Rambus (RMBS)

[cordob]

Hi Paul,

I think the microthreading is indeed an issue and it is for the PS3. This is where it originated no doubt. They have been workign with it (Nvidea, Sony) and have been finding that they are not getting the real bandwidth anywhere near the theoretical bandwidth with the memory due to two issues: latency and granularity.

You concentrate exclusively in your arguments on (max) bandwidth, but there are other parameters which have an effect on the overall performance.

So I think (just a supposition), that Nvidea/Sony would very much like to have microthreading added to XDR. Whether those who have to spend the money redesigning the chips, redoing the masks, documentation etc. (Samsung, Elpida, Toshiba) are willing to do it, remains to be seen.

(We will know for sure within about 6 months, the PS3 demonstrating in a week's time will have "ordinary" XDR which will be interesting enough

You say:
FSB speeds are increasing faster than memory bus speeds.

I assume you meant to say that cpu speeds are increasing faster. The FSB speeds are not increasing all that fast. (maybe they need a Flexio bus like Cell The question is whether for normal PC computing everything is not fast enough anyway as it is and they should not concentrate more on getting the power down and the cost lower.
(AMD has just launched a 1 Watt processor at 1 GHz, maybe that becomes the trend)

I imagine that there is a market for it though. That market will likely be Game consoles, servers, and HDTVs.

I don't think servers need it. below a link of a record sustained performance for a 2U server by Tyan; as you can calculate it's not the memory which holds it back, probably the disks. (sustained bandwidth for 48 hours
http://www.tyan.com/html/pr05_nna.html
The 6.5 gigabit data rate was measured as 96 client computers of Neal Nelson's Multi-Node Test Bed performed FTP "Gets" and "Puts" of multi-megabyte files. The client traffic was evenly spread over six separate gigabit ports in the server machine.
...
"Never before has a single 2U chassis run real application programs like FTP at sustained data rates that are this high," continued Nelson. "When this is viewed in the context of continuous operation for 48 hours and a total of 1 petabit of data transferred this is truly a record-setting accomplishment."

As for HDTV, most HDTV chips now use sdram and some DDR. They don't need the extra bandwidth. TV is actually quite slow. TV also wants to use the cheapest product possible. It's like the car industry, every cent is turned around a few times.

Workstations maybe. They could use the bandwidth, but also need low latency and large capacity memory. (for example for cracking large matrix operations to keep them in memory, think also of second order Laplace eqations etc that sort of thing)
I remember a few years ago somebody on the AMD board who desperately needed something which would support a memory of 2 GBytes, kit speeded up his calculations by several orders of magnitude.

Anyway I am waiting "with bated breath" for the workstation with Cell (and hence XDR), which IBM has apparently promised.

Cheers
Cor