scratchpad
A form of on-chip memory that is explicitly managed by software. Unlike caches controlled automatically by hardware, the compiler must decide how data is brought in, kept resident, and evicted.
Timeline (last 90 days)
Each Spyre core has an independent scratchpad with limited capacity, which makes data placement planning a first-class task for the compiler.
A scratchpad is on-chip storage explicitly managed by software, unlike the hardware-controlled cache hierarchy in GPUs.
On architectures that use scratchpads, the compiler must explicitly decide when each block of data is brought in, how long it resides, and when it is swapped out; the hardware does not make cache-hit decisions for the programmer.
An end-to-end encrypted scratchpad can to some extent limit the propagation path of prompt injection attacks, but genuine protection still requires combining access control and content validation mechanisms.
All Facts (4)
Each Spyre core has an independent scratchpad with limited capacity, which makes data placement planning a first-class task for the compiler.
50%UnverifiedA scratchpad is on-chip storage explicitly managed by software, unlike the hardware-controlled cache hierarchy in GPUs.
50%UnverifiedOn architectures that use scratchpads, the compiler must explicitly decide when each block of data is brought in, how long it resides, and when it is swapped out; the hardware does not make cache-hit decisions for the programmer.
50%UnverifiedAn end-to-end encrypted scratchpad can to some extent limit the propagation path of prompt injection attacks, but genuine protection still requires combining access control and content validation mechanisms.
50%