Hierarchical Paging Telemetry Profile
Overlaps active decoding GEMMs with asynchronous GPUDirect Storage DMA transfers, eliminating host-device IO stalls and sustaining 10M+ token context windows at sub-10ms page fault latency.
Page Fault Latency: Sub-10ms (6.4ms Avg)
Inter-Token Latency: Sub-8ms Real-Time
Max Context Sustained: 10,000,000 Tokens