Tag
This blog post details the hardware path of a GPU memory read instruction on an NVIDIA RTX 4090, explaining how a CUDA kernel accesses memory through components like L1 cache and DRAM for performance insights.
Reveals that the 748GB unified memory advertised for the NVIDIA DGX Station actually only has 252GB of high-speed HBM available. The remaining 496GB of slow LPDDR5X is essentially useless for large model inference, reflecting NVIDIA's precise product differentiation strategy.
A CT scan of the Intel 386 processor's ceramic package reveals complex internal wiring, multiple power networks, and layered interconnect design, offering a rare look at the packaging technology of the 1985 chip.
A detailed technical analysis of Super Nintendo cartridges, covering CIC copy protection, ROM size distribution, SRAM with battery backup, and enhancement processors like the Super FX chip.