Products & Tools

Cerebras releases Tungsten for wafer-scale programming

Cerebras released Tungsten and Aviary on November 16, 2026. Its compiler manages a 900,000-core wafer, and an agent wrote a plasma simulation using 98% of it.

Cerebras Releases Tungsten for High-Level Wafer-Scale Programming
Cerebras Releases Tungsten for High-Level Wafer-Scale ProgrammingBru-nO / pixabay
By Elena Vasquez6 min read

Why it matters

  • Cerebras released Tungsten and Aviary on November 16, 2026.
  • An agent-written plasma simulation ran across 98% of a 900,000-core wafer.
  • SNAP reached 1.6 PFLOP/s on one CS-4, ahead of 10,800 GPUs on El Capitan.
  • HPCG reached 1.04 PFLOP/s on one system.
  • The plasma run advanced 60 million particles at 7,200 time steps per second and 430 billion particle updates per second.

By Elena Vasquez

Cerebras released Tungsten and Aviary on November 16, 2026, and says an agent-written plasma simulation ran across 98% of one 900,000-core Wafer-Scale Engine. Cerebras describes Tungsten as a high-level language and compiler for that processor. It presents Aviary as a portfolio of scientific applications and tutorials. The company says both are live through Cerebras Labs.

Michael James, Cerebras co-founder and chief architect of advanced technologies, still encourages users to learn Tungsten themselves. He said: "Someone has to keep the agents on track."

The release targets the programming burden, not a change to the processor itself. Cerebras says its goal is to let scientists program at a high level without becoming hardware specialists or giving up performance. Cerebras frames that shift as the practical issue for scientific computing, where specialized hardware remains difficult to use without low-level configuration.

Cerebras reports that SNAP, a standard proxy for deterministic nuclear radiation transport, reaches 1.6 PFLOP/s on one CS-4, ahead of 10,800 GPUs on the El Capitan supercomputer. The company also reports 1.04 PFLOP/s for HPCG, a sparse-workload supercomputing benchmark, on a single system.

What does Tungsten automate?

Cerebras says earlier compilers exposed non-scalar vector, stream and event capabilities through intrinsics. Under that model, developers had to configure descriptors, connect tasks to events and route the fabric manually.

Tungsten maps high-level code onto access-pattern registers, event-driven tasks, fabric calls and loop-bearing instructions, giving programmers a model for the wafer's unusual hardware operations without manual routing.

In Cerebras's Jacobi stencil example, each source line represents a loop over an entire block and compiles to one vector instruction. The example needs no descriptor setup or task wiring. Its <+- notation accumulates into a destination, similar to +=.

Under Cerebras's model, the compiler allocates access patterns, loop definitions, fabric sockets and microthreads across the whole program. Cost models and optimization passes swap those resources in and out during execution. Cerebras compares the approach with CPU register allocation and spilling. It says that ownership turns later restructuring into an edit to high-level code rather than a manual rewrite.

Cerebras describes the Wafer-Scale Engine as a dataflow machine with 900,000 cores, local memory and an on-wafer communications fabric. In that model, arriving data triggers the work that consumes it.

Access-pattern registers walk multidimensional arrays and fabric streams, so address arithmetic consumes no instructions. One instruction can take an entire loop as its operand. Cores can call functions on one another through fabric messages. Event-based control flow replaces polling and branching.

Cerebras says Tungsten exposes these capabilities at a higher level. The release leaves the processor unchanged; it changes the layer through which researchers address it.

What does Aviary include?

Cerebras says Aviary spans compact tutorials and full scientific applications. The portfolio includes:

  • Compact heat-equation and Gray–Scott reaction-diffusion stencils.
  • Particle transport, sparse solves and fast Fourier transforms.
  • Molecular dynamics, plasma simulation and high-order hydrodynamics.

Together, the examples cover both compact stencils and computationally heavier scientific workflows. Cerebras presents the range as evidence that high-level programming can reach multiple kinds of scientific work. The practical claim is broader than one successful stencil or benchmark.

The Gray–Scott tutorial contains about 50 lines of Tungsten, according to Cerebras. The company presents it as a smaller entry point alongside the full scientific codes. Its molecular-dynamics applications cover short-range Lennard–Jones interactions and tabulated embedded-atom potentials for metal alloys.

Cerebras says many of the applications grew out of work with Sandia, Lawrence Livermore and Los Alamos national laboratories under NNSA's Advanced Memory Technology program. The molecular-dynamics codes build on the collaboration's 2024 Gordon Bell Prize finalist work but remain separate codes. The earlier result used an early version of Tungsten.

How did the agent write the plasma code?

Cerebras presents the particle-in-cell application as the release's central agentic-coding example. One researcher and one coding agent produced it in about two weeks through a chat interface. Cerebras says the agent wrote every line.

The agent worked from published algorithms and validated its reference implementation against the open-source XOOPIC code. It modeled the problem on the EMPIRE plasma code. Cerebras says the agent also produced tests, figures and documentation while finding and fixing its own bugs.

The model did more than generate a code fragment. According to Cerebras, it assembled the application and its supporting work around published scientific references and existing plasma software.

Cerebras describes an electron beam entering argon plasma. In the simulation, particles move under electromagnetic fields and migrate between cores. An implicit solver computes fields on a decomposed mesh from the particles' positions and velocities.

Cerebras reports that the full run advanced 60 million particles at 7,200 time steps per second, with throughput reaching 430 billion particle updates per second on one system.

Cerebras says the Tungsten reference implementation is complete enough to support this agent-led workflow. That claim sits alongside James's insistence that people remain responsible for keeping agents on track.

Can the same code scale beyond one wafer?

Cerebras says the same Tungsten code runs on one wafer or across a cluster. Its global tsunami simulation used 64 wafers within an AI deployment. The company says such deployments scale to 2,048 systems.

Cerebras lists a design optimizer as forthcoming, not among the released applications. It would couple an AI model to a physics simulation acting as its world model. Both would run on the wafer, letting the AI model test designs against the simulation.

Both scale figures describe the existing programming model, not a new processor design. Cerebras is trying to make its specialized system easier to drive with scientific code and agents.

What has to prove next?

Cerebras says the practical significance is the programming model around the reported throughput. The key test is whether researchers and agents can move from Aviary's examples to their own codes without returning to intrinsics or hand-routed fabric. Cerebras says that capability would determine how broadly its specialized hardware can support scientific computing and AI-guided design.

The next proof point is whether Aviary users can carry their own scientific codes onto the wafer while leaving descriptors, events and fabric routing to Tungsten. Performance must remain intact for that easier workflow to matter, according to Cerebras's stated goal. The company will also need to show that its compiler decisions hold up as researchers' programs change and grow more complex.

For Cerebras, success now depends less on adding processor features than on making its existing 900,000-core system easier to program at scale.

Source: unite.ai

Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.

708 articles

Related articles

  1. Cerebras CEO Andrew Feldman to Tackle AI Scaling at TechCrunch Disrupt 2026
  2. Cerebras puts disaggregated AI inference into production
  3. KORA Unveils Linux Desktop With On-Device AI for Snapdragon PCs
  4. Cerebras Q3: Cloud Now 63% of Revenue, Microsoft Signs On
  5. Supermicro Ships 20-Plus H15 Servers Built on AMD EPYC 9006 Chips

« Previous articleNext article »