that kind of performance.

IBM talked at (PC)²
Originally posted: 2005 Oct 11 01:21 PM
missile
Post new reply

Greetings!

Today 'Dr. Juan José Porta' has given a talk at the
'Paderborn Center for Parallel Computing'.
http://wwwcs.uni-paderborn.de/pc2/
http://wwwcs.uni-paderborn.de/pc2/talks/talk135.html
-- thanks for coming!

Beside the things we already know, it *cannot* be
expected to get developer workstations from IBM. He
has referred, more or less, to the PS3 for such things.

The good news: CELL software and libs. are ready to be
released at the end of this month.

The tenor of the talk was that the design of CELL was
mainly driven by Sony and now IBM tries to get the
best out of it (beside games).

There are already some modification going on within
CELL. There is some progress in replacing the entire
FPU inside the SPU with a full-blown DP (double
precision) unit. The estimated performance should be
about 1:2 against the current SP unit, which would be
a major improvement compared to the current situation
of about 1:10 - 1:14.

cu,
Robert

http://www-128.ibm.com/developerwor...123&message=13760013&cat=46&q=DP+FPU#13760013

 

[BlueTsunami]

There are already some modification going on within
CELL. There is some progress in replacing the entire
FPU inside the SPU with a full-blown DP (double
precision) unit. The estimated performance should be
about 1:2 against the current SP unit, which would be
a major improvement compared to the current situation
of about 1:10 - 1:14.

Wow so the new revision willl up the DP performance of Cell to 1:2 SP performance besides the 1:10 - 1:14 they currently have. The reason they didn't first replace or flat out just put a better Double Procision unit (in Cell) a cost reason? or just deemed not important for the certain tasks it most likely to be doing?

 

[Heinrich4]

In this benchmark speaks early 217MVtx/sec per SPU/E, but under how many processed lights, therefore David Kirk (Nvidia) said certain time the SPE could reach about 800 Mvtx/sec?

(217Mvtx/sec = with 2 lights?)

 

[Shifty Geezer]

Emh..no

Sorry, failed to notice these are vector parameters . Also what about the 'inout' term? Does that basically pass the data like a pointer, so q is passed as an object and worked on directly rather than a copy of q's data?

 

[Carl B]

Let me ask you xbdestroya, with Apple going to Intel, how many customers do the 970 have?

970 is dead. IBM is taking care of not getting any benchmark against Power5, for example.

True, IBM really wouldn't have any incentive not to offer the 970 up as a sacrifice for the sake of Cell.

 

[Titanio]

In this benchmark speaks early 217MVtx/sec per SPU/E, but under how many processed lights, therefore David Kirk (Nvidia) said certain time the SPE could reach about 800 Mvtx/sec?

(217Mvtx/sec = with 2 lights?)

Possibly, but I guess we can't know for sure. That detail would have been nice. I guess since it was a TnL demo, though, at least one light needs to be involved. Does your rate halve with every light..?

 

[Npl]

Wow so the new revision willl up the DP performance of Cell to 1:2 SP performance besides the 1:10 - 1:14 they currently have. The reason they didn't first replace or flat out just put a better Double Procision unit (in Cell) a cost reason? or just deemed not important for the certain tasks it most likely to be doing?

both. They opted to get the SPE-die as small as possible, while still having the option of DP ( beeing code-compatible with future iterations where DP is of concern ).

 

[nAo]

Also what about the 'inout' term? Does that basically pass the data like a pointer, so q is passed as an object and worked on directly rather than a copy of q's data?

It's like passing q as a reference (you don't have pointers in Cg).
AS you can see that shader is doing a lot of dynamic branching (there are breaks everywhere ).
Obviously you can't tell from the source code if it's going to pay or not.

 

[Panajev2001a]

Wow so the new revision willl up the DP performance of Cell to 1:2 SP performance besides the 1:10 - 1:14 they currently have. The reason they didn't first replace or flat out just put a better Double Procision unit (in Cell) a cost reason? or just deemed not important for the certain tasks it most likely to be doing?

A mix of both I'd guess.

 

[mckmas8808]

A mix of both I'd guess.

So what would the DP number be raised too?

 

[Carl B]

So what would the DP number be raised too?

Mckmas c'mon man how lazy are you?

The DP ratio would be raised to 1:2, so.... half the SP power. I mean c'mon!

 

[London Geezer]

Mckmas c'mon man how lazy are you?

The DP ratio would be raised to 1:2, so.... half the SP power. I mean c'mon!

How much is that?








/runs

 

[mckmas8808]

Mckmas c'mon man how lazy are you?

The DP ratio would be raised to 1:2, so.... half the SP power. I mean c'mon!

It's the morning my brain is still in 1st gear. So now we are talking about 100+ Gflops. That's pretty good.

 

[Shifty Geezer]

Very good indeed even, if they're IEEE Flops. I'd like to not what the increase in size is though. Presumably quite a hefty jump in size and costs, else why didn't they use improved DP on the current Cell version? Unless it's a simple 'trick' that wasn't obvious but can get greater DP boosts at little extra cost.

 

[aaaaa00]

This line

if( dot( q, q ) > ESCAPE_THRESHOLD )

dot(q,q) = Sqr(q^2) or Sqr(2)*q IIRC, so why do a dot, and why have a qp variable if you're not using it? I'm guessing this ought to be :

if( dot( q, qp ) > ESCAPE_THRESHOLD )

q is a vector not a scalar.

The dot product of a vector with itself is the square of the length of the vector.
He's trying to drop out of the loop when this value is bigger than ESCAPE_THRESHOLD, which is a magic constant he came up with through trial and error.

qp is an inout variable. The result is modified and passed back out of the function, probably for the next iteration.

I still don't see a bug.

Edit: oops nAo too fast for me.

 

[Shifty Geezer]

Yes, my mistake which I already mentioned above
http://www.beyond3d.com/forum/showpost.php?p=639967&postcount=164

 

[aaronspink]

Very good indeed even, if they're IEEE Flops. I'd like to not what the increase in size is though. Presumably quite a hefty jump in size and costs, else why didn't they use improved DP on the current Cell version? Unless it's a simple 'trick' that wasn't obvious but can get greater DP boosts at little extra cost.

A pipelined DP multiplier will run about 2-4x the size of a SP multiplier depending on the techniques used. In addition, an extra cycle or two will be needed as well. Like they'll have to get rid of 2 SPE's to fit within the same die footprint with a fully pipelined DP SPE.

Aaron Spink
speaking for myself inc.

 

[Edge]

A pipelined DP multiplier will run about 2-4x the size of a SP multiplier depending on the techniques used. In addition, an extra cycle or two will be needed as well. Like they'll have to get rid of 2 SPE's to fit within the same die footprint with a fully pipelined DP SPE.

And in a forum labeled "Console Technology", this is important for what?

 

[dantruon]

You know Cell wasnt design to be perfect and excel in all areas, it was design at the start for graphic and visuallisation but more specifically it was design for the PS3, so what is the point of arguing about the DP it pushes when console dont need that to make games.

 

[Carl B]

Guys I have to say, I think we're all interested in the DP performance of Cell's future iterations. Truthfully if this forum *only* discussed matters related to consoles, half the topics would be closed. And unfortunately, it would be the more interesting half!

Cell and Blu-ray, Xenos and unified architectures, these are things that have found their home here in the console forum because the people most interested in them hang out here. Don't think of it as "CONSOLE technology," think of it as "console TECHNOLOGY," for the appropriate emphasis. For example I don't follow this technology due to a love of consoles, I follow these consoles due to a love of technology.