Sony Cell BE / RSX shrink & integration roadmap

Also, based on all that's written above, I wonder: If a die merger for Cell and RSX is this hard to realize (because of increasing cost, diminishing returns, lay-out difficulties, node & process differences, ownership of the IP, use of different foundries), would combining them in a single package as a SiP be a reasonable and feasible alternative? It would eliminate the need for them to be on the same node and the same process while I suppose it could still give substantial savings in PCB real estate, and consequently at set level, in logistics, etc.?
 
Last edited by a moderator:
Are nodes this strongly linked to the applications of ICs? Have GPUs skipped 45nm in favor of 40nm, and will they skip 32nm to go straight to 28nm? Whereas the Cell as CPU would go from 45nm to 32nm?

It's not so much about the applications, but the foundries producing the devices. TSMC, which presently makes all the high end GPUs, uses the "half-node" shrinks as their flagship products. They had 40nm out very soon after 45, and they skipped 32nm alltogether. So products that come out from their primary production lines are now either 40nm or 28 nm. (They do have an active 45nm for low-power products). On the other hand, the Cell is produced at fabs belonging to the Common Platform group (IBM, GF, Samsung, others), which concentrate on the primary nodes, like 45nm and 32nm.

Note that moving products from a fab family to another is a very complicated and expensive process -- even though they sort of do the same stuff, they use various different technologies, and the differences typically mean that to go from GF to TSMC, you need to redesign you entire chip if you want to keep it fast.

Also, the nominal node does not tell the whole story. As I understand it, the Intel 32nm process is actually tighter for logic (but not SRAM) than the TSMC 28nm, so the names can be deceiving. Of course, Intel's processes generally seem to be built from magical fairy dust -- that's the only way to explain why they are so much better than their peers using the same technologies...
 
Worse in terms of yield?

Worse in terms of how much scaling they get from the shrink. EiB doesn't shrink very well at all. At 45nm, IBM was discussing what they could do with Cell at the 32nm shrink to make it not worthless to shrink it, one suggestion was double the local memory pools to 512k, which would increase performance in local-memory heavy tasks by 60% or more. However, this doesn't apply to PS3s Cell, as it has to stay at the defined performance and spec, which leaves them with all the problems IBM was trying to fix by changing the hardware on Cell, which, at least IBM thought at the time, made it pretty useless to shrink it.

Edit: To elaborate on the first part, here is a excellent article from Real World Tech: http://realworldtech.com/page.cfm?ArticleID=RWT022508002434&p=3

Die size has only gone down by half in shrinks that should have reduced it by 4x, theoretically. Thats half as good as expected normal scaling.
 
Last edited by a moderator:
Stop calling it PSX. Not even the original Playstation was called PSX; that was its pre-release development code-name (or so gaming press rumors say anyway.)

No playstation EVER has been called PSX.
 
Last edited by a moderator:
No playstation EVER has been called PSX.

300px-Console_psx.jpg


http://en.wikipedia.org/wiki/PSX_(DVR)

Tommy McClain
 
Based on all that's written on page 1 of this thread I'm left wondering:

If a die merger for Cell and RSX is this hard to realize (because of increasing cost, diminishing returns, lay-out difficulties, node & process differences, ownership of the IP, use of different foundries), would combining them in a single package as a SiP be a reasonable and feasible alternative?

It would eliminate the need for them to be on the same node and the same process while I suppose it could still give substantial savings in PCB real estate, and consequently at set level (more compact cabinet, less powerful power supply, perhaps no more need for a fan)?
 
No clue still whether a shrink or integration of Cell/RSX is due but at least two hints in firmware v4.20 that a new console design is imminent:
  • [Eject Disc] has been renamed [Remove Disc]. This points at a top loader (like PS2 Slim) as opposed to the current slot loader of PS3 fat and slim.
  • [Hard Disk] has been renamed [System Storage] — a strong suggestion that (some versions of) the next PS3 will use solid state storage instead of an HDD
http://bit.ly/NHxOZE
 
Last edited by a moderator:
That's interesting, and follows a path they've taken before. A switch to flash has been an obvious move, as soon as you could get a large enough pool to be useful for significantly less than a mechanical hard drive. 64GBs has always been the amount to target. That much is probably cheaper than a 2.5 hard drive, and it's also larger than the biggest launch hard drives so it won't seem like a big downgrade. It would be a shame if a new model lost the ability to add a hard drive, though. PS Plus members need all the storage they can get!
 
New model if coming with flash would probably have flash then a hdd bay but no hdd. Seems like it might be reading a bit much into name changes though.
 
A switch to a tray optical drive would be highly disappointing. A top-loaded drive even more so. Never that I'd buy such an abomination, so unless PS4 can play PS3 games I won't be touching my current games library ever again. I sure wish my Euro launch PS3 Phat was still working... :(
 
If rumors of it being announced at gamescon are true it shows how far E3 has fallen on the totem poll.
 
Back
Top