As organizations grapple with skyrocketing AI training costs and GPU shortages, Phison is bringing a compelling value proposition to SC25 with a clear message: do more AI with less GPU. Michael Wu, GM and President of Phison US, explores how the company’s 25 years of NAND flash innovation is addressing one of the industry’s most pressing challenges—extending effective GPU memory using high-capacity flash storage to enable larger model training and faster inference without adding expensive GPUs.
At Booth 4532 in St. Louis this November, attendees will see live demonstrations of Phison’s aiDAPTIV+ platform, which offloads memory pressure from GPUs onto flash storage, and the company’s Pascari Gen5 enterprise SSDs delivering up to 14.8 GB/s read speeds and 3.2 million IOPS. With new portfolio additions launching at the show and partnerships featuring AI systems from Intel, NVIDIA, Newegg ABS, and MAINGEAR, Phison is positioning itself as a practical solution for budget-conscious HPC organizations looking to achieve faster time to value, predictable costs, and improved price-performance in their on-premises AI deployments.
VMblog: If you were giving an SC25 attendee a quick overview of your company, what would you say? How would you describe your company’s evolution since SC24?
Michael Wu: Phison has led NAND flash storage innovation for 25 years. We’ve evolved from ODM-focused controllers to a full portfolio: enterprise SSDs under our Pascari brand, client and industrial SSDs, and aiDAPTIV+, our on-premises AI training and inference solution. Since SC24, we expanded high-capacity Pascari Gen5 enterprise drives, introduced new computational storage capabilities, and grew aiDAPTIV+ to help organizations run larger models on smaller systems.
VMblog: What key message will SC25 attendees hear from you this year? What concrete value proposition will they take back to convince their C-suite and budget holders in today’s tighter economic climate?
Wu: Do more AI with less GPU. Phison’s aiDAPTIV+ lowers total cost by offloading memory pressure from GPUs onto high-capacity flash, improving price-performance and making on-premises AI practical. The result: faster time to value, predictable costs, and supply-chain stability from a partner with a 25-year track record and ultra-fast enterprise grade Pascari SSDs. New offerings also debuting at SC25 for the Pascari enterprise portfolio, stay tuned.
VMblog: Given the current challenges facing HPC organizations-from budget constraints to talent shortages-what specific problems does your technology solve for SC25 attendees?
Wu:
- Budget:Train larger AI models and accelerate inference without adding more expensive GPUs. aiDAPTIV+ extends effective GPU memory using flash.
- Utilization:Keep GPUs busy with high-throughput data staging and caching.
- Complexity:IMAGIN+ customization aligns firmware, endurance, and form factors to each workload.
- Supply risk:Multi-source NAND and joint ventures help ensure continuity.
VMblog: With so many vendors claiming AI-HPC leadership, what are your true differentiators?
Wu:
- Depth in flash:25 years of NAND leadership, more than 2,000 patents, and first-to-market milestones.
- Unique offering: Industry’s only AI solution which speeds inference and enables larger model training without adding expensive GPUs.
- End-to-end:Controllers, enterprise SSDs (Pascari), aiDAPTIV+ all-in-one AI toolset for training, fine-tuning, RAG, and inference.
- Customization:IMAGIN+ embeds partner requirements from firmware to endurance under your control.
- Proven scale:Strategic collaborations with leading NAND vendors, OEMs, and integrators.
VMblog: How can attendees find you at SC25? What hands-on demonstrations, live benchmarks, or interactive experiences will you have at your booth this year?
Wu: Visit Phison at Booth 4532 in the Quantum Junction area of Hall 5.
See:
- aiDAPTIV+ :Faster inference response and training of bigger AI models on an on-premises node showing GPU memory extension via flash.
- Pascari Gen5 Enterprise SSDs:Up to 14.8 GB/s read, 8.7 GB/s write, and 3.2 million IOPS in common AI/analytics I/O patterns.
- Launch Previews: New high-performance Pascari portfolio additions
- AI Systems: Intel integrated GPU-based AI laptop PC, NVIDIA discrete GPU-based desktop AI PC from Newegg ABS and a concept demo of an NVIDIA discrete GPU-based AI laptop PC from MAINGEAR.
VMblog: For returning sponsors, what keeps bringing you back to SC year after year?
Wu: SC is where AI and HPC practitioners validate what works. It’s the right forum to show significant data capacity increases and performance gains to solve the biggest computing challenges and collaborate with partners and customers on real workloads.
