rokkerkory
Regular
I guess the real question is if MS knew GDDR5 would be available at 8GB in time for launch, would they even have went with ESRAM.
*spin-off* Choice of RAM configuration
- Thread starter rokkerkory
- Start date
MS probably told developers how they think is the best way to use it.
I suspect that middleware engines unreal etc will be updated to take advantage of the esram, may not have a choice in the matter.
As ever it will be 1st parties that may really find interesting ways to use it.
That's one we wont know, but there is a chance that even if they knew they could use 8 GDDR5 they may have still gone the esram route. Simply due to cost.
Although have no idea how much it cost to develop that method, it's not like its something new to them.
That's also dismissing the fact that they may actually believe that there is other benefits to going the esram route.....
Or, if XBO's APU had 18 CUs and 32ROPs would its memory subsystem design ultimately be more flexible and better performing than a single pool of GDDR5?
Not really a practical question, more of looking at the memory subsystem from a different perspective.
People can't separate the ESRAM choice from the lower number of CUs but I wonder that if there was more compute resources on the chip, if people would look at the DDR3/ESRAM choice as having more flexibility and greater performance over more varied workloads than a single pool of GDDR5. I wonder if it might actually be viewed as preferable if it wasn't always framed as as a CU tradeoff.
I've always seen their need for 8GB of memory as driving the decision to use ESRAM and Kinect as the reason for get the cost down elsewhere in the design which led to less CUs. Any idea what CU and ROP parity would have done to cost?
Silent_Buddha
Legend
Highly unlikely. GDDR5 is quite expensive compared to DDR3, not only in monetary costs but in power costs as well.
Regards,
SB
The ESRAM was there before the 8GB of RAM. The size of the RAM pool was not responsible for the decision to use ESRAM.
I can't find any reliable information on GDDR5 being significantly more power hungry given they're using the same manufacturing process.
If anything, the manufacturers actually claim that GDDR5 is less power consuming, like on page 19 of this document
http://www.elpida.com/pdfs/E1600E10.pdf
GDDR5 can also down-volt itself at lower usage, so it could be less power-consuming than DDR3 if anything.
Indeed... Now if we only knew when the move engines became part of the design....
Well, that could still mean it was part of the design due to it being DDR3 not GDDR5, in that regards the amount of main memory is less important than the BW.
Although I still subscribe to the view that it was part of the design more so than some sort of band aid.
Silent_Buddha
Legend
Well, the ESRAM was more tied into the type of ram used instead of the capacity chosen due to the bandwidth in the first place anyway.
It was pretty clear that it was either a large pool of high speed GDDR5, or a huge pool of DDR3 + an eSRAM. 8GB of it was just a choice of "how large is enough."
Except that the decision to use DDR3 vs GDDR5 was likely driven in part by the amount of memory and the relative cost and complication in MB design or at least this is how I viewed it till now.
"System Bandwidth" could actually include more than even the aggregate memory bandwidths as procedurally generated data over internal data paths is possible and IMO eligible for consideration as well. I guess it all depends on how you like to define that. But, as other have stated, the intricacies involved with leveraging these paths to their fullest potential running complex software is non-trivial making direct comparisons difficult.
Silent_Buddha
Legend
Here's PR numbers from the same company...
GDDR5 (Samsung)
http://originus.samsung.com/us/business/oem-solutions/pdfs/Green-GDDR5.pdf
2 Gb module (Samsung 40 nm "class")
8.7W with a 256 bit interface
4.3W with a 128 bit interface
2 GB would then be 69.6W for a 256 bit interface and 34.4W for a 128 bit interface.
DDR3 (Samsung)
http://www.samsung.com/global/business/semiconductor/file/support/memory/green_ddr3_jun_11.pdf
DDR3 (choosing the numbers for the 40 nm "class" chips to keep it equivalent to the GDDR5 sample)
48 GB of 1 Gb chips - 41W - 1.7W per 2 GB or 0.2125W per Gb
48 GB of 2 Gb chips - 34W - 1.4W per 2 GB or 0.175W per Gb
So...at the 40 nm "class" node for Samsung, 2 Gb DDR3 is 24.6x to 49.9x more power efficient depending on whether you use low bit width GDDR5 or high bit width GDDR5.
