Skip to main content
v2026.11,986 entries · CC-BY 4.0
CASRAIAsk CASRAI · AI assistant for researchAsk a research question. Answers come with their sources.Funder rules, compliance, methods and publishing. Answers link the sources they rest on.

Editorial · CASRAI · AI and ML research outputs

WaferTetris Preprint Optimizes Freeform Workload Mapping on Wafer-Scale AI Chips

Researchers from Edinburgh, Waterloo, and Toronto unveiled WaferTetris, a freeform compiler that boosts wafer-scale AI chip throughput by up to 4.42x.

Published 16 Nov 2026· 8 minute read

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

CASRAI is the reference for research administration — bookmark it for the next question.

Computer systems researchers from the University of Edinburgh, the University of Waterloo, and the University of Toronto released a preprint on November 13, 2026, detailing WaferTetris, an automated freeform mapping system designed to eliminate spatial fragmentation and memory underutilization on wafer-scale AI accelerators like the Cerebras Wafer-Scale Engine 3 (WSE-3).

Automated Freeform Placement Replaces Rigid Rectangular Grid Templates

Wafer-scale processors integrate hundreds of thousands of independent processing elements (PEs) across a single monolithic silicon wafer. Programming these architectures has historically required developers to rely on rigid rectangular templates known as PE islands. In their research paper submitted to the arXiv preprint repository on November 13, 2026, lead author Dayou Du, alongside Congjie He, Andrew Boutros, Luo Mai, Vaughn Betz, and Jianyi Cheng, demonstrated that rectangular layouts leave large volumes of compute and on-chip memory stranded, bottlenecking large language model (LLM) inference and scientific simulations.

WaferTetris removes geometric layout restrictions through a flexible, Tetris-like workload packing strategy. Rather than forcing operations into predefined rectangular tiles, the compiler generates fully unrolled layout descriptions in Cerebras Software Language (CSL). This allows individual processing tasks to occupy non-rectangular, interlocking geometries across the wafer fabric while preserving existing PE-task kernel code. By decoupling spatial allocation from rigid shapes, WaferTetris enables workloads to consume previously idle compute cores and harvest local static random-access memory (SRAM) across the wafer.

Technical Architecture: Recurrence Throughput Modeling and Per-Color Gating

To automate spatial mapping without requiring days of physical hardware compilation or simulation, the research team solved two core algorithmic challenges: predicting dataflow pipeline performance and enforcing wafer-scale network-on-chip routing rules.

First, existing spatial mapping tools commonly evaluate candidate designs using a slowest-unit bottleneck proxy, assuming pipeline stages execute independently. While this assumption holds for simple linear pipelines, modern neural architectures feature branch-and-join topologies, such as transformer residual connections where tensors fork across attention and multi-layer perceptron (MLP) blocks before rejoining. The authors showed that slowest-unit proxies overpredict throughput by 199% to 227% because waiting times propagate across dependent paths. WaferTetris replaces this heuristic with a recurrence-based throughput model grounded in classic dataflow analysis. By formulating pipeline scheduling around the Maximum Cycle Mean (MCM) across directed graph recurrences, WaferTetris predicts end-to-end execution throughput within 8.4% of hardware measurements, guiding search algorithms directly toward high-performance configurations.

Second, the team bridged the gap between wafer-scale routing constraints and mature field-programmable gate array (FPGA) place-and-route CAD algorithms. The authors adapted the open-source Verilog-to-Routing (VTR 9.0) toolchain developed under Professor Vaughn Betz at the University of Toronto. On Cerebras WSE architectures, each processing element contains an integrated 2D-mesh router governed by 21 hardware virtual channel colors, and any message traversing a PE consumes that router’s color resources. Direct application of standard FPGA routers fails because FPGA logic blocks feature interchangeable pins and routing channels that bypass logic blocks without consuming internal switches.

