Explainers
Illustrated introductions to problems in memory systems, written from our papers.
Memory for AI
Memory for AI Memory for AI: why inference and training wait on memory Why generative decode is limited by memory bandwidth, how Raptor's 3D-DRAM moves the Pareto frontier of throughput against interactivity, and how Keyformer, FAE and Slipstream reduce the data that must move. Read the explainer
- Why decode is memory-bound
- Two phasesThe KV cacheA roofline for decodeBatchingBandwidth or capacity
- Raptor and 3D-DRAM
- DesignStream blockingStream flippingRedundancyThermal behaviorRefresh at 105 °CSilicon
- Performance
- DeploymentThroughput against interactivityTokens per second per kWHeterogeneous deployments
- Moving fewer bytes
- Keyformer
- Recommendation training
- Embedding tablesFAESlipstream
- Summary
- At a glanceThe papers
Rowhammer
Memory security Rowhammer: when reading memory corrupts it Why activating one DRAM row flips bits in its neighbors, and how a decade of our work defends against it, one paper at a time. Read the explainer
- The problem
- Rows of chargeThe disturbanceA falling thresholdWhy it matters
- Tracking the aggressor
- CRA and PRAHydraDAPPER
- The mitigating action
- Randomized Row-SwapScalable and Secure Row-SwapAQUASafeGuard
- Counting inside the chip
- PRACQPRACCnC-PRACCounterpointPRACLeak and TPRACPrISM
- Summary
- At a glanceThe papers