That's a rather massive difference. Of course, the testing methodology is different (server workload versus graphics workload) and the GDDR5 is obviously massively faster.
The server as "estimated" is with 8 hours active and 16 inactive. So lets say that the GDDR5 is shown with 100% load while the DDR3 power numbers roughly equates to 33% uptime.
That makes it slightly better at only 8.2x to 16.6x more power efficient depending on the interface on the GDDR5.
Regards,
SB
GDDR5 (Samsung)
http://originus.samsung.com/us/business/oem-solutions/pdfs/Green-GDDR5.pdf
2 Gb module (Samsung 40 nm "class")
8.7W with a 256 bit interface
4.3W with a 128 bit interface
2 GB would then be 69.6W for a 256 bit interface and 34.4W for a 128 bit interface.
DDR3 (Samsung)
http://www.samsung.com/global/business/semiconductor/file/support/memory/green_ddr3_jun_11.pdf
DDR3 (choosing the numbers for the 40 nm "class" chips to keep it equivalent to the GDDR5 sample)
48 GB of 1 Gb chips - 41W - 1.7W per 2 GB or 0.2125W per Gb
48 GB of 2 Gb chips - 34W - 1.4W per 2 GB or 0.175W per Gb
So...at the 40 nm "class" node for Samsung, 2 Gb DDR3 is 24.6x to 49.9x more power efficient depending on whether you use low bit width GDDR5 or high bit width GDDR5.
That's a rather massive difference. Of course, the testing methodology is different (server workload versus graphics workload) and the GDDR5 is obviously massively faster.
The server as "estimated" is with 8 hours active and 16 inactive. So lets say that the GDDR5 is shown with 100% load while the DDR3 power numbers roughly equates to 33% uptime.
That makes it slightly better at only 8.2x to 16.6x more power efficient depending on the interface on the GDDR5.
Regards,
SB
You're seriously not suggesting PS4 with 8GBs will run at 200W+ for the memory alone.
by default 256bits = 8 modules or 16 modules in clamshell.
46nm 2Gb GDDR5 @ 256 bits will be 2GB of RAM, consuming 8.7W, around 1W~.5W per module.
PS4 utilizes 16 of the 4Gb GDDR5s, so we can expect the power draw to be around 8W~ 16W depending on how Samsung did their tests.
I stand corrected as 20nm 4Gb DDR3 @ 8GB only takes ~ 4W or so, and 40 nm ones would take approxmately 1.5 of that, or around 6 Watts or so (we're talking 2133 Mhz for the ones used in Xbox One, and yes these probably require 1.5V and consume more)
(http://www.samsung.com/global/business/semiconductor/minisite/Greenmemory/products/green-ddr3)
In the end the difference would at most would be probably 6 W against 16 W. The argument of if that's a lot or not would probably be up to the criteria. Looking from a Wattage/bandwidth perspective GDDR5 is probably equal/better even when comparing 40nm to 20nm modules. Standing from Wattage/size perspective GDDR5 is probably ~half of DDR3.
Last edited by a moderator:
Silent_Buddha
Legend
4 Gb chips should be more power efficient then 2 Gb chips. As well, I'm assuming it's on a newer process node which will reduce power consumption as well. But then again the DDR3 modules will be on larger capacity chips on a lower process node as well. So the ratio should remain relatively the same.
And yes GDDR5 does use a LOT of power. It's a rather large power consumer on desktop graphics boards.
And those numbers that Samsung used are pretty much inline with desktop graphics cards using those chips at that time. So, I don't see anything wonky with their numbers.
[edit] For example in the DDR3 PDF. A 2 Gb DDR3 module uses 71% more power than a 4 Gb DDR3 module for the same amount of memory. So while each individual 4 Gb chip uses more than a 2 Gb chip, you need 2x 2 Gb chips to have the same memory capacity.
Regards,
SB
a 2Gb module is not going to eat up 8.7W or have a 265 bit interface. 8 of them, sure.
