---
type: "post"
id: "high-level-code-wafer-scale-performance-f05cd0"
slug: "high-level-code-wafer-scale-performance-f05cd0"
title: "High-level code, wafer-scale performance"
description: "Cerebras has released Tungsten, a language and compiler for its Wafer-Scale Engine, alongside Aviary, a portfolio of scientific applications and tutorials."
retrieval_nugget: "Cerebras has released Tungsten, a language and compiler for its Wafer-Scale Engine, alongside Aviary, a portfolio of scientific applications and tutorials. The release aims to let scientists express algorithms at a higher level while the compiler handles specialized hardware resources—a significant change in how teams may approach programming wafer-scale systems."
published_at: "2026-11-17"
updated_at: "2026-11-17"
record_date: "2026-11-17"
date_kind: "published_at"
topics: ["compute","ai-infrastructure"]
entities: ["Cerebras","Tungsten","Aviary","Wafer-Scale Engine","XOOPIC","Sandia National Laboratories","Lawrence Livermore National Laboratory","Los Alamos National Laboratory"]
related_patterns: ["agent-ready-software-exposes-capabilities"]
novelty_label: "new_fact"
what_changed: "Cerebras released Tungsten, a high-level language and compiler for its Wafer-Scale Engine, alongside Aviary applications and tutorials."
prior_refs: []
source_url: "https://www.cerebras.ai/blog/high-level-code-wafer-scale-performance"
source_urls: ["https://www.cerebras.ai/blog/high-level-code-wafer-scale-performance"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:high-level-code-wafer-scale-performance-f05cd0"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/high-level-code-wafer-scale-performance-f05cd0/","markdown":"https://newruntime.com/posts/high-level-code-wafer-scale-performance-f05cd0.md","json":"https://newruntime.com/posts/high-level-code-wafer-scale-performance-f05cd0.json"}
markdown_url: "https://newruntime.com/posts/high-level-code-wafer-scale-performance-f05cd0.md"
---

# High-level code, wafer-scale performance

## Retrieval answer

Cerebras has released Tungsten, a language and compiler for its Wafer-Scale Engine, alongside Aviary, a portfolio of scientific applications and tutorials. The release aims to let scientists express algorithms at a higher level while the compiler handles specialized hardware resources—a significant change in how teams may approach programming wafer-scale systems.

## Summary
Cerebras has released Tungsten, a language and compiler for its Wafer-Scale Engine, alongside Aviary, a portfolio of scientific applications and tutorials. The release aims to let scientists express algorithms at a higher level while the compiler handles specialized hardware resources—a significant change in how teams may approach programming wafer-scale systems.

## What happened
Cerebras announced Tungsten and Aviary on November 16, 2026. Tungsten is a language and compiler for programming the company’s Wafer-Scale Engine. Aviary is a collection of applications and tutorials written with Tungsten. The announcement says the tools are live at labs.cerebras.ai; it does not state a price or describe eligibility or access limits.

The announcement describes the Wafer-Scale Engine as a dataflow machine with 900,000 cores, local memory, and an on-wafer fabric. Data arriving at a core triggers the work that consumes it. Access pattern registers handle multidimensional arrays and fabric streams, while event-based control flow replaces polling and branching. Cerebras says that earlier programming methods exposed these mechanisms through intrinsics, leaving programmers to configure descriptors, connect tasks to events, and route the fabric manually.

Tungsten’s compiler models those hardware features and maps high-level code onto them. In the company’s description, programmers can write code closer to the mathematical form of an algorithm; the compiler assigns resources such as access patterns, loop definitions, fabric sockets, and microthreads. Cerebras uses a Jacobi stencil tutorial to illustrate the approach, saying each line represents a loop over a block and compiles to a single vector instruction. It compares compiler management of wafer resources to how compilers allocate CPU registers.

Aviary includes scientific applications, benchmarks, and smaller tutorial examples. Cerebras reports that SNAP, a proxy for nuclear radiation transport codes, runs at 1.6 PFLOP/s on a single CS-4; HPCG reaches 1.04 PFLOP/s on a single system; and its wafer-scale molecular-dynamics code runs 748 times faster than Frontier. A design optimizer coupling an AI model with a physics simulation is described as coming soon.

