Updated Thermal Model Description Rev 01 6/6/2017 Yueming Li, Thermal Engineer, Facebook John Fernandes, Thermal Engineer, Facebook Jia Ning, Hardware Engineer, Facebook
Changes based on community’s feedback ASIC Use a generic 2-resistor model by modifying junction to case & junction to board resistances Rjc = 0.4 C/W Rjb = 10 C/W PCB Rather than layer definition, PCB is modeled as bulk cuboid with orthotropic conductivity In-plane conductivity: 34 W/(m-K) Normal conductivity: 0.33 W/(m-K) QSFP model Update to a more representative/detailed model Use a blocker to prevent air from bypassing underneath the baseboard Based on the changes, updated models were uploaded with initial simulation results Special thanks to Dell, Mellanox & Intel for their suggestions and help with the model!
Updated Cases for CFD - Input
Simulation Results – Output Grid independence study was conducted for each model “D” cases report lower ASIC temperatures and higher QSFP temperatures compare to “A” cases Cold-aisle cases seem feasible even with smaller form-factor 7a QSFP temperature is the gating factor for hot-aisle cases DRAM appears to not be a gating issue
Simulation Results – Hot-aisle cases D5: hot-aisle, no DRAM, 20W ASIC, 150LFM
Simulation Results – Hot-aisle cases Approach temperature to QSFP modules is around 74 - 77°C D5: hot-aisle, no DRAM, 20W ASIC, 150LFM
Discussion Topics Bypass in between Mezz PCB top clearance (2.9mm) can be optimized by I/O shield (to be discussed) Stand-by power and cooling solutions Estimates place this value at ≤ 50% of max power Stand-by power for optical module? Case temperature spec of 85°C for QSFP modules seems improbable Mellanox propose to consider 70°C as a target during thermal simulations Large fail region for hot-aisle cases indicates need for better cooling of QSFP modules Consider adding heat sink to QSFP port/cage? Any existing design that can be leveraged? Wind tunnel specification and standard monitor points for vendors to share test data with adopters (system vendors)