Thwarting Key Extraction and Supply Chain attacks by Detonating GPUs

The talk asks why a device owner might want a built-in self-destruct for GPUs and other accelerators, since whoever physically holds the hardware is not always a party you trust. The concerns covered include key extraction through physical side channels, tampering along the supply chain, and deploying expensive hardware in regions where political relationships may change. The goal is to destroy the chip fast enough to deny access, and to do it at data center scale.

The speaker walked through the options tested. Nanothermite deposited on a chip was difficult to ignite, especially with a heat sink attached, and it left the chip largely intact, so it was ruled out. Small detonators placed inside the heat sink, using the copper cold plate as the striking surface, produced complete destruction of the GPU die, while the memory stacks were less thoroughly destroyed. Detonators were also aged at elevated temperature to check they would stay stable over a long service life. Layers of metal foam, carbon fiber and a Kevlar sleeve were used to contain the blast on both SXM and PCIe form factors.

To close, the speaker outlined next steps: cleaner packaging, regulatory exemptions for handling and shipping, and integration with tamper detection or an emergency scram trigger. For practitioners, the takeaway is that any destruction design has to balance effectiveness, cost, containment and legal limits, not just speed.