XDR, GDDR3 and Latency.

[Urian]

I don´t know if this was discussed anytime in the past inside the forums, but something came to my curious mind.

Forgot for a moment the SPE and 2 extra cores of the Xenon and imagine a situation where we only have the PPE as the main CPU of the PS3 and one of the Xenon cores as the main CPU of 360, memory continue being the same.

Which type of memory will affect more the performance of both solutions XDR or GDDR3?

 

[archie4oz]

I don´t know if this was discussed anytime in the past inside the forums, but something came to my curious mind.

Forgot for a moment the SPE and 2 extra cores of the Xenon and imagine a situation where we only have the PPE as the main CPU of the PS3 and one of the Xenon cores as the main CPU of 360, memory continue being the same.

Which type of memory will affect more the performance of both solutions XDR or GDDR3?

You're still not eliminatiing all the variables. The PPE benefits from having it's memory controller on-board whereas on the 360 you're traversing an FSB to an external memory controller. OTOH you've got more L2 on the 360, so it's still not a strict memory comparison...

 

[Rolf N]

These processors have stupidly long cache lines (128 bytes), which masks the effects of latency and shifts the prime memory characteristic more in the bandwidth direction.

IOW whenever either of these processors needs just one uncached byte from main memory, they'll initiate a burst transfer of at least 128 bytes into their caches. You really don't count the latency unitl the first byte arrives, but the latency until the cache-line is filled. And with long cache lines, bandwidth dominates this measure. It's the same old Pentium 4 trick.

 

[phat]

You really don't count the latency unitl the first byte arrives, but the latency until the cache-line is filled. And with long cache lines, bandwidth dominates this measure. It's the same old Pentium 4 trick.

Aren't you forgetting about critical-word-first?

 

[Gubbi]

Aren't you forgetting about critical-word-first?

Yeah, the only way the time to fill the cacheline would impact latency is when you have back-to-back cache refills.

The reason why there seems to be strong correlation between bandwidth and latency is simply because most benchmark suites use constant stride access patterns which with modern prefetchers turns a latency bound test into a bandwidth bound test.

Cheers

 

[Rolf N]

Aren't you forgetting about critical-word-first?

Yes
However, as by design the long lines will need to be filled, the first miss would be accelerated by chf, but the next access can only be made after that is all done. Of course, all else being equal, having a shorter memory latency is always preferable, however it wouldn't improve performance with Cell nor Waternoose nearly as much as it would with, say, a Pentium III.

In practice these systems should saturate their memory buses almost all the time. If you're not ordering and anticipating memory accesses well enough to cover most of the misses up with a prefetch, you're not going to get satisfactory performance either way. Closed boxes let you work like that. But that is orthogonal to the theoretical improvements of lower memory latency.

 

[MBDF]

Is Cell in a better position having direct access to it own XDR memory pool than if it was dropped into the X360 and replaced Xenon... this was the argument for split memory pools, was it not?

 

[N3xtG3nGam3r]

As far as i know (which is not entirely far, lol)...

...using GDDR3 for the Cell, instead of the XDR memory, would basically make the architechture and strenghts of the Cell obsolete. GDDR3 would cause a bottleneck, would it not? I dont think the case is the same for the 360 however. If GDDR3 does cause a bottleneck on the 360 bandwidth wise for the CPU, i couldnt it imagine it hurting it significantly, or making the processor architechture obsolete.

Also, there is EDRAM which has a 256 gb/s bandwidth on 360, and that can be used for pretty important stuff.

However, to answer your question, i dont think the Cell could perform well enough without the XDR, and therefore, is why it was introduced into the PS3.