Citrini Analyst Questions SanDisk HBF Performance Comparison: Displayed Metrics May Underestimate HBM's Actual Capability

August 14th, Citrini analyst Zephyr questioned SanDisk's performance comparison between HBF and HBM presented at the Investor Day, suggesting that there is a parameter setting deviation. Zephyr believed that the company "deliberately underestimated the performance of HBM" in the demonstration.

He pointed out that SanDisk set the total bandwidth of both HBM and HBF to 12.8TB/s in the comparison, equivalent to 1.6TB/s per stack; and used bfloat16 to run the Qwen3-480B-A35B model. Zephyr stated that the current model inference relies more on FP4 or FP8, so the model's required capacity under different quantization methods is about 240GB to 480GB.

Zephyr also stated that SanDisk fixed the HBM single GPU capacity at 192GB, but a 16-layer HBM4E can reach a capacity of 512GB in an 8-stack configuration, providing approximately 32TB/s of bandwidth, roughly three times the bandwidth of the parameters in SanDisk's demonstration.

He believed that the parameter choices above resulted in SanDisk's demonstration showing that "HBF only needs 1 to 4 GPUs, while HBM requires 8 GPUs," which cannot fully represent the actual performance of higher-spec HBM configurations.