radeonr420 said:
I agree that 3 CPUs / 6 cores (and now the "4 or more PowerPCs" is
probably way too much CPU power for a $300~$400 console.
I expect a single CPU, perhaps even a duel-core CPU at most. but not
more than 1 CPU die, or more than 2 cores.
I think duel GPUs (ala NAOMI 2 arcade board) would be *better* than
lots of CPUs. but even that is unlikely.
I assume that raw design of xbox2 will be similar to xbox. That is CPU
and GPU will share one memory pool (UMA), nothing like data throughput
machine like PS2 is or PS3 probably will be.
Given that this is machine intended for playing videogames any kind of
multiprocesoring doesn't make much sense. Even as simple as dual cores
or SMT.
I am not familiar with Naomi but as far as I know it is nothing like
xbox with additional cpu. It is like two xboxs working together on
delivering unified gaming experience. It features two separated pipes
for crunching code much closer in design ideas to Saturn than to Xbox.
Every pipe has its own CPU and own gpu. Parts of hardware which are
shared are main memory and T&L chip which probably had enough
performance to fill both pipes or else every pipe would use its own T&L
chip.
Multiprocesoring in xbox hardware doesn't have much sense for me. Xbox
is GPU centred architecture, meaning that most of videogame code depends
on feature set and strength of GPU; and directx 10 will just strengthen
that idea. CPU is something which process rest of code like AI, I/O,
physics and synchronization of all parts of systems. Even most of sound
is done outside of CPU. Memory is shared between all parts of system so
balanced memory performance is more crucial than any other part of xbox.
Strong CPU and gpu do not make sense if memory is slow. This
automatically means that cache is something of big importance. In normal
situations game which constantly has cache misses will work few times
slower than it should. On xbox this effect is amplified. Just compare
performance of Serious Same on P3 and P3 Celeron. Engine was too big for
128k cache which Celeron had and because of that P3 was much better
performer although P3 and P3C are more or less same chips.Performance of
code on xbox highly depends on size of cache of its cpu(s) especially
because main memory of xbox is not exclusively cpus memory.
Dual cores effectively halves amount of cache on xbox and it halves
amount of memory bandwidth per cpu core. Negative aspects of this
behaviour can be clearly seen on Penitum4 chips where Hyperthreeding
sometimes cripples performance instead improving it. On power4 enabling
second core improved performance only for abut 60%. However power4 is
richer in cache than one xbox CPU will probably ever be. Also computers
built around power4 do not have many plagues which xbox2 will have
(xbox2 = $300, power4 computer = $ X0000 and more) so one cannot expect
60% power improvement with second core on xbox's cpu. And as with
additional core stress on memory will be bigger even fall performance
can occur (UMA remember). Also chip is bigger (higher cost for chip but
also for cooling and main board).
SMT doesn't have any sense. SMT has sense in p4 which is intended for
general processing. Most of time PC struggles with more than one program
and SMT has sense in that kind of situations. However SMT will not
improve performance if one program has big emphasis on one part of CPU
as most of games do. Hyperthreading doesn't improve gaming performance
or rendering and similar stuff. So there are no benefits in gaming code
with SMT.
Multiple cpus are out of question. This is wrong way because it will
kill memory performance and at same time price will skyrocket. And for
what one needs 2 or 3 G5+ cpus in xbox? There is not real use for that
unless R500 is very very very fast gpu which will need extraordinary CPU
to keep its pipes filled. But I doubt that because G5 is excellent
performer.
Maybe they will do something radically different so all I said can be
ignored. I don't know.
I think the basic ATI R500 core will be beefed up much like the NV20
was, into NV2A, which doubled the Geometry & Lighting unit (Vertex
Shader) of the NV20/GeForce 3.
Hopefully Xbox 2 will have very fast performance for VS/PS 2.0 and 3.0
(DX9) - as well as having shaders 4.0 thanks to it being a DX10
machine, much like you said in your post.
Framebuffer integrated with R500 is good idea because in dx10
framebuffer is much more written and read from than in previous dxs.
Also part of integrated memory can hold textures and geometry which will
improve performance of R500 and be kinder towards using memory bandwidth
at same time. How R500 will be tweaked I don't know.
Dreamcast and Gamecube both had very efficient single CPUs for their
day. I think Xbox2 will have a strong single CPU chip, with the
possibility of 2 cores on the one chip.
I agree. Both consoles are fine examples of efficient designs.