Was 360's CPU alpha plan of Cell?

Well they may have some bits removed and some bits added as well. The cache is shared in the X360 CPU so they've saved some trainies there.

 

That's a vanilla Power5 I believe. I remember seeing that pic before on some webpage on gamespot or something, then checked out IBM's media photo pages. Just a Power5. I doubt they'll show the Xenon die anytime soon.

 

Oh just wanted to also add that the 970FX has 58 million transistors and 512KB cache.

 

Does anyone think the 1Mb of cache is enough for the 3 cores? I know that price and performance are a factor here but from some of the things I have read about what "might be" the Xbox 360's CPU may depend on the cache alot. I would think 2Mb or even 3Mb would of been better. I know most would agree that die space and price is a huge factor but couldnt this end up being a big bottleneck when all the cores are being pushed to their max?

How much in transistor count, die size, (heat?) and price would of been added if 2Mb or 3Mb of cache was used? Sorry I have very limited knowledge to cost vs performance etc. in the silicon world

 

Cache takes up a lot of realestate.

The Arstechnica article discussed this some. More cache IS better. But consoles are not desktop PCs. Specifically, 3D gaming does a lot of streaming and the XeCPU has specific features for those. Hannibal gives the example of Quake 3 which played as good on Celerons with no cache simply because the game did not need it.

So while more cache is better, the issue is more of how what is there is used. I think the bigger question is how long before developers really dive into using all the resources on these CPUs. Epic got UE3 up on one PPC core (that PPE in the PS3, which only has 512K cache). So I am not too worried about it. Yeah, you always want more but you have to make concessions at some point. A lot of desktop PCs are shipping with 512K (or even 256K) of cache, so a dedicated closed box for gaming with 1MB for 3 cores does not sound too bad. It only pales next to the top end expensive CPUs. And the return on performance for the extra cost is not worth it. You pay a premium for it (and those are the sacrifices of a $300 box).

 

Actually given the in-order nature of Xenon CPU and its L2 access times, increasing L1 sizes on each core could well have more impact on performance then larger L2.

 

But bigger caches also means slower caches. The question is if the increased hit rate could offset the increased latency. Given that the XeCPU is an in-order part, probably not.

Cheers
Gubbi

 

How big is the L1 cache on each core?

Nite_Hawk

 

The size is probably fine. More is always better, but Xbox doesn't exactly thrash it's cache with 128Kb, and GC with 256Kb's hardly ever sees a miss. PC's and servers see large improvements from increases in cache size largely because the OS and other Apps make memory access patterns unpredictable.

As long as the cache configuration is reasonable (N-way set associative where N is resonably large) it should be fine. The danger with a shared cache is that one poorly behaved process can cause the cache to thrash for all running processes. However it also makes interprocess communication through memory much faster.

The Xenon CPU also as the ability to lock cache sections down, which allows it behave more like a segmented memory system, if that's what the developer wants.

 

I believe the XeCPU cores each have a total of 64K L1 cache each.