This OCP Global Summit 2025 session from Solidigm and Iceotope argues that liquid cooling for SSDs is no longer optional — PCIe Gen 5 SSDs already strain air cooling limits, and PCIe Gen 7 devices projecting 60-70W will make air cooling impossible. Solidigm presents engineering solutions for hot-swap cold plate integration and dual-sided PCB cooling, while Iceotope demonstrates its precision immersion dielectric fluid approach achieving up to 950+ kW per rack. The session concludes with a call for new industry standards, noting a Solidigm-led draft specification is under SNIA review.
Key Takeaways
- SSDs are the last air-cooled holdout in otherwise liquid-cooled GPU servers — PCIe Gen 5 at ~25W already struggles with ASHRAE A2 limits; PCIe Gen 7 at 60-70W will make air cooling infeasible.
- SSD thermal throttling has direct GPU utilization consequences: at 77°C SSDs drop to ~58% performance; at 79°C to under 1%, directly impacting tokens/second and customer economics.
- MLCommons MLPerf Storage 2.0 data across 150 submissions shows a strong linear relationship between storage read bandwidth and GPU utilization — there is no shortcut to GPU efficiency without faster storage.
- Liquid cooling SSDs introduces three unique challenges not present for CPUs/GPUs: hot-swap serviceability, dual-sided PCB heat sources, and specifications written exclusively for air-cooled environments.
- Solidigm's spring-activated cold plate integration solves the hot-swap problem — cold plate auto-engages on insertion, disengages on removal — and their single cold plate design cools both PCB sides simultaneously.
- Liquid cooling provides approximately 100x better cooling energy efficiency for SSDs versus air: 1000x lower flow rate needed, only 10x higher impedance, net 100x energy advantage.
- Iceotope's precision immersion system (small circulating mineral oil) supports 280 kW usable compute per rack standalone and over 950 kW combined with cold plates, versus ~120 kW for cold plate + rear door heat exchanger hybrids.
- Iceotope's Meta JBOD partnership (5+ years ago) demonstrated 30% drive failure reduction from eliminating fan vibration, and customers report zero performance loss during full HVAC failure events.
- Solidigm has submitted a draft specification through SNIA (OCP Alliance partner) addressing surface flatness tolerances, cold plate attachment surfaces, and angled insertion edges for hot-swap compatibility — currently under member review.
Sign in to read the full article.
Sign In