Mr. Camel (a@b.c) on 8/31/07 wrote:
---------------------------
>I am trying to compare Beckton's and Tukwila's CSI bandwidths per socket (per socket
>bandwidth = PSB for the rest of this post). I know that the following calculations
>will be wrong, but I am hoping that some knowledgeable person will post a reply and correct me.
>
>David's article stated that Beckton's per socket bandwidth will be up to 10X that
>of it's predecessor. I am assuming that the predecessor is Tigerton (or the 45
>nm version of Tigerton). So from currently available information, Tigerton's PSB
>will be 34/4 = 8.5 GB/s. This means that Beckton's PSB will be 85 GB/s. The article
>gives a number for Tukwila's bandwidth but I am not sure if it is the per core bandwith
>or the PSB. If we assume David is referring to the PSB (160 GB/s), the Tukwila's
>platform will have roughly 2X the bandwidth of Beckton's platform.
>
>Somebody please correct me if I'm wrong.
>

Full CSI link a 6.4MT/s delivers 12.8 GB/s in the read direction or 11.4 GB/s in the write direction. With right mixture of reads and writes you probably can extract over 15 GB/s.
If Benton has 3 full links + 2 half links + 2 FBD800 channels than the total bandwith = 15*4+12.8*2 = 85.6 GB/s, i.e. almost exactly 10 times higher than a single 1066 MT/s bus of Tigerton.

Is a compaison meaningful? Of course, it isn't, because the system architectures too different. However that's how raw numbers add up.