Having listened to the whole thing now, here's the edited highlights including verbatim quotes where applicable too so people don't need ot listen to it to understand the wierd things said.
DISCLAIMER : Before I say any more, I'm not on an MS bashing mission! Some stuff they pointed out was valid, like the saving on bandwidth of eDRAM or SPE's not having a branch predictor. I'm including valid points they raised (ignoring some stuff already known) and points of these I question, but also their total baloney.
About Cell's SPE's :
"They're double precision math is extremely slow"
At first that strikes me as total bunk. 20+ DP GFlops for a single chip is
good, let alone
extremely slow. But now I wonder if taken as a single core, how does SPE's DP performance compare with a core of PPC or Pentium 4 maybe? Anyone got any comparisons for DP performance?
They explained their take on SPE's not having any cache.
SPE's have no cache, only 256 kb local RAM. In a CPU you write to main RAM and the data is stored in cache in case you need it again. SPEs don't have this. SPEs can only read/write to local ram and if they want data that's not in the local store, they have to fetch from main RAM.
"It's almost as if the main memory is a Filesystem and you have to read from it and write to it when your done...For General programming this is pretty terrible."
This is first class bunk. It's the same with any processor. If the data isn't available on the fast local storage you have to fetch it from main memory. And in this respect Cell ahs the advantage as it has more local memory than then XeCPU. Don't know
what they're talking about here!
Asked what it meant for gamers they replied
perhaps longer development times, or it could mean the SPE's are underutilised.
Talking about eDRAM on Xenos
"We're able with that to get, to...to get all the way up to 256 Gigabytes a second to main memory."
There it is. 256 GB/s to main memory according to these guys
To be fair I think it more likely they don't interview well than set out to present such false info as this (I hope!)
Regards HDTV rendering
the interviewees said that 1080i renders evey 1080 lines every frame in virtually every case. The only difference is the little chip at the end and whether it's sending 520 lines per field or 1080.
They basically say 1080i and 1080p rendering is exactly the same resolution/throughput requirement, just a bit different output whether you send interleaved fields or the whole frame. In this respect I guess they assume 1080p is 30fps, in which case they're right.
Regards system performance:
MS took a very conservative approach on the 1 teraflop system performance.
"Again our teraflop was extremely conservative."
This was a funny part of the interview as they said teraflop figures was basically stretching the truth until you were making them up!
About Launch titles and exploiting hardware.
They observed the programmaing challenges. For XB360 they described development beginning on a single thread. Then when you want more power you could take some code from out of your thread and port it to one of the other cores. Being symmetric that wasn't a problem. On Cell you need to write for asymmetric cores.
I understand this and it makes sense, though would development really work that way, porting bits of code? Shouldn't multithreaded code be designed as such from the ground up?
Regards what was shown at E3
They said demos were running on Alpha kit Dual G5
"and it has the ATi R300 on it."
All in all, there was plenty of wrong information and total nonsense in with the truths. I don't know if it's just they don't know what they're talking about or if they actually are trying to mislead people, but seriously, the complaints about Cell's local storage? Writing data to and from system RAM being bad for general purpose code? What are they on?!
