Is the PS3 on track to deliver full specs?
- Thread starter Rash'
- Start date
Yeah... I can't help but think that they are still finalizing the last few options. The Mhz reduction and the rumored PS3 secret for instance are but 2 more examples.
Well, as I understand it, there's 128bit to GDDR3, and then there's 128bit to XDR2. Of course, I could be wrong ...
And how that relates to 64GB/sec and if that number is correct, I have no idea about.
mckmas8808
Legend
nAo can you tell us if this new secert actually exist? Is it fake?
MasterDisaster
Regular
He is under NDA
but I believe that there's no secrets, just a way to hype the people after the last news
mckmas8808
Legend
Yeah but isn't it public knowledge that a secert is still left to be seen with the PS3. I thought he could confirm that that secert exist, while not telling us what the secert is.
ShootMyMonkey
Veteran
The XDR controller on Cell is 64bit, not 128. And the same applies to the FlexIO connection between them. FlexIO by nature is divided into 8-bit lanes (which is apparently the finest granularity you can get at the controller, if you believe Rambus), so 40 GB/sec with 35 GB/sec dedicated to RSX and 5 GB/sec to the SB (as the specs imply) suggests that each lane gives you 5 GB/sec... which in turn implies that RSX's connection to XDR is 56-bit. I'm sure more than a handful of people here should have arrived at the same conclusion.
In any case -- 700 MHz * 2 (DDR) * 128 bit = 22.4 GB/sec, vs. 650 * 2 * 128bit = 20.8 GB/sec (7.14% drop). And of course, assuming that it does not affect the FlexIO at all, you've got total bandwidth going from 57.4 GB/sec down to 55.8 GB/sec (2.78% drop). Other than nAo apparently using 8-bit FP precision, nothing really surprising.
What?!?
ShootMyMonkey
Veteran
However he got his 64 GB/sec, it seems he took the 2.5% you posted as the exact value.
ShootMyMonkey
Veteran
5 mantissa (effectively 6 with normalization), 2 exponent, 1 sign, then.i'v heard it was only 6bits...
Eeek! no5 mantissa (effectively 6 with normalization), 2 exponent, 1 sign, then.
In 4 bytes 3 positive values are packed, 2 are stored as 8 bit fixed point values, the last one is stored in a kind of custom-fp format (since it's a logarithm) which uses 4.5 bits for the exponent and 11.5 for the mantissa.
Marco
From our point of view (as developers) there was never a downgrade. The devkits were gradually upgrading all the time. So, you see, if there was anything short of promised , we just never got to use it.
Like in this post?
http://www.beyond3d.com/forum/showpost.php?p=767079&postcount=13
Anyway, I'm looking forward to seeing the final hardware layout.
so your ARE basicly CONFIRMING the gpu speed is dropped from the initial 550mhz figure...
Angelcurio
Regular
Did you miss the "IF" in his comments?so your ARE basicly CONFIRMING the gpu speed is dropped from the initial 550mhz figure...
scooby_dooby
Legend
But when you're developing, don't you target certain specs knowing that they will eventually be delivered? I mean if you were counting on a 22% increase in GPU clockspeed, from 450 to 550, but only recieved a 11% increase couldn't that sort of throw a wrench into the plans? Especially if you're really pressed for time?
We might develop with some target in mind, but (speaking personally here) I never expect an explicit X-percent increase based on some promised clock change. There are too many variables for me to be able to comfortably predict exact performance until someone drops a bit of hardware onto my desk that does what it's supposed to do that I can test on.
These days enough things are scalable that if more performance appears from somewhere, it can be put to use. If not, never mind.
It is kind of nice to know that more performance will be available, because it gives you a little headroom should designers or artists get a bit crazy, bu I never tell them that.
developer's first rule: keep the artists at bay.
