an investigation into Huawei's copes for lacking robust HBM supply. Not sure about details on HiZQ 2.0 process. They'll determine how the bulk of the Chinese compute in the next 2 years is built, so it's important. I fail to see how anything below "millions" makes sense.
What is HW doing w/ HBM on Ascend? HiBL1.0 uses 8-die stack & 2GB/die, but lower bandwidth/die for 1.6TB/s
HiZQ makes up for lower Bus width (due to organic substrate constraint) thru higher CXMT DDR5 data rate.
How 950DT achieved 4.0 TB/s vs 3.35 TB/s for H100 & 4.8TB/s for H200