返回新聞
產品展示AI Understanding 簡報

Cerebras 推出 Tungsten,這是一種用於其晶圓級引擎的高級語言和編譯器

Cerebras Systems 推出了 Tungsten,這是一種新的高階程式語言和編譯器,可抽象化其晶圓級引擎的複雜性,使科學家能夠在沒有低階硬體內在函數的情況下編寫程式碼。

6 min readRead the linked source
Source-provided image accompanying Cerebras launches Tungsten, a high‑level language and compiler for its wafer‑scale engine
來源參考來源記錄
出版商
unite.ai
來源類型
連結來源-主要來源狀態尚未確定。
背景60 秒內了解這一點

關鍵術語

基準測試
用於測量和比較模型性能的標準化測試或資料集。
推理
經過訓練的模型產生預測或輸出的運行時階段。
引文
模型回應中包含的來源段落或文件的引用,以支持其主張。
測試一下自己AI 模型解釋測驗

發生了什麼事

Cerebras Systems released Tungsten, a high‑level language and compiler for its Wafer‑Scale Engine, together with the Aviary portfolio of scientific applications and tutorials. The announcement highlights a plasma simulation written entirely by a coding agent that ran across 98% of a Cerebras wafer. Tungsten abstracts the wafer’s dataflow architecture—access‑pattern registers, event‑driven tasks, fabric calls and loop‑bearing instructions—into a concise syntax, allowing developers to express whole‑loop operations in a single line. The compiler automatically allocates resources, performs cost‑model‑driven optimizations, and maps code to the wafer’s 900,000 cores. Aviary includes examples such as heat‑equation stencils, Gray‑Scott reaction‑diffusion, particle transport, FFTs, molecular dynamics, and a plasma simulation that achieved 430 billion particle updates per second on a single system. Cerebras reports results: SNAP reaches 1.6 PFLOP/s on a CS‑4 system, surpassing 10,800 GPUs on the El Capitan supercomputer, and HPCG hits 1.04 PFLOP/s on a single wafer. The release is live through Cerebras Labs, and the same Tungsten code can run on a single wafer or scale across clusters, with a reported 64‑wafer global tsunami simulation and potential scaling to 2,048 systems.

Cerebras announced Tungsten, a new high‑level programming language and compiler designed specifically for its Wafer‑Scale Engine (WSE). The language abstracts the WSE’s dataflow architecture—access‑pattern registers, event‑driven tasks, fabric sockets, and loop‑bearing instructions—into a concise syntax that allows developers to express complex operations, such as whole loops, in a single line of code. The compiler automatically handles resource allocation, cost‑model‑driven optimizations, and mapping of code to the wafer’s 900,000 cores, eliminating the need for developers to manually configure descriptors, connect tasks, or route the fabric.

Alongside Tungsten, Cerebras released Aviary, a portfolio of scientific applications and tutorials covering heat‑equation stencils, Gray‑Scott reaction‑diffusion, particle transport, sparse solves, FFTs, molecular dynamics, plasma simulation, and high‑order hydrodynamics. The Gray‑Scott tutorial is about 50 lines of Tungsten, providing a low‑entry point for new users. The molecular‑dynamics examples include Lennard‑Jones and embedded‑atom potentials for metal alloys.

Cerebras highlighted a plasma simulation written entirely by a coding agent via a chat interface. The agent produced the full application—including code, tests, figures, and documentation—over roughly two weeks, validating against open‑source XOOPIC and modeling the problem on the EMPIRE plasma code. The simulation advanced 60 million particles at 7,200 time steps per second, achieving a throughput of 430 billion particle updates per second on a single wafer.

Performance benchmarks were presented: the SNAP proxy for deterministic nuclear radiation transport reached 1.6 PFLOP/s on a CS‑4 system, surpassing the combined performance of 10,800 GPUs on the El Capitan supercomputer; the HPCG achieved 1.04 PFLOP/s on a single wafer. Cerebras noted that the same Tungsten code can run on a single wafer or scale across clusters, citing a global tsunami simulation that used 64 wafers within an AI deployment and indicating potential scaling to 2,048 systems.

Cerebras mentioned a forthcoming design optimizer that would couple an AI model to a physics simulation running on the wafer, enabling AI‑driven design testing. The release is currently available through Cerebras Labs, but pricing and licensing details were not disclosed.

來源詳情: unite.ai

為什麼這很重要

Tungsten represents a significant shift in how researchers can harness Cerebras’s wafer‑scale hardware. By removing the need for low‑level intrinsics and manual fabric routing, the language lowers the barrier to entry for scientific computing on a platform that traditionally required hardware‑specialist expertise. The reported performance—1.6 PFLOP/s on SNAP and 430 billion particle updates per second—demonstrates that high‑level abstractions can retain the raw throughput advantages of the Wafer‑Scale Engine. This could broaden adoption of wafer‑scale processors beyond niche AI workloads to a wider range of HPC and simulation tasks, potentially accelerating AI‑guided design cycles and large‑scale scientific research. Moreover, the integration of a coding agent that can generate full applications via a chat interface showcases an early example of AI‑assisted software development on specialized hardware, hinting at future workflows where AI agents co‑design and optimize code in real time. However, the practical impact will depend on how easily external researchers can adopt Tungsten, the availability of the language and compiler, and whether the performance gains hold across diverse real‑world applications beyond the showcased benchmarks.

The high‑level abstraction provided by Tungsten could dramatically lower the expertise barrier for using wafer‑scale hardware, which has traditionally required deep knowledge of low‑level intrinsics and fabric routing. This opens the platform to a broader scientific community, potentially expanding its use beyond AI to HPC and simulation workloads.

The reported performance figures suggest that the abstraction does not sacrifice the raw computational power of the WSE. Achieving 1.6 PFLOP/s on SNAP and 430 billion particle updates per second demonstrates that high‑level code can still leverage the wafer’s massive parallelism and dataflow efficiencies.

The integration of a coding agent that can generate full applications via a chat interface showcases an early example of AI‑assisted software development on specialized hardware. If reliable, this could accelerate development cycles, reduce human error, and enable rapid prototyping of complex simulations.

Cerebras’s collaborations with national labs (Sandia, Lawrence Livermore, Los Alamos) and the reference to prior Gordon Bell Prize work lend credibility to the scientific relevance of the release, suggesting that the platform may be adopted for high‑impact research projects.

However, the impact will hinge on external accessibility: without clear licensing, pricing, or cloud‑based access options, adoption may remain limited to existing Cerebras customers or partner institutions.

Interactive Mechanism

互動機制:它實際上是如何運作的

以互動方式探索這項發展背後的基礎技術。

Agent Lifecycle Stage:
1
User Intent & Planning: "Audit customer refund request #4092 and settle payment."
2
Tool Calling: Emits structured JSON call crm_get_transaction(id='4092').
3
Guardrail & Verification:Paused: High-value action requires human operator sign-off.
4
Final Settlement: Refund recorded, email receipt dispatched, and audit log stored.
Core takeaway: An AI agent is not just a language model—it is a closed loop of planning, tool invocation, and environment feedback. Production systems require self-healing retries and strict human approval guardrails.
互動式概念檢查+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

接下來看什麼

Key indicators to monitor include the rollout of Tungsten to external research institutions, pricing and licensing terms for the language and compiler, and the availability of the promised design‑optimizer AI model. Adoption metrics such as the number of published papers or projects using Tungsten, as well as performance validation from independent labs, will reveal whether the high‑level abstraction truly preserves the wafer’s performance advantages. Additionally, the evolution of the coding‑agent workflow—its reliability, debugging capabilities, and integration with existing development tools—will be critical for broader acceptance. Finally, any announcements of expanded hardware support, such as multi‑wafer clusters or cloud‑based access, could further influence the ecosystem.

Announcements regarding the pricing model, licensing terms, and whether Tungsten will be offered as a cloud service or remain limited to on‑premise Cerebras hardware.

Independent validation of the results from academic or industry labs not affiliated with Cerebras, to confirm that the high‑level language truly preserves performance across diverse workloads.

Adoption metrics such as the number of external research groups publishing results using Tungsten, or of the Aviary examples in peer‑reviewed literature.

Progress on the promised design‑optimizer AI model, including its release timeline, integration details, and early performance or usability reports.

Updates on scaling capabilities, particularly real‑world deployments that use multiple wafers (e.g., the 64‑wafer tsunami simulation) and any cloud or data‑center offerings that enable broader access.

相關指引和測驗

覺得有用嗎?