VMblog: As AI training costs have skyrocketed, how are your solutions helping customers achieve better price-performance while maintaining competitive AI capabilities?
Wu: aiDAPTIV+ shifts part of the model state and datasets to scalable flash memory, reducing dependence on high-cost GPU VRAM. Combined with aiDAPTIVCache high endurance flash memory SSDs, teams can shorten inference response times and fine-tune larger AI models within existing budgets for on-premises deployments while being easy-to-use and ensuring privacy.
VMblog: Power efficiency has become critical. What breakthrough technologies are you showcasing to deliver superior performance per watt, and what real-world gains can customers expect?
Wu:
- Offload Power-Hungry GPU cards: aiDAPTIV+ reduces the need for more GPU cards to get more VRAM. This shrinks the number of AI servers and enables AI processing on more power-efficient desktop computers.
- High-density Enterprise SSDs:Up to 122.88 TB per SSD reduces server count and idle power.
- Controller efficiency:Phison controllers sustain high throughput at lower active power per gigabyte moved.
- Workload fit:Matching endurance and form factor to I/O patterns cuts overprovisioning.
VMblog: The HPC-AI convergence has accelerated in 2025. How are your solutions enabling seamless workflows across simulation, generative AI training, and inference on unified infrastructure?
Wu: Pascari enterprise SSDs provide consistent performance across mixed I/O profiles from checkpointing and scratch for simulation to streaming batches for training and low-latency reads for inference. aiDAPTIV+ adds a software layer for caching, dataset staging, and memory extension, so teams keep one flash-centric pipeline across the lifecycle.
VMblog: With the emergence of foundation models for scientific computing and multi-modal AI, how is your product strategy evolving?
Wu: aiDAPTIV+ adds no-code tools for on-premises fine-tuning and RAG so domain teams can iterate on scientific and multi-modal models without cloud lock-in. We’re investing in controllers and firmware tuned for large sequential and small random blends, higher queue depths, and fast metadata ops.
VMblog: As sovereign AI and data residency requirements reshape deployment strategies, how are you helping organizations navigate on-premises, cloud, edge, and hybrid quantum-classical architectures?
Wu: aiDAPTIV+ runs on premises and at the edge, keeping sensitive data local. Pascari’s broad form factor support (U.2, E1.S, E3.S, 2.5-inch SATA) and endurance options let architects build consistent storage tiers across sites and providers.
VMblog: Memory bandwidth and storage latency remain bottlenecks. What innovations will you unveil at SC25?
Wu:
- aiDAPTIVCache:Flash-based caching and staging to feed GPUs efficiently.
- High-IOPS Gen5 Pascari SSDs:Up to 14.8 GB/s read and 3.2M IOPS to accelerate batch delivery, checkpoints, and feature stores.
- SCM option:Pascari X200Z in pSLC mode, up to 60 DWPD, for log- and write-heavy phases.
VMblog: With HPC democratization efforts expanding, what new solutions or programs are you introducing to make advanced computing accessible to mid-market organizations, developing nations, and specialized research communities?
Wu: aiDAPTIV+ enables budget-friendly training and inference on modest GPU footprints. We pair that with partner bundles through OEMs and integrators such as Newegg ABS and MAINGEAR AI desktop PCs and workstations to help schools, labs, and SMBs can deploy quickly with predictable costs.
VMblog: As cyber threats intensify, what comprehensive security frameworks and zero-trust approaches are you implementing without compromising performance?
Wu: Pascari enterprise SSDs support built-in encryption and secure erase, with firmware signed and validated through our supply chain. IMAGIN+ lets customers align policies such as secure boot and crypto erase to their compliance needs while maintaining throughput.
VMblog: What prizes, exclusive demos, or special giveaways will you have at your SC25 booth to engage attendees?
Wu: Demos with the Intel integrated GPU-based laptop running accelerated inference applications. Swing by our booth for giveaways.
VMblog: Are you hosting any networking events, customer appreciation gatherings, or technical deep-dive sessions during SC25?
Wu: Yes, we are hosting customer and partner discussions in Meeting Room# ES 26.
VMblog: How does Phison’s IMAGIN+ service reduce deployment risk for AI-HPC projects?
Wu: We co-design endurance, firmware, and form factor around your datasets and I/O patterns, then validate with your integrator before install. That shortens production time and avoids costly re-spins.
