Samsung GDDR6 expected with 14-16 Gbps from 2018

Well publicly Micron is now showing their GDDR6 is in sampling status; 12-14Gb/s.

Like I said earlier, I am not convinced Samsung will be releasing the 16Gb/s for a little while as that performance IMO was more related to lab conditions/prototypes.
Still 14Gb/s speed from multiple manufacturers is nice to start with for next gen of upper GPUs.
 
Well publicly Micron is now showing their GDDR6 is in sampling status; 12-14Gb/s.

Like I said earlier, I am not convinced Samsung will be releasing the 16Gb/s for a little while as that performance IMO was more related to lab conditions/prototypes.
Still 14Gb/s speed from multiple manufacturers is nice to start with for next gen of upper GPUs.
Afair Micron was able to run G5X with 16 Gbps in 2016, too. But it was not reliable enough, i think.

Yeah, Micron ran GDDR5X at 16 Gbps in lab conditions back in June.

https://www.micron.com/about/blogs/2017/june/what-drives-our-commitment-to-16gbps-graphics-memory

They also posted an approximate roadmap that shows them getting to 16 Gbps no earlier than 2019. I expect Samsung will see a similar ramp-up despite what that CES award suggested.

blog_image_gddr_3.png
 
Ah, thanks. My brain kept the wrong date since the last months past by so quickly for me. :D G6 needs a (very) different MC since the i/o config is x8/x16 instead of x16/x32. A conventional MC has 2x 32 bit lanes for each channel and module (correct me if i'm wrong - and let's keep outlaws such as the 660 Ti off the track :D). So there will be no mix with either G5X or G6 - analogue to G5 and G5X.
 
G6 needs a (very) different MC since the i/o config is x8/x16 instead of x16/x32. A conventional MC has 2x 32 bit lanes for each channel and module (correct me if i'm wrong - and let's keep outlaws such as the 660 Ti off the track :D). So there will be no mix with either G5X or G6 - analogue to G5 and G5X.
Does that mean a GDDR6 GPU cannot also support GDDR5 or GDDR5X?
 
I'm pretty sure there won't be a mix of GDDR5X and GDDR6 on one PCB. The pinout and the pincount are different, so you will need at least a modified PCB. The controller side is a different question. I think it should be possible, but not sure <yet>. Have to check specs back again, but i'm really busy in theses weeks. ^^
 
Yeah they are very different not just in layout in terms of ball assignments but also number of assignments generally for Data/CA/misc signals.
Also another major difference is the structure/use of the CA (Command-Address) with GDDR6 compared to GDDR5X when looking at system view and flow control; it occurs after Byte 0 and 1 transfer for each channel with GDDR6.
 
Last edited:
Just to be clear: i meant one type of layout for a PCB, not a single PCB, with a GPU and GDDR5X+GDDR6. :D I read my sentence twice and felt it should be mentioned...

Happy new year. :)
 
heh. :cool:

hm... GT 1030 uses 2x1GB on a 64-bit bus @ 6Gbps for 48GB/s. Would it be particularly cheaper than 1x2GB @ 12Gbps (in say... another year for hypothetical product release)?
 
did gddr5x end up reaching 14ghz? if so are there any differences or just in effect a rename?
Only in lab/walled conditions.
Highest offered was 12Gb/s, and there are subtle differences between the standards from channels to structure-number of pins for data/control/etc.
In their Micron technical notes manuals they mention 13Gb/s achievable for GDDR5X although not offered to public.
The biggest fundamental difference is what they call access granularity (read/write in a single flow cycle) and here GDDR6 is same as GDDR5 while GDDR5x is double and also channel count difference of 2 for GDDR6 vs 1 for GDDR5x that is part of the reason.
If you look at the data-control flow you will notice GDDR5x is not the same with command/address relation with data - not the data clocking frame diagram.
The double data sounds good on paper for GDDR5x with regards to access granularity, but one needs to consider this also increases the flow control mechanisms or longer delay for CA in terms of communication so has trade-offs and especially so if certain functions/applications require optimised memory in way of GDDR5 characteristics; we can see some real world examples that are impacted by GDDR5x.
Maybe it also impacts/complicates the development of higher transmission rates and why only 12Gb/s are offered for GDDR5x; context also includes the 2 channels with GDDR6 that are part of this in the way I mention.

I am surprised they are able to offer 16gbps and higher now for GDDR6 from Samsung, would had expected it to be around 12 months for mass production at those speeds.
Price must be a premium.
But even Micron are able to offer higher GDDR6 to their clients than their GDDR5x.
 
Last edited:
Back
Top