Hard Disk DriveNever Die. Request Technical Brief

More room
for intelligence.

A magnetic memory layer for the model data AI keeps coming back to.

Explore the applications

Store in magnetism.
Design beyond the chip.

Thermal resilience
for Processing-in-Memory.

Unlike charge-based DRAM cells, the platter retains data as magnetic orientation without refresh. High-anisotropy recording materials can provide strong resistance to thermal fluctuations, making the magnetic layer a promising foundation for thermally robust model storage.

For Processing-in-Memory, this separates bulk data retention from the charge-based memory cells beside hot compute logic. Media selection and cooling can be tailored to each layer, with thermal limits validated for the chosen material, retention time, and read electronics.

Radiation resilience
for space data centers.

Space radiation can upset semiconductor memory when a particle deposits charge in a sensitive circuit. A magnetic platter stores its bits through magnetization, giving the storage medium a different susceptibility from charge-based memory cells.

Historical NASA-hosted research identifies magnetic media as radiation-tolerant while support electronics limit the disk system. For space data centers, this offers a potential media-level advantage: pair a persistent magnetic model store with protected read and compute circuitry, and qualify both for the mission’s radiation environment.

Keep live state close.
Bring model data into reach.

Activations and KV cache need fast working memory. Weights, embeddings, and experts need capacity and repeated access. Magnetic HBM explores a dedicated layer for that model data.

Magnetic HBM

Weights · embeddings · experts

Prefetch

HBM / DRAM

Active layers · activations · KV cache

Compute

GPU / NPU

Run the model

A proposed memory hierarchy. Effective throughput depends on the read plane, interface, scheduling, and workload.

Read often.
Think bigger.

From AI inference to processing near the read plane: explore what a large magnetic memory layer could enable.

Stream the next layer.
Keep the model moving.

For models whose weights exceed accelerator memory, stage upcoming layers from Magnetic HBM while the current layer computes. Keep active tensors and live state in HBM or DRAM.

Model data
Large, reusable weight blocks
Access pattern
Layer-by-layer reads with a prefetch window
What to validate
Whether layer transfer fits within available compute time

More experts in reach.
The active ones up close.

Explore storing an expert-weight pool in Magnetic HBM and keeping frequently used experts in working memory. Batch routed requests and prefetch selected experts where the execution schedule allows.

Model data
Expert weights beyond the resident cache
Access pattern
Grouped expert reads with reuse across tokens
What to validate
Routing lead time, cache hit rate, and transfer stalls

A larger lookup space.
A more deliberate read.

Explore large embedding tables and recommendation models where lookups can be grouped into useful blocks. Keep hot entries close to compute, and schedule reads from the larger magnetic pool in batches.

Model data
Embedding tables and reusable vector blocks
Access pattern
Batched, locality-aware reads
What to validate
Read amplification, batch size, and tail latency

Bring processing
to the read plane.

Processing-in-memory (PIM) brings work closer to stored data to reduce data movement. Magnetic HBM offers a direction to explore: add processing logic beside the parallel read array, so selected operations happen before data travels to the host accelerator.

Local filtering, decompression, and reductions could turn large input streams into smaller outputs. The platter retains the data; nearby semiconductor logic performs the computation. This is a proposed near-memory extension that needs its own hardware and software validation.

Candidate work
Filtering, decompression, and blockwise reductions
Potential benefit
Less off-device traffic when local processing reduces output size
What to validate
Compute throughput, precision, power, and read-plane heat

A model-memory layer.
Beyond Earth.

Platter

Persistent weights & embeddings

Read plane + compute

Radiation-tolerant logic · ECC · recovery

Orbital AI infrastructure would need large, persistent model stores alongside reliable compute. Magnetic HBM is a candidate to investigate because its platter separates magnetic data retention from the semiconductor circuits used to access it.

Combine that media-level radiation opportunity with protected electronics, error correction, and recovery. Near-memory processing could also reduce traffic between the model store and accelerator when operations can be completed locally.

Potential role
Persistent model weights and embedding pools for orbital inference
Design priorities
Radiation qualification, heat rejection, sealed operating environment, and launch loads
What to validate
Media retention, total dose, single-event effects, bearings, and read-plane clearance

Space operation also requires pressure and mechanical engineering; a terrestrial disk is not automatically space-ready. NASA: Mass Data Storage for Space Missions.

The right data.
The right rhythm.

The strongest fit is large model data that changes infrequently, is read repeatedly, and can be scheduled ahead of computation.

  • Enough useful data per read to use parallel bandwidth.
  • A predictable schedule or enough batching to prefetch.
  • Working memory for active tensors and latency-sensitive state.

Tiny, unpredictable random requests and frequently rewritten live state are less suitable. These applications are architecture directions to validate, not measured deployment results.

See how the
numbers line up.

Explore bandwidth and latency assumptions in the architecture calculator, or start a technical conversation about your workload.

Open Architecture Calculator Request Technical Brief