"How does it compare to HBM2 ?

Could bring good news to the ridiculous prices we have today
if we do the basic calculation

rx 580 with 8gbps gddr5 memory we have
(2ghz real frequency) x(4 ,quad data rate) x (256bit bus) /8 = 256GB/s

for 18Gbps gddr6 on a 256bit bus
(4.5ghz real frequency)x(4 ,quad data rate)x(256bit bus) /8 = 576 GB/s

for 18Gbps gddr6 on a 384bit bus (like a titan Xp)
(4.5ghz)x(4)x(384)/8 = 783GB/s

for hbm 2 2.4gbps on 2 stacks (feasibly a vega 64 card could be equiped with this)

(1.2ghz real frequency)x ( 2 ,double data rate) x (2048 bit bus) /8 = 614GB/s

for hbm2 on 4 stacks (vega 20 machine learning gpu)

(1228.8gb/s)

so i'd say that the gddr6 has an advantage if its cheaper than hbm2 particularly in the consumer range of products, since its offers similar bandwidth and doesn't need an interposer.

however i will say once you go past a 384bit bus things get more expensive, and require alot more space than a similar hbm2 solution,
hbm is still superior interm of its charateristics (power consumption (pretty sure), space efficency, memory density ect) but its cost is the inhibiting factor

what i would expect to see from gddr6 is the return of 64 bit and 128bit bus gpus to mainstream, since the bandwidth provided will actually be enough to be useful, will either make cards cheaper , or more profitable ." - [link]

"Cuccok: csattogós lepke és tiki-taki" └[ʘヘʘ]┘