This website uses cookies that are necessary to deliver an enjoyable experience and ensure
its correct functionality and cannot be turned off. Optional cookies are used to improve the page with analytics, by
clicking “Yes, I accept” you consent to this use of cookies. Learn more
Welcome to the Huawei Zurich Tech Arena, where innovation meets challenges in the world of AI and network technology!
This year, we’re excited to present two groundbreaking challenges designed to push the boundaries of model compression and network scalability. Compete with the best minds in tech as you explore solutions that could redefine the future of large-scale AI processing.
With a Docker environment ready to go, all the tools are at your fingertips. Join us at Huawei Zurich Tech Arena to show off your skills, compete for the top spot, and make your mark on the future of AI and network innovation!
with industry leaders and top innovators from across Europe.
Showcase your expertise
in front of a leading tech company and its decision-makers.
Get hands-on experience
solving challenging problems in the world of AI and network technology!
Gain insights
into the latest technologies and trends shaping the future of AI.
Compete
for a chance at a €28,000 prize pool and potential career opportunities within Huawei.
Network
with like-minded individuals across multiple countries.
WHO CAN PARTICIPATE?
Students
Apply only if you are in a Master's or PhD program.
Students from Switzerland, the UK, and Germany are invited to register! Conquer alone, no teams allowed.
Studying...
Networking & Communication Systems
Computer Science
Computer Engineering
Software Engineering
Data Science
Electrical Engineering
Computer Systems
Information Technology
Systems Engineering
Machine learning
Artificial intelligence
Computational Mathematics
CHALLENGES
Challenge 2: Communication affined Direct Topology of NPUs
Participants will be asked to define a general direct network topology framework that defines the connections among switch nodes, such that the cluster can achieve the maximum network scale（supports maximum number of GPUs in the cluster / approximately approaches the Moore bound).
Modeling the communication efficiency of the network topology for AllReduce and AlltoAll primitives；Devise the AllReduce/AlltoAll algorithm that achieves the ideal/optimal performance.
Describe the modularity of the network topology (how it can be constructed into practice).
Compare the proposed network topology with classical topologies (e.g., CLOS, Dragonfly, Dragonfly+) regarding their advantages, disadvantages, and application scope.
The challenge consists in implementing one-shot compression of the Llama-3.1-B model with the goal of achieving the highest compression rate at the lowest accuracy degradation with respect to the original model.
To implement one-shot compression, participants are allowed to implement pruning and weight-only quantization, but cannot re-train and fine-tune the compressed model to improve its accuracy.
Participants are required to integrate their solution in the popular lm-evaluation-harness benchmarking framework. A docker container is provided that contains the model development and evaluation environment.
Code that implements the model compression algorithm (estimated size in the order of the KBs). This is needed to check that the reduced model is programmatically generated and is not the result of fine-tuning and retraining.
Compressed model (estimated size in the order of the GBs), i.e, weights, biases, activations etc. This is needed to run inference on the platform.
Code to dequantize the model (estimated size in the order of the KBs). This is needed together with the compressed model to know how to load the model parameters.
Deliverables:
Describe how a direct network topology can be constructed with switch nodes of Radix n, calculate the maximum scale the network can achieve.
Modeling the communication efficiency of the network topology for AllReduce and AlltoAll primitives；Devise the AllReduce/AlltoAll algorithm that achieves the ideal/optimal performance.
Describe the modularity of the network topology (how it can be constructed into practice).
