The discussion of HBM(2) versus traditional memory keeps on going, two weeks ago Hynix announced GDDR6 memory for graphics cards in 2018. Over at GDC Hynix has now shown a wafer and shared specs of the new GDDR6 chips. It's twice as fast.
Now, I started the paragraph with the HBM2 discussion. While HBM might seem to offer better memory-bus width, latency and lower voltages, it seems hard to fab and implement and likely thus is more costly and complicated to use than expected. I mean look at the massive AMD Radeon RX Vega delay, this likely is due to the limited availability of HBM2. Meanwhile Nvidia has been plastering GDDR5X in their high-end SKUs and is laughing rather hard. GDDR5X still rocks hard this year. Next to HBM2, GDDR6 is the new thing for 2018. The new graphics memory type will offer up to 16Gb (2GB) per IC. So eight ICs would already get you to 16 GB of graphics memory.
GDDR5 offers up-to 8 Gbps performance. If you look at that newer GDDR5X memory, these offer 11~12 Gbps speed, GDDR6 could offer speeds up-to 16 Gbps (thus a 16 GHz effective data-rate). It’s not stated that the introduction ICs will run at that speed though. But realistically seen from say a GeForce GTX 1070, it could offer twice the memory bandwidth from 256 towards 512 GB/sec. It seems clear that GDDR types of memory might be more energy inefficient opposed to HBM, but this memory type remains to be to implement and has a faster market adoption rate. The photos below are taken by hardwareluxx (see source link) and come from GDC.