[RC5] Cell processor documentation

Décio Luiz Gazzoni Filho decio at decpp.net
Sun Aug 28 16:08:12 EDT 2005


On Aug 28, 2005, at 4:23 PM, Taka Khumbartha wrote:

> it seems too powerful for it's
> own good, if you ask me.  think of the military/AI applications.  I
> Robot anyone? =]

I'm not familiar with these kinds of workloads, but I know that Cell  
won't flex its muscle in every application. It relies too much on  
parallelism (both ILP, TLP, and whatever you call the kind of  
parallelism that enables SIMD coding). Previous processors have  
relied on ILP, true, but they had the resources to effectively  
exploit it -- mostly thinking about out-of-order execution here. Cell  
relies on proper scheduling by the compiler/programmer. Moreover, it  
was designed for very high frequencies (David Wang of RealWorldTech  
estimates a 5-8 FO4 critical path) and it shows: even the simplest  
instructions have a 2 cycle latency, while many others have 4-6  
cycles (source: http://www-306.ibm.com/chips/techlib/techlib.nsf/ 
techdocs/E815CC047A60914687256FC000734156/$file/ISSCC-07.4- 
Cell_SPU.PDF). At 3.2 GHz, the frequency rumoured for introduction  
with the PS3, such high latencies are disappointing -- the P4 ALU  
runs at 2 x 3.8 = 7.6 GHz, and could easily reach more if it weren't  
for power dissipation concerns. The huge register file (128  
registers) allows the programmer to mask these latencies, but only if  
the application has that much parallelism available.

Cell is basically a glorified multi-core digital signal processor.  
Not that I'm complaining, but then again, my workloads aren't really  
typical.

Décio
-------------- next part --------------
A non-text attachment was scrubbed...
Name: PGP.sig
Type: application/pgp-signature
Size: 186 bytes
Desc: This is a digitally signed message part
Url : http://lists.distributed.net/pipermail/rc5/attachments/20050828/6c07ea49/PGP.bin


More information about the rc5 mailing list