Beyond2.5D Stacking, Interposer, lego

I'm not sure I understand your statement here, if Sony wanted ddr4 they would've gone with ddr4. There's no guarantee that ddr4 will be less expensive than gddr5 and there's certainly no guarantee that ddr4 in a superwide bus will be cheaper to implement than gddr5.

What Sony wants is the most bandwidth possible for a reasonable price that they can launch with in 2013.
Stacked memory offers the best option for bandwidth and price over the long term but it looks too risky to launch in 2013 with. DDR4 will do well price-wise over the long term, but it's a low bandwidth option and it availability in 2013 at volume is a risk. GDDR5 offers very good bandwidth and is available in quantity for 2013. Price should be managable via contract but no cost savings over the life of the console.
Overall, I don't think GDDR5 wasn't that hard of a choice for them but I'm pretty sure they held out to the last minute for a stacked option.

They can't, reasonably.

 

I am aware that there are no guarantees when it comes to DDR4 price, however, considering that JEDEC (eventually) wants 512 and 1024 bit bus in DDR4, it could outperform GDDR5 in future. However, you are right, it's probably too great of a risk to implement DDR4 now considering that initially available solutions would get them nowhere near 192 GB/s.

In another forum I wrote the same thing regarding GDDR5 - regulating the price with a special contract, so I agree with you there.

 

This is just a theory, but with the new rumor going around that PS4's 4 GB of RAM have a bandwidth of 176 GB/s, I'm getting the inkling that Sony may actually have been able to pull off Wide IO RAM stacked on an interposer. This would benefit them because, in addition to having high bandwidth, they'd have significantly lower latency to deal with than GDDR5 RAM which would increase CPU (and likely any non-GPU Co-processor) utilization. It would also have the added bonuses of having a lower per unit cost over time, producing less heat, and using less power.

Of course it could just be 6 Gbps GDDR5 RAM downclocked from 1500 MHz to 1350 MHz (i.e., 192 GB/s down to 176 GB/s). But as someone else already pointed out, why would you do that just to save a measly 3 Watts of power consumption? That spec GDDR5 RAM (i.e., 192 GB/s) has been out for some time and is a known quantity. If 192 GB/s was the target, why not just stick to it? Why would you incur the same expense to buy the same chip and then downclock it? I've heard that it is the memory controller that produces most of the heat anyway. It seems the only way to remedy that is to use a different kind of RAM--thus my hypothesis.

 

(1375MHz)

Choosing a lower bin is going to be about money and/or volume.

 

Oops, 1375 MHz is the right number. My bad.

That's the answer I expected to hear. Stacked RAM in these consoles is a pipedream at this point perhaps, although it would provide a stream of benefits down the road if they could get it out of the gate ahead of the rest of the industry. It just seems that the tech is not ready for primetime just yet.

Oh well, lower binned GDDR5 RAM it is (not that this is a bad thing at 176 GB/s).

 

it is not completely out of the window I think ... Vita does have stacked RAM, but not yet a bus that really takes advantage. But important will be price, slso of production lines, licencing etc an availability.

 

Low latency is important and more ram speed is not equal to low latency ! That means 176gb/s stacked ram is better than 192gb/s gddr5 ram?

 

Well, what we probably shouldn't forget is that regardless of stacking or not, doubling the density or doubling the # of chips (due to the 2GB -> 4GB decision) is going to have a not-so-insignificant impact.

 

With an ~HD7850 class GPU and unified/shared RAM that's simultaneously accessed by the CPU (and additional computing units, potentially), ~190GB/s of total bandwidth sounds like a very reasonable ballpark, though.

The GPU alone would need ~150GB/s (around the bandwidth AMD uses on their Pitcairn cards) to avoid bigger bottlenecks; and an additional ~20-30GB/s for the CPU and some headroom for potential custom compute stuff seems very adequate, too.

I'd be very suprised if Orbis, given the current spec rumors, ended up with less than 150GB/s (potential bottlenecks) or more than 200GB/s (probably not very cost/performance effective) of unified bandwidth - irrespective of how they eventually achieve it.