NVIDIA's B300 raises per-GPU HBM3e memory to 288GB, up 50% from B200's 192GB, with dense FP4 rising from ~9 to ~15 PFLOPS and doubled attention performance. For long-context and large MoE inference, memory capacity often matters more than peak compute. Note that B300 sharply cuts FP64 (~1.25 TFLOPS), making it unsuitable for double-precision HPC.