The release also recounts a particle-in-cell plasma simulation built by one researcher with a coding agent over about two weeks. Cerebras says the agent wrote tests, figures, and documentation, validated a reference against the open-source XOOPIC code, and fixed bugs. The full run advanced 60 million particles at 7,200 time steps per second, or 430 billion particle updates per second, on one system. The same Tungsten code can run on one wafer or across a cluster, according to the announcement.

## What the archive adds
New Runtime’s earlier coverage, [How Cerebras Deploys Trillion-Parameter AI Models](https://newruntime.com/posts/how-cerebras-deploys-trillion-parameter-ai-models-06db8d/) (October 22, 2026), described Cerebras’s multi-wafer deployment method for Kimi K3. Tungsten and Aviary are a new event: this release shifts the focus from deploying large models to programming scientific workloads on the wafer-scale platform.

## Trend signal
This release supports [agent-ready software exposes capabilities](https://newruntime.com/patterns/agent-ready-software-exposes-capabilities/) only weakly. Tungsten’s higher-level abstractions could make specialized hardware more accessible to coding agents as well as scientists, and Cerebras reports an agent-built simulation. But the announcement does not describe a permissioned interface or a general capability boundary for external agents, so this is not yet strong evidence of the pattern.

## Why now
Cerebras says a researcher ran a plasma simulation across 98% of a wafer last month, with a coding agent writing every line. The company presents Tungsten and Aviary as the tools behind that demonstration, turning an internal result into a public release.

## Connections to other events
Cerebras says Aviary’s molecular-dynamics applications build on work with Sandia, Lawrence Livermore, and Los Alamos national laboratories under the NNSA’s Advanced Memory Technology program. The release also connects to the company’s earlier multi-wafer deployment work: it says a global tsunami simulation ran on 64 wafers, and that deployments scale to 2,048 systems.

## Why it matters
For scientific teams, the compiler’s central promise is a lower barrier to writing code for specialized hardware without giving up its performance. If the abstraction works across real applications, researchers may spend less time managing descriptors, event wiring, and communication routes—and more time expressing and testing the scientific algorithm. The fact that the same code is said to run on one wafer or a cluster could also make it easier to move from a local experiment to larger deployments.

The agent-built plasma simulation is a useful signal that coding agents can contribute to highly specialized scientific software when paired with an adequate reference and validation workflow. It is not, by itself, evidence that the approach generalizes: teams still need domain expertise to judge the model, validate results, and decide whether a performance claim fits their workload.

## What happens next
- **Cerebras will publish or demonstrate the promised AI-coupled design optimizer in Aviary.** (6 months, 65%). Watch for: an Aviary release or demonstration showing the model and physics simulation running together.

## Limits and uncertainty
The performance figures and comparisons in the announcement are Cerebras claims; workloads, configurations, and measurement methods are not fully detailed for every result, and they have not been independently checked here. The article does not establish how broadly Tungsten’s abstractions apply across scientific workloads, or what expertise is still required to debug and validate programs. It also does not state pricing or access terms. The agent-written simulation is one reported project, not a measured study of coding-agent reliability or productivity.

## Skill to build now
A team evaluating Tungsten can start with Aviary’s smaller tutorial examples and compare the high-level code with the compiled behavior and performance on a representative stencil or simulation workload. Keep a known reference implementation and correctness checks in the loop; Cerebras’s plasma example explicitly describes validation against XOOPIC. This makes it possible to assess both whether the abstraction reduces hardware-specific work and whether the resulting scientific output remains trustworthy.

## Reader takeaway
Tungsten and Aviary package Cerebras’s attempt to make wafer-scale scientific computing accessible through a higher-level compiler and reusable examples. The release is promising for teams with relevant hardware, but its strongest performance and agent-coding claims remain vendor-reported demonstrations rather than independently established general results.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Compute](https://newruntime.com/topics/compute.md): Explore the compute topic hub.
- topic hub: [Ai Infrastructure](https://newruntime.com/topics/ai-infrastructure.md): Explore the ai-infrastructure topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
