Samsung's GDDR6W Doubles Performance and Capacity

1_rick

2[H]4U
Joined
Feb 7, 2017
Messages
3,418
https://www.tomshardware.com/news/samsung-gddr6w-doubles-performance-and-capacity

"Samsung has introduced its all-new type of GDDR6 memory that doubles the DRAM package's capacity and increases interface width to double its peak bandwidth."

It's got twice as many dies inside the package, with a more advanced 3D interconnect, has two physical interfaces to get the double bandwidth, and is thinner than regular GDDR6, so it should be easier to cool. They claim it approaches the bandwidth of HBM2E.
 

scottypippin

Weaksauce
Joined
Feb 28, 2022
Messages
69
dang, someone's gotta tell me what a gig per bit-second is:
1669928714419.png

ludacris results all things considered
 

LukeTbk

2[H]4U
Joined
Sep 10, 2020
Messages
3,349
dang, someone's gotta tell me what a gig per bit-second is:
Same has Gb/s
Gigabits per second (Gbps), sometimes abbreviated Gb/s,
https://www.techopedia.com/definition/2704/gigabits-per-second-gbps

512 * 22 / 8 give us 1,408
4096*3.2 / 8 give us 1,638

I not sure how impressive we should be here a 384 bits GDDR6X on lovelace does 1,008 GB/s, 512 /384 * 1 = 1.33 TB/s

Is it a way to get a bit better (5%) than some of the best GDDR6x around without the significant added power-price ? Which I would imagine would be nice, but not sure special versus the upcoming GDDR7

Or maybe the cost of the 512 bits is similar to a regular will be mid range in the future 256 bits (a 6800xt) because of some free doubling going on, maybe that what that part is saying, I do not know enough to understand:
DRAM devices featuring and therefore features two 32-bit interfaces, thus doubling capacity (from 16Gb to 32Gb per chip) as well as interface width (from 32-bits to 64 bits). To do so, Samsung's GDDR6W chips use the company's Fan-Out Wafer-Level Packaging (FOWLP) technology that replaces traditional printed circuit board with a redistribution layer (RDL) that is thinner and has significantly finer wiring patterns.

If it is a change that will make 512 to 768 bits bandwith common at reasonable price that would be quite something and cheaper card having nice 256-384 bits performance.
 

KazeoHin

[H]F Junkie
Joined
Sep 7, 2011
Messages
8,724
Still, When was the last time anyone saw a GPU with a 512bit-wide bus?

By the math, a 384bit-wide bus would still be over 1TBp/s which is HBM1 levels of bandwidth so, wow...

I can't wait to see this in consumer products never.
 

LukeTbk

2[H]4U
Joined
Sep 10, 2020
Messages
3,349
Still, When was the last time anyone saw a GPU with a 512bit-wide bus?
Thats what it mean, a doubling of the bus with similar trace ? yeah it would be a possible revolution (if the price cut get passed down because of all the competition in high end GPU)
 

Grebuloner

[H]ard|Gawd
Joined
Jul 31, 2009
Messages
1,491
Same has Gb/s
Gigabits per second (Gbps), sometimes abbreviated Gb/s,
https://www.techopedia.com/definition/2704/gigabits-per-second-gbps
Read his post and chart again... ;)

Anyway, while there are some definite advances made, here, particularly in size and cooling potential, the headline is overhyping it. It's not that amazing that you can get twice the everything with twice the stuff. 512-bits of GDDR6X will make the same bandwidth. Cost will be the determining factor, just as it is with HBM.
 

Armenius

Extremely [H]
Joined
Jan 28, 2014
Messages
35,357
Still, When was the last time anyone saw a GPU with a 512bit-wide bus?

By the math, a 384bit-wide bus would still be over 1TBp/s which is HBM1 levels of bandwidth so, wow...

I can't wait to see this in consumer products never.
Radeon R9 290X and its sixteen 2Gb GDDR5 memory chips.
 

Lakados

Supreme [H]ardness
Joined
Feb 3, 2014
Messages
6,908
The important take away I see here are the 6W has the same power and thermal properties of GDDR6 but twice the density.

GDDR6x has a pin speed of 21Gbps but 6W runs at 22Gbps so very slightly faster, but Samsung says it’s cheaper than their 6X competitors from Micron.

So it’s cheaper, cooler, “faster”, and double the capacity.
Yay for progress.
 

DanNeely

Supreme [H]ardness
Joined
Aug 26, 2005
Messages
4,340
If I'm understanding correctly what Samsung's effectively done is to allow making a 512bit card while using the PCB space of only a 256bit one by halving the number of packages needed. As preposterously huge as NVIdia's current generation of cards are saving PCB space itself is probably not a major concern.

It still requires 512 PCB traces; and AFAIK cramming them in is the expensive part because you have have more GPU pins, more EMI/cross talk, and need more PCB layers to fit the extra traces and wrap everything in grounds to suppress the EMI.

For narrower bus widths halving the number of packages would let them keep all the ram on the same side for cooling while being able to put it all closer to the GPU for shorter trace lengths. If anything I suspect this will be the part most attractive to NVidia and AMD.
 

Armenius

Extremely [H]
Joined
Jan 28, 2014
Messages
35,357
If I'm understanding correctly what Samsung's effectively done is to allow making a 512bit card while using the PCB space of only a 256bit one by halving the number of packages needed. As preposterously huge as NVIdia's current generation of cards are saving PCB space itself is probably not a major concern.

It still requires 512 PCB traces; and AFAIK cramming them in is the expensive part because you have have more GPU pins, more EMI/cross talk, and need more PCB layers to fit the extra traces and wrap everything in grounds to suppress the EMI.

For narrower bus widths halving the number of packages would let them keep all the ram on the same side for cooling while being able to put it all closer to the GPU for shorter trace lengths. If anything I suspect this will be the part most attractive to NVidia and AMD.
Ampere and Lovelace PCBs are smaller than Turing and earlier boards. 2080 Ti on top, 4090 on bottom:
1669995015321.png


Regardless, you want as little physical distance as possible between the memory packages and the GPU, so board size would not even matter.
 
Top