Citrini Analyst Questions SanDisk's HBF vs. HBM Performance Claims

Huoxing Finance reports that on August 14, Citrini analyst Zephyr published a post questioning the parameter settings in SanDisk’s investor day presentation comparing HBF and HBM performance, suggesting the company deliberately underrepresented HBM’s capabilities. He noted that SanDisk set the total bandwidth for both HBM and HBF at 12.8 TB/s, equivalent to 1.6 TB/s per stack, while running the Qwen3-480B-A35B model using bfloat16. Zephyr pointed out that current model inference primarily uses FP4 or FP8, meaning the model would require approximately 240 GB to 480 GB of memory under different quantization schemes. Zephyr also stated that SanDisk fixed the HBM capacity per GPU at 192 GB, whereas a 16-layer HBM4E configuration with eight stacks can achieve 512 GB of capacity and approximately 32 TB/s of bandwidth—about three times the parameters used in SanDisk’s demonstration. He argued that these chosen parameters render SanDisk’s conclusion—that “HBF requires only 1 to 4 GPUs, while HBM requires 8 GPUs”—inadequate to represent the actual performance of higher-spec HBM configurations.