Phison Electronics Corp. · TPEX:8299

Phison Electronics — NVIDIA GTC 2025

32 slides from Phison Electronics's conference presentation, Mar 15, 2025

  1. Slide 1: Affordable LLM Training & Inference

    Affordable LLM Training & Inference CUSTOMER PRESENTATION GTC 2025

  2. Slide 2: Executive Summary

    Executive Summary Problem Target Phison’s Statement Market Solution Enables… • High cost of entry and scaling • SMB – Small budgets & infrastructure • Up to 405B parameter models on-prem • Lack of control and data privacy • Regulated, privacy conscious • IoT to PCs to Workstations to Servers • LLM t

  3. Slide 3: What Phison is Announcing for aiDAPTIV+ at GTC 2025

    What Phison is Announcing for aiDAPTIV+ at GTC 2025 Demo of World’s 1st Support for NVIDIA LLM Training Laptop Jetson IoT Devices Maingear laptop PC built for AI Edge computing and robotics powered by Phison aiDAPTIV+ powered by Phison aiDAPTIV+ • Full Model Fine-Tuning • LoRA Model Fine-Tuning • En

  4. Slide 4: Quick Facts About Phison

    Quick Facts About Phison Experience 23+ $1.5B 80%+ 2023 Revenue R&D Expense / OPEX Years 4000+ Business Focus 18%+ ➢ Enterprise (SSD, Signal IC) HC Global SSD Market Share ➢ AI ➢ Industrial 75% R&D ➢ Client 1. Phison Technical Marketing Apr 2024 Incl. Enterprise and Client CUSTOMER PRESENTATION GTC

  5. Slide 5: Phison Trusted by Highly Respected Companies

    Phison Trusted by Highly Respected Companies Enterprise NAND Wafer/SSD Controller Space Missions Require: Customers Partners Zero Failures Partnered with Phison to supply their Mars Perseverance Enterprise SSD brand Rover Depends on Phison’s SSD 2 Missions to the ISS and Counting New Missions to Lun

  6. Slide 6: Corporations Need AI

    Corporations Need AI AI isn’t just for individuals; more companies are using it to improve processes, marketing predictions, and business operations… Market Product Sales Smart Evaluation Creation Prediction Manufacturing CUSTOMER PRESENTATION GTC 2025

  7. Slide 7: Today's LLM Market Segment

    Today’s LLM Market Segment GPU Cards Requirements Pre-training Build it Text >1000 Understanding Model Sentence human language Massive GPUs for Creation Article High Computing Power RAG and Fine-tuning On Premise Domain Train it Training Small datasets Question Answering Key 10~100 Information Extra

  8. Slide 8: Technical Constraints for Generative AI On-Premises

    Technic al Constraints f or Generative AI On -Premises Fine-Tuning • Rapid growth in size of models ◦ Insufficient memory capacity ◦ Difficulty in scaling • High machine costs slows adoption Inference Generative AI • Insufficient memory for tokens leading to: ◦ Limited context for chat and prompt ◦

  9. Slide 9: Skills Constraint for AI Adoption On-Premises

    Skills Constraint for AI Adoption On-Premises 1 LLM deployments are limited by the lack of skilled LLM developers. 2 These developers lack affordable infrastructure on which to learn. Source: Lack of LLM Developers Impacting AI Ecosystem itprotoday.com CUSTOMER PRESENTATION GTC 2025

  10. Slide 10: Challenges for Organizations Who Want to Train LLMs

    Challenges for Organizations Who Want to Train LLMs On-Prem Cost of Entry and Scaling is Too High Domain Doesn't Fit Training in How I Work the Cloud Due to Power Violates Demands, Control and Ease of Use Privacy & Skills CUSTOMER PRESENTATION GTC 2025

  11. Slide 11: Phison aiDAPTIV+ Solves this Dilemma

    Phison aiDAPTIV+ Solves this Dilemma Makes AI Affordable Fits Home, Office or Provides Full Control by 8X Reduction in Classroom for and Privacy of Your Cost and Power Domain Training Data & IP CUSTOMER PRESENTATION GTC 2025

  12. Slide 12: How aiDAPTIV+ Does It

    How aiDAPTIV+ Does It Builds a Larger Memory Pool for LLM Training By Tiering Expensive VRAM and Affordable Flash Memory Traditional Approach Phison Approach AI Training Architecture AI Training Architecture GPU CPU GPU CPU HBM DRAM HBM DRAM or GDDR SSD or GDDR SSD Extends GPU Memory CUSTOMER PRESEN

  13. Slide 13: Phison aiDAPTIV+ LLM Domain Training Solution

    Phison aiDAPTIV+ LLM Domain Training Solution AI Application PyTorch AI Framework Middleware Library GPU Specialized SSD Middleware AI-Series SSD Family Pro Suite UI Coordinates the swapping Seamless Integration End-to-End User Experience between HBM/DRAM and with VRAM/DRAM Spanning Training to Infe

  14. Slide 14: aiDAPTIVLink Memory Management Middleware

    aiDAPTIVLink Memory Management Middleware Expands Capac ity to Affordably Fine-Tu ne Growing Mode ls Data coordinator AI Application Memory management Middleware PyTorch AI Framework middleware offloads fine-tuning data to aiDAPTIVCache SSD: Middleware Library • Expands model size capacity GPU SSD G

  15. Slide 15: High-Endurance Specialized SSD for LLM Training

    High-Endurance Specialