Citrini analysts question SanDisk HBF performance comparison: demonstrate parameters or underestimate HBM's actual capabilities

In comparison, Citrini analyst Zephyr posted an article questioning the parameter setting bias in the HBF and HBM performance comparison shown by SanDisk on Investor Day, believing that the company deliberately underestimated HBM's performance during the presentation.

He pointed out that in the comparison, SanDisk set the total bandwidth of HBM and HBF to 12.8 Tb/s, which is equivalent to 1.6 Tb/s per stack; at the same time, bfloat16 was used to run the Qwen3-480B-A35B model. According to Zephyr, the current model inference uses more FP4 or FP8, so the model requires about 240GB to 480GB of capacity under different quantization methods.

Zephyr also stated that SanDisk fixed the HBM single GPU capacity at 192 GB, but the 16-tier HBM4E can reach 512 GB in 8 stack configurations and provide about 32 Tb/s of bandwidth, which is about 3 times the SanDisk demo parameters.

It believes that the above parameter selection makes the HBF shown by SanDisk only require 1 to 4 GPUs, and the results of HBM requiring 8 GPUs are not sufficiently representative of actual performance under higher specification HBM configurations.