Early Access · Proof of Concept

Unvertical NVMC

Non-Volatile Memory Cache — next-generation caching middleware that replaces expensive DRAM page cache with cost-effective NVM in disaggregated storage architectures.

Trade DRAM cost for NVM capacity

Modern data centers separate compute from storage for scalability, but this creates a latency problem. The standard fix — buffering data in DRAM page cache — is expensive and getting more so.

NVMC takes a fundamentally different approach.

Observe the mismatch

In disaggregated storage, DRAM used for caching is vastly overprovisioned in speed relative to the remote backend it fronts. That performance headroom is wasted money.

Replace with cheaper media

NVM (SLC NAND) is ~25× cheaper per GB than DDR5 but still orders of magnitude faster than remote storage. More capacity means higher hit ratio, which means better aggregate performance.

Use the spare bandwidth

Latency-bound workloads can't saturate network links to storage nodes. NVMC uses that idle bandwidth for speculative prefetching onto local NVM — trading bandwidth for latency.

Why this matters now

DRAM costs are rising sharply as manufacturers shift capacity from DDR5 to HBM for AI workloads. A technology that reduces DRAM dependency addresses an increasingly urgent need.

~25×
Cost difference
DDR5 at ~$25/GB vs SLC NAND at ~$1/GB. More cache capacity for less money.
+50%
DRAM price increase
Spot prices rose ~50% in Q4 2025, with contract prices projected to follow at 45-50% QoQ.
30-50%
DRAM share of server CAPEX
Memory is one of the largest line items in server costs. Reducing it has outsized impact on TCO.

The first NVM-first compute-side cache

Every existing caching solution treats NVM as a secondary tier behind DRAM. NVMC inverts this hierarchy — NVM is the primary caching layer, not an afterthought.

⚡

NVM-First Cache Engine

Purpose-built for NVM characteristics: read/write asymmetry, higher latency than DRAM, I/O granularity. Not a DRAM engine with NVM bolted on.

📡

Bandwidth-Aware Prefetching

Real-time analysis of I/O patterns and network utilization. Proactively fetches data onto local NVM using spare link bandwidth — possible only with NVM's capacity advantage.

🧠

Minimal DRAM Footprint

Redesigned metadata structures to minimize the engine's own DRAM consumption. The whole point is to reduce memory costs — the engine itself can't be the problem.

🔌

Transparent Block-Level Integration

Plugs in as a middleware layer compatible with Ceph, KVM/QEMU, Proxmox, HDFS, and distributed databases. No application changes required.

Built for disaggregated storage

NVMC is designed for environments where compute and storage are separated — the architecture pattern where its NVM-first approach delivers the most value.

☁

Distributed Storage Clusters

Ceph, NFS-over-RDMA, NVMe-oF — reduce latency from remote storage nodes by caching and prefetching on local NVM at the compute layer.

🖥

Virtualization with Remote Storage

KVM/QEMU, Proxmox, OpenStack — accelerate VM disk I/O transparently while cutting the DRAM budget allocated to host page cache.

🗄

Databases on Shared Storage

Reduce query latency and DRAM costs for databases running against remote or shared storage backends. Particularly effective for read-heavy and mixed workloads.

📊

Big Data & Analytics

HDFS, Spark, data lakes — cache iteratively accessed datasets locally on NVM. Speculative prefetching is especially effective for sequential scan patterns.

Help shape the future of compute-side caching

NVMC is currently at proof-of-concept stage. We're looking for 2–3 design partners with disaggregated storage environments who want to validate the technology on real workloads and influence the product roadmap.