Modeling the next-generation supercomputer for year 2030

• Profiled AMD Mi250 Accelerator using AMD's supercluster server, Omniperf, RocProf Compute
• Performed roofline analysis and compute to memory ratio calculations for different performance benchmarks (rocBLAS, oLLaMa, nBody)
• Mathematically modeled the performance improvement, power & cost trade offs using valid assumptions and given resource constraints
• Came up with 3 different supercomputer system suited for different tasks while meeting the cost and power requirement