Neocortex Call for Proposals – Spring 2023

Important dates

Eligibility

Eligibility criteria are as follows. Please contact us at neocortex@psc.edu if you have questions about your eligibility to apply.

• Projects must be non-proprietary.
• Applicants must be a researcher or educator at a US academic or non-profit research institution.
• A principal investigator (PI) is required. PIs may not be a high school, undergraduate or graduate student; a qualified advisor must serve in this capacity. Postdocs are eligible to serve as PI.

Successful applicants must agree to abide by the Neocortex Acceptable Use Policy in order to gain access and use the system.

Awardees are wholly responsible for conducting their research project activities and preparing results for publication.

For more information about Neocortex, see the Neocortex home page.

How to Submit

Submissions are being accepted in four tracks. Once you have identified the track that your project falls under, follow the EasyChair submission link above to address some general questions about your project. Submit your application document in PDF format. In the application document, you are expected to address all the pertinent track-specific questions listed in the section below. 

Make sure to address the corresponding track-specific questions listed in the next section in your application document. Failure to do so may subject your application to being rejected without review.

If you have any questions or need additional information, feel free to contact us at neocortex@psc.edu.

Tracks

Track 1: ML workflows leveraging model(s) as-is from the Cerebras Modelzoo
Track 2: ML workflows leveraging model(s) similar to those of Cerebras Modelzoo
Track 3: General purpose projects leveraging the Cerebras SDK
Track 4: Domain specific programming on structured grids leveraging the specialized WFA API

Track 1: ML workflows leveraging model(s) as-is from the Cerebras Modelzoo

This track is for researchers that intend to use at least one of the DL models as-is offered in the Cerebras Modelzoo. The models are supported via TensorFlow and PyTorch.

Some of the ML models available in the Cerebras modelzoo are BERT (standard, classifier, name entity recognition, summarization, question answering), GPT-2, GPT-3, GPT-J, Linformer, RoBERTa, T5, Transformer, FC-MNIST, 2D Unet.

If your project falls under this category, make sure to address the following questions in your application document:

1. Please indicate which model(s) from the modelzoo do you intend to use. Do you anticipate being interested in adjusting the model architecture?
2. Please describe the dataset you are intending to use.
3. How big is the dataset of interest (total dataset size, number of samples, and sample size in MB)?
4. Please elaborate on the readiness of the dataset of interest. Is it fully available at this time? If not, how soon would it be fully available?
5. Please specify the shapes of the input and output tensors for your model/s.
6. If possible, please specify the name of the dimensions for your input and output tensors from the previous question. E.g. (batch, input channels, height, width)
7. Please specify the loss function that you would like to use.
   
8. Please list the libraries complementary to standard PyTorch and/or TensorFlow distributions that you would need to train your model(s).
9. Please list the key libraries that you would need for data preprocessing.

Track 3: General purpose projects leveraging the Cerebras SDK

This track is for researchers that intend to leverage the Cerebras SDK to develop custom compute kernels. The SDK is for researchers who want to explore the fundamental mapping of algorithms to a gigantic rectangular mesh of independent processors with a unique memory architecture. Ongoing research work with the SDK includes a collaboration with an energy company to implement a finite difference method for seismic wave equations and implementation of Graph 500 benchmark algorithms.

Some potential research avenues of interest are the implementation of HPC benchmark algorithms, such as HPL (high performance LINPACK) and HPCG (high performance conjugate gradient), optimal mappings of sparse tensor operations, linear solvers, particle methods, or graph algorithms, for instance.

The SDK can also be used to explore the implementation of custom machine learning kernels on the CS-2. Note however, that the SDK is not currently interoperable with the CS-2’s PyTorch and TensorFlow frontends, and can only be used to write standalone code. Also note that existing code written in other languages cannot be used as-is.

To better understand what leveraging the SDK entails, it can be helpful to envision what CUDA development would look like for a specific GPU architecture. 

If your project falls under this category, make sure to address the following questions in your application document:

1. Using the SDK requires programming in CSL, a C-like language designed specifically for the problem of massively parallel programming on the CS-2. What is your experience with HPC programming paradigms and languages, such as MPI, OpenMP, CUDA, OpenCL, etc.?
2. What are the underlying computational algorithms you’re interested in exploring? What existing software packages or libraries use these algorithms?
3. How is this problem bottlenecked on current hardware? Is the problem more bottlenecked by memory bandwidth, or communication costs associated with scaling up across distributed compute nodes?
4. What range of problem sizes are you interested in addressing? For example, how much memory does your problem use? How does memory usage or program run time scale with problem size?
5. What portion of your algorithm do you plan to port to the CS-2? Why are you interested in exploring this part of your algorithm?
6. The CS-2 offers native half and single precision data types. What precision does your algorithm or use case need?
7. The CS-2 is a network-attached accelerator. At a high level, the CSL programming model is similar to that of CUDA, in which data is moved between host CPU nodes and the device (CS-2) on which computational kernels are launched. How often will data need to be moved between the wafer and the worker nodes?
8. Describe your general plan to map your problem onto 850,000 cores. In answering this question, it might be helpful to recall some details of the CS-2 architecture. The 850,000 cores are laid out in a mesh, with each core connected on fabric to its four nearest neighbors on the East, South, North, and West. Memory is distributed among the cores, with each core having 48 KB of local memory. 64-bit local memory reads and writes take roughly a cycle, as does sending and receiving 32-bit messages between the four neighboring cores.

Track 4: Domain specific programming on structured grids leveraging the specialized WFA API 

This track is for researchers that intend to advance a project that would map to a regular grid and involve data movement patterns that are nearest-neighbor only on the grid, leveraging the WFA API (collaboratively developed by NETL and Cerebras). An example of the kind of work and performance gains achievable via the WFA API  is the recent work enabling unprecedented CFD simulations. The ideal project would involve a set of PDEs discretized to first/second order in implicit or explicit methods. The best performance gains will be with highest data intensity (or lowest algorithmic intensity). Applications could include CFD, geomechanics, electrodynamics, and some materials problems.

As you prepare your application document, it might be helpful to recall some details of the CS-2 architecture. The 850,000 cores are laid out in a mesh, with each core connected on fabric to its four nearest neighbors on the East, South, North, and West. Memory is distributed among the cores, with each core having 48 KB of local memory. 64-bit local memory reads and writes take roughly a cycle, as does sending and receiving 32-bit messages between the four neighboring cores. The memory capacity of the entire WSE is 40 GB.If your project falls under this track, make sure to address the following questions in your application document:

1. Can your problem lay out on a Hex grid (3d or many 2d parallel)?
2. Does your problem involve Spatial Locality?
3. Is your project Data Intense? Please, elaborate.
4. Is your problem compatible with single precision and with a total data volume not exceeding 40 GB?
5. Which of the following applies to your problem/project?
   • Computational Fluid Dynamics (FVM, FDM, FEM, LBM)
   • Structural Mechanics
   • Geomechanics
   • Weather/Climate
   • Materials – Ising Model, Density Functional Theory
   • CNN/RNN inference
6. Are you willing to commit to the following as you advance your project?
   • Build a Python class that imports the WFA and contains a “Library” to solve your scientific problem
   • Post on a public GitHub

References

 Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … & Polosukhin, I. (2017). Attention is all you need. In Advances in neural information processing systems (pp. 5998-6008).