# Thermal Kit for GPU HBMs

### Reduce Thermal Throttling, Unlock 3X ROI

**Higher Sustained GPU Performance, Improved Overall Cluster Stability**

Phononic’s GPU HBM cooling solution delivers up to 0.15 PUE savings, 5X lifetime improvement, 40% greater compute performance, and a 3X ROI over five years. It’s a solution specifically engineered for rapidly evolving AI training and inference workloads, and the unique requirements of HBM 3 and later generations. The GPU cooling solution maintains stable operating conditions under dynamic thermal loads, unlocking higher performance through reductions in thermal throttling.

This solution is readily deployable with existing nodes and racks.

### Benefits of Phononic GPU HBM Cooling

- Energy Efficiency  
- Performance Gains  
- Device Life

### Up to 0.15 PUE Savings

Millisecond action times enables cooling precisely when, where and for the exact duration required. Cooling overshoot can now be eliminated, and a new era of precision thermal control is unlocked. Maintain optimal operating envelopes during sustained high-compute workloads through throttle-free performance that significantly improves PUE.

### Up to 40% Greater Compute Performance

The software integration of Phononic’s Thermal Kit enables precision control of HBM temperatures, in real-time. By decoupling xPU and HBM temperatures, and maintaining those temperatures as workload increases, the millisecond response of TECs unlocks more performance.

Avoidance of pre-throttle performance divergence and recovery of throttle window further pay performance dividends, unlocking potential previously hindered by thermal throttling.

### Up to 5X Device Life Extension

Thermal telemetry and system-level connectivity improves operational predictability, enabling better workload orchestration and enhances rack-level efficiency, predictively. The avoidance of these device stresses unlocks up to 5X device life extension, and improved uptime.

## Applications

### xPUs and Switch ASICs

Increase TDP while Maintaining Today’s Thermal Management Architecture

Most data centers still rely on air cooling, using fans and conventional heat exchangers that are nearing their thermal limits as rack densities and chip power rise. The Thermal Design Power (TDP) of xPUs and switch ASICs continues to increase with higher core counts and port densities, driving the need for more efficient thermal management. Traditional upgrades—larger heat exchangers, taller chassis, and increased airflow—come with trade-offs like reduced rack space and higher fan noise. Thermoelectric cooling offers a compact alternative, boosting chip TDP by up to 20% without changing airflow or system architecture. These millimeter-thick devices fit standard server and switch layouts, activating only when needed to maintain safe operating temperatures. This approach extends the viability of existing air-cooled infrastructure while supporting next-gen semiconductor performance.

### Improve performance with selective cooling of multi-chip modules and memory

Advanced semiconductor packaging is enabling multi-chip modules with compute and memory die, each operating at distinct optimal temperatures. Traditional cooling methods often overcompensate or accept performance trade-offs to manage this thermal complexity.

Cooling strategies like Direct Liquid Cooling (DLC) with cold plates are commonly used, but maintaining ideal temperatures across diverse components often requires lowering coolant temperatures, increasing energy consumption. Phononic’s approach uses thermoelectric chips (TECs) integrated with cold plates and control software to thermally decouple compute and memory die. This allows HBMs to operate at higher bandwidth and dissipate more power without exceeding thermal limits.

## Contact Us to Unthrottle Your Data Center Performance