WaferTetris resolves this mismatch by constructing an architecture model with a crossbar lookup table and a color-gate bank. Each modeled PE router interface contains unit-capacity gates for every virtual channel color. Any network traffic produced, consumed, or forwarded through a PE must pass through that specific color gate, converting physical WSE routing rules into unit-capacity constraints that PathFinder negotiated-congestion routing algorithms resolve through rip-up and reroute cycles. To ensure compilation scalability across 900,000 PEs, WaferTetris partitions the wafer into an 8 × 4 grid of balanced subgraphs, completing placement, routing, and CSL code generation for a 13-billion-parameter model in approximately 37 minutes.

Empirical Benchmarks on Cerebras WSE-3 Hardware

The research team evaluated WaferTetris on commercial Cerebras WSE-3 hardware, comparing performance against CU.POKer (a template-based automated placer presented at ICCAD) and WaferLLM (a manually tuned wafer-scale inference engine published at USENIX OSDI 2025). Benchmarks spanned full-model LLM decoding across Qwen2.5-7B, LLaMA3.1-8B, and LLaMA2-13B, as well as multi-physics scientific stencils.

Against CU.POKer on matched fabric allocations, WaferTetris achieved 3.24× to 4.42× higher system decoding throughput, increasing PE utilization from 53.7–85.0% up to 93.2–98.0% (reserving a 2% margin for routing resources). On scientific computing stencils, WaferTetris achieved a 1.26× speedup on Gray-Scott reaction-diffusion and 1.18× to 1.21× speedup on elastic wave propagation, raising PE utilization to over 96%.

Benchmark Workload Baseline System WaferTetris PE Utilization Baseline PE Utilization Throughput / Capacity Gain
LLaMA3.1-8B Decoding CU.POKer (Template) 98.0% 59.8% 3.24–4.42× system throughput
LLaMA2-13B (4k Context) CU.POKer (Template) 98.0% Out-of-Memory (Failed) 76.8k tokens/s (CU.POKer OOM)
LLaMA3.1-8B Decoding WaferLLM (Manual) 98.0% 60.1% 1.04–1.10× throughput; 1.8× context (80k vs 44k)
Gray-Scott Stencil CU.POKer (Template) 96.3% 75.0% 1.26× lattice updates/s
Elastic Wave Stencil CU.POKer (Template) 97.6% 83.3% 1.18–1.21× lattice updates/s

Compared with WaferLLM, WaferTetris delivered 1.04× to 1.10× per-user decoding throughput while supporting 1.8× the context window length for LLaMA3.1-8B (80,000 tokens versus 44,000 tokens). On LLaMA2-13B under constrained fabric sizes where WaferLLM failed due to out-of-memory errors from rigid rectangular allocations, WaferTetris successfully scheduled the workload by harvesting unutilized SRAM blocks.

Background: Wafer-Scale Integration and Spatial Fragmentation

Standard GPUs and accelerator ASICs are constrained by lithography reticles to roughly 850 square millimeters, requiring multi-chip clustering over networks that introduce latency. Cerebras Systems bypassed this boundary with its Wafer-Scale Engine series, etching an entire silicon wafer as a single contiguous compute fabric measuring 46,225 square millimeters.

The Cerebras WSE-3 features 900,000 AI vector cores manufactured on TSMC’s 5-nanometer process, packing 44 gigabytes of on-chip SRAM with 21 petabytes per second of memory bandwidth and 214 petabits per second of fabric bandwidth. Because the chip contains no external DRAM, each core relies strictly on its local 48 kilobytes of SRAM (roughly 40 kilobytes usable for data). In spatial architectures, usable memory capacity is directly proportional to the number of physical cores occupied.

Under template-based mapping, when an application cannot fit rectangular PE islands onto the mesh, developers must shrink island dimensions, leaving tens of thousands of PEs idle. On WSE-3, the authors noted that even manual frameworks like WaferLLM left approximately 344,000 out of 862,500 usable PEs unallocated during certain execution phases. This spatial fragmentation created an artificial memory wall, causing key-value caches to exceed local storage limits prematurely.

Research Community Reaction and Architectural Analysis

Computer systems researchers and electronic design automation specialists highlighted the preprint as a major milestone in unifying compiler toolchains with spatial supercomputing hardware. Demonstrating that mature FPGA place-and-route algorithms can adapt to wafer-scale fabrics establishes an automated physical compilation pathway for extreme-scale chips.

Hardware architects noted that treating the Cerebras WSE as an FPGA variant with virtual channel colors mapped to physical routing tracks establishes a blueprint for future spatial architectures, including multi-die wafer-level packaging and coarse-grained reconfigurable arrays. Systems researchers emphasized that the recurrence-based throughput model provides a mathematical alternative to ad-hoc heuristics, ensuring spatial parallelism searches do not stall at false bottleneck estimates when running a synthetic benchmark or production deep learning model.

Implications for Academic Supercomputing and Research Computing Facilities

Automated freeform mapping carries operational implications for academic research institutions, high-performance computing centers, and university research computing offices managing wafer-scale systems.

National computing facilities, including the Pittsburgh Supercomputing Center (hosting the NSF-funded Neocortex system) and the Argonne Leadership Computing Facility, deployed Cerebras systems to accelerate computational fluid dynamics and foundation model training. Academic researchers accessing these clusters through programs such as the NAIRR Pilot frequently encounter steep programming hurdles when porting domain-specific kernels. Developing custom CSL layouts previously demanded months of assembly optimization and geometric floorplanning.

WaferTetris lowers this barrier by automating layout compilation from high-level operator graphs. Research software engineers working with performance-portability libraries like Kokkos can target wafer-scale fabrics without manually orchestrating message routing coordinates or sacrificing PE occupancy. Furthermore, the 1.8× expansion in usable context length demonstrated on LLaMA3.1-8B directly enables universities to host long-context scientific document analysis and protein sequence modeling on on-premise wafer-scale hardware without resorting to multi-node GPU clusters.

Development Roadmap and Upcoming EuroSys 2027 Presentation

The authors confirmed that WaferTetris has been accepted for formal peer-reviewed presentation at EuroSys 2027, the ACM European Conference on Computer Systems, scheduled for April 19–23, 2027, in Rabat, Morocco. The researchers plan to expand the framework to support distributed training pipelines and dynamic tensor parallelism across multi-wafer SuperPods.

The team is also exploring integration with the MLIR compiler ecosystem to establish an end-to-end compilation flow from PyTorch down to native CSL binaries. By releasing their mapping abstractions and recurrence models to the academic research community, the authors aim to accelerate software-hardware co-design across next-generation wafer-scale and chiplet-based AI computing architectures.

Sources

  • Du, D., He, C., Boutros, A., Mai, L., Betz, V., and Cheng, J. (2026-11-13). WaferTetris: Efficient Freeform Mapping on Wafer-Scale Accelerators. arXiv preprint arXiv:2611.18527. https://arxiv.org/abs/2611.18527
  • Cerebras Systems. (2026). The Future of AI is Wafer Scale: Cerebras Wafer-Scale Engine 3 Architecture and Specifications. https://www.cerebras.ai/chip
  • Cerebras Systems. (2025). Cerebras Software Development Kit (SDK) and CSL Architecture Documentation. https://sdk.cerebras.net/
  • He, C., Huang, Y., Mu, P., Miao, Z., Xue, J., Ma, L., Yang, F., and Mai, L. (2025-07-10). WaferLLM: Large Language Model Inference at Wafer Scale. 19th USENIX Symposium on Operating Systems Design and Implementation (OSDI 25). https://www.usenix.org/conference/osdi25/presentation/he
  • Betz, V., and Rose, J. (2026). Verilog-to-Routing (VTR): Open-Source CAD for FPGAs and Spatial Architectures. University of Toronto. https://verilogtorouting.org/
  • Pittsburgh Supercomputing Center. (2026). Neocortex: Exploring Wafer-Scale AI and High Performance Computing for Academic Research. National Science Foundation. https://www.psc.edu/resources/neocortex/

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →