a very strange solution with HBM memory, because when video cards had GPU + memory on the same substrate (r9 nano, r9 fury on HBM and amdrx vega on HBM2) there was a monstrous percentage of defects due to the HBM memory architecture itself on the same substrate with the GPU, since the line differential between the memory and the chip were so thin that the percentage of memory dump was monstrous. I do not think that the situation will be different when this solution is on network equipment. There will be a monstrous percentage of RMA returns in the first years of sales and the further the percentage of returns will be higher.
Thanks for the comment. But these memories, especially the second generation onwards (HBM2 and above) and their lanes to the core die are very robust. The robustness and quality are quite evident with the very limited number of RMAs we have seen on product shipping. Moreover, all high-end GPU and networking chips are using HBMs, which shows the reliability and performance of HBM are unquestionable.
great video!! (just received the 10k4 chassis with dual 9600 line cards)
Informative video
a very strange solution with HBM memory, because when video cards had GPU + memory on the same substrate (r9 nano, r9 fury on HBM and amdrx vega on HBM2) there was a monstrous percentage of defects due to the HBM memory architecture itself on the same substrate with the GPU, since the line differential between the memory and the chip were so thin that the percentage of memory dump was monstrous. I do not think that the situation will be different when this solution is on network equipment. There will be a monstrous percentage of RMA returns in the first years of sales and the further the percentage of returns will be higher.
Thanks for the comment. But these memories, especially the second generation onwards (HBM2 and above) and their lanes to the core die are very robust. The robustness and quality are quite evident with the very limited number of RMAs we have seen on product shipping. Moreover, all high-end GPU and networking chips are using HBMs, which shows the reliability and performance of HBM are unquestionable.