How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.
1.At a glance
SimpleStart here
The Wafer-Scale Engine, unveiled 19 August 2019, was the first chip to turn an entire silicon wafer into a single AI processor: 1.2 trillion transistors and 400,000 cores on 46,225 mm² of silicon 1. The system built around it, the CS-1, reached its first customer three months later — Argonne National Laboratory, on 19 November 2019 2.
Deep diveThe technical detail
Argonne’s Rick Stevens, associate laboratory director, said deploying the CS-1 “dramatically shrunk training time across neural networks, allowing our researchers to be vastly more productive” 2. This is the earliest Cerebras chip with a named, confirmed customer deployment rather than only a public unveiling — the bar this site applies before giving any chip its own page.
2.Launch and history
SimpleStart here
Cerebras spent its first years on one bet: that a whole silicon wafer, left uncut, could beat GPUs at training AI models. WSE-1, paired with the CS-1 system, was that bet’s first public proof 1.
Deep diveThe technical detail
A second named customer followed within a year. The National Science Foundation funded “Neocortex,” announced 9 June 2020, coupling two CS-1 systems (800,000 cores combined) with a conventional HPE Superdome Flex server for shared memory 3. It reached user access in mid-February 2021, built with Carnegie Mellon’s Pittsburgh Supercomputing Center 4.
3.What’s inside it
SimpleStart here
400,000 AI-optimized cores connected by Cerebras’s on-wafer “Swarm” communication fabric, with memory distributed next to the cores instead of sitting off-chip 1.
Deep diveThe technical detail
Cerebras’s own 2019 comparison: the WSE carried “3,000 times more” on-chip memory and “10,000 times more” memory bandwidth than “the largest graphics processing unit” available at the time, without naming that GPU specifically, and the WSE itself was “56.7 times larger” 1.
4.Full spec table
SimpleStart here
| Spec | Cerebras WSE-1 |
|---|
| Transistors | More than 1.2 trillion 1 |
|---|
| Cores | 400,000 AI-optimized cores 1 |
|---|
| On-chip memory | 18 GB SRAM 1 |
|---|
| Memory bandwidth | 9 petabytes/s 1 |
|---|
| Fabric bandwidth | 100 petabits/s aggregate (Swarm fabric) 1 |
|---|
| Chip size | 46,225 mm² 1 |
|---|
| Process node | TSMC 16 nm 1 |
|---|
Deep diveThe technical detail
Every headline figure roughly doubled or better within two years: WSE-2 (2021) reached 2.6 trillion transistors and 850,000 cores on the same 46,225 mm² wafer size, on a smaller 7 nm process (see that page on this site) 5.
5.Where it’s made
SimpleStart here
Designed by Cerebras; the wafer itself made by TSMC on its 16 nm process 1.
Deep diveThe technical detail
Cerebras’s 2019 materials do not compare its yield approach to competitors the way its later launches do (see the WSE-2 and WSE-3 pages on this site for those comparisons) — no equivalent yield claim was found for this generation in the sources checked for this page.
6.Which systems use it
SimpleStart here
Sold as the CS-1 system. Two confirmed early customers: Argonne National Laboratory (from 19 November 2019) 2, and the NSF-funded “Neocortex” system at the Pittsburgh Supercomputing Center, pairing two CS-1s with an HPE Superdome Flex server 3.
Deep diveThe technical detail
At Neocortex’s February 2021 user-access launch, PSC’s AI and Big Data director Paola Buitrago called it a “formidably strong collaboration between users, vendors and our outstanding PSC team” 4. Carnegie Mellon’s John Wohlbier framed the goal as testing “to what extent will AI-specific ASICs…enable model training and parameter studies on a much faster timescale” than typical data-centre or cloud resources 4.
7.Official pricing
SimpleStart here
Not published by the company. Neither Cerebras’s own 2019 unveiling nor the Argonne or Pittsburgh Supercomputing Center announcements give a CS-1 system price.
Deep diveThe technical detail
Cerebras has never published a system list price for any generation found on this site — the CS-3 and CS-4 pages on this site record the same “not published” result, so this is a consistent company policy, not a gap specific to the earliest chip.
9.What came before, what came next
SimpleStart here
Came before: nothing — this was Cerebras’s first chip.
Came after: WSE-2 (CS-2), covered on its own page on this site.
Deep diveThe technical detail
WSE-2 replaced this chip in April 2021, less than two years after WSE-1’s August 2019 unveiling — in that time this chip reached at least two named deployments (Argonne, then Neocortex) 5 2 3.
10.Hidden in plain sight
SimpleStart here
Cerebras’s first AI supercomputer needed help from a conventional computer to work: Neocortex paired two CS-1 wafer-scale systems with an ordinary HPE Superdome Flex server to handle shared memory 3 — an early version of the external memory unit (MemoryX) Cerebras still pairs with its wafers for large-model training today (see the WSE-3 page on this site).
Deep diveThe technical detail
Argonne’s CEO-level framing at deployment was broad, not narrow: Andrew Feldman said the CS-1 there was being used “to better understand everything from cancer drug interactions to the properties of black holes” 2 — general-purpose scientific computing, not a single named AI model or product, unlike the model-serving claims Cerebras makes for its current Inference cloud (see the Cerebras brand guide on this site).