High performance, high capacity and high bandwidth memory solutions are helping bring the virtual realm to a closer match with reality. To meet this growing market demand, Samsung Electronics has deve... Samsung GDDR6W: Twice the Bandwidth and Density of GDDR6 (stacking)
Have I understood this correctly? This is only packing 2 individual mem chips on top of each other? A 3090 could have used the double stack to keep memory on one side. The 3090 ti solved that problem with double sized memory modules. So more then 24GB memory is needed before stacking makes sense and it still needs to be cheaper then just going back to putting memory on both sides of the PCB again like the 3090. AMD and Nvidia already use HBM for the top pro cards, so the stacking would not really be needed here, unless it is a cheaper alternative. The GPU would need a wider bus to address the 2 individual memory modules in the stack if double mem speed is required, just like when 2 memory modules was just side by side on a PCB? GPU manufactures would mostly cheap out on the mem bus, so cost would prevent this from doubling the mem speed, just like today. The only place I can see this really shine would be for small form factor GPU, laptop GPU or SSD cache.
it doesnt need to be wider necessarily, they just need to increase the number of data per clock cycle, gddr6 is quad data rate memory so 4 bits per cycle. octal data rate memory (8 bits per cycle) has been done before , so its definitely possible. the wiki page on gddr6 actually mentions that while the data rate is qdr , the command rate is actually odr, so seems like updating the spec to support odr data rates too would be a natural evolution.
If this is basic stacking of 2 chips on top of eachother, DDR should work fine with stacking too, maybe even stacking of memory on top of the main flash storage. A "fairly cheap" PS 5 has a GDDR controller on the CPU so I do not see why a new server grade SSD with gen 5 at 4-16 lanes could not have the same ability. It could be a mmo player data server or youtube login server or something like that.
So those modules is 4GB each with double the bandwidth per chip, would make use of only 6 chips to achive the same 24GB on a 384bit memory bus config. That would require less PCB space and a more simple power supply.
GDDR7 will likely appeared on consumer cards from 2024 onwards. 2023 will be busily occupied with standard/old GDDR6, new 24Gbps+ GDDR6, GDDR6X, GDDR6 + L3/Infinity Cache, and GDDR6W.
