The programme explores how model, quantisation, context, traffic and hardware choices interact. Each workstream targets measurable improvements while preserving security and workload requirements.
Research questionHow should model and runtime settings adapt to a specific GPU topology?
Target outcomeA reproducible configuration profile for each hardware class.
Measurable metricsThroughput, latency, utilisation, cost per token
Research questionHow can very large models serve long context while minimising memory and preserving response quality?
Target outcomeHigher usable context and concurrency within fixed capacity.
Measurable metricsTTFT, GPU memory, maximum context, concurrent sessions
Research questionWhich partitioning strategy best balances memory, communication and compute?
Target outcomeEfficient execution of models that exceed one GPU.
Measurable metricsInterconnect traffic, tokens/s, balance, memory headroom
Research questionWhat precision reduction is acceptable for a specific task and model?
Target outcomeLower memory and compute cost within a defined quality budget.
Measurable metricsQuality delta, memory, throughput, energy/token
Research questionHow can cache allocation adapt to context and concurrency?
Target outcomeMore stable capacity under mixed request shapes.
Measurable metricsCache hit/use, memory fragmentation, concurrency, latency
Research questionWhen does a draft model reduce latency without excess overhead?
Target outcomeFaster generation on suitable workloads.
Measurable metricsAcceptance rate, inter-token latency, energy/token
Research questionWhich configuration meets service requirements with the least energy?
Target outcomeLower energy and infrastructure cost per useful output.
Measurable metricsEnergy/token, utilisation, quality, total cost
Research questionCan candidate configurations be compared consistently before deployment?
Target outcomeA workload-specific recommendation supported by evidence.
Measurable metricsBenchmark coverage, repeatability, quality, latency, throughput, cost