Flash Memory Summit 2020: Intel Keynote

## Overview

Intel's 2020 Flash Memory Summit keynote presents a one-year progress update on Optane persistent memory technology. The speaker addresses the critical memory bottleneck facing modern data centers: while CPU core counts on Xeon processors have grown exponentially through architectural innovations like chiplets, DRAM scaling has dramatically slowed from 4x density increases every three years (mid-1990s) to only 2x density every four years currently, with per-bit cost declining even more slowly at roughly 10-12% annually. This creates an unsustainable gap between compute demand and memory capacity that aligns with Amdahl's Law—memory must scale proportionally with compute power.

## Solution Architecture

Intel's response is a re-architected two-tiered memory and storage hierarchy. For memory, this pairs a performance tier (DRAM accessed via standard cache/memory hierarchies) with a capacity tier (Optane persistent memory attached to the DDR bus, accessed through load/store instructions at near-DRAM speeds but significantly cheaper than DRAM). For storage, Optane serves as a high-performance non-volatile tier while traditional SSDs provide cost-effective capacity, enabling both performance and cost objectives simultaneously.

## Ecosystem and Market Adoption

Over one year, Intel developed a robust ecosystem spanning all major OEMs, systems integrators, cloud service providers, and ISVs with qualified and optimized software. Market adoption significantly exceeded typical new product performance: approximately 200 Fortune 500 companies have deployed persistent memory or are in proof-of-concept stages, with an 85% POC-to-deployment conversion rate, resulting in over 270 production installations. Customer deployments fall into three benefit categories: TCO savings (30-40% system cost reduction through DRAM replacement without performance loss), throughput improvements (2-4x job increases by eliminating memory bottlenecks), and faster time-to-insights (order-of-magnitude improvements via persistent storage features and instantaneous recovery).

## Second Generation Performance

Intel released the 200 series persistent memory with a new media controller delivering 25% performance improvement over first generation. Benchmark comparisons show dramatic advantages: 1000x lower latency versus NVMe SSDs at 1GB/s traffic, 3.8x higher bandwidth than SSDs, and sustained performance under heavy load.

## Customer Deployments and Applications

SAP HANA testing demonstrated compelling value: doubling system memory capacity while maintaining similar DRAM-level costs, with 6x faster database restart times; alternatively, reducing server counts by 50% for large databases while maintaining capacity and achieving 52% lower hardware costs. The DAOS (Distributed Asynchronous Object Store) software stack, co-developed with HPC customers, combines persistent memory with traditional storage by placing small, frequently accessed metadata and indices on persistent memory while aligning larger data structures to block storage. This eliminates serialization overhead and block contention from concurrent multi-process access. DAOS achieved record results on the IO 500 storage benchmark, scoring 2x the next solution overall and claiming the top three positions (Intel DAOS, Texas Advanced Computing Center, and Argonne National Labs) in the 10-node challenge.

Oracle Exadata X8M exemplifies deep collaboration, integrating persistent memory with RDMA over converged ethernet for the hot storage tier. Results include 2.5x higher IOPS (16 million total), 10x better latency (under 19 microseconds for database reads), and availability on Oracle Cloud as a competitive advantage. RDMA-based replication to persistent memory demonstrated 10x+ latency improvements compared to NVMe SSDs over TCP.

## Developer Resources and RDMA Democratization

Intel released libRPMA libraries within PMDK (Persistent Memory Developer Kit) to democratize RDMA integration, enabling developers to use popular APIs without rewriting applications. The speaker highlights the free "Programming Persistent Memory" book and online resources. The keynote emphasizes collaborative innovation and thanks ecosystem partners, customers, and audience contributors to the persistent memory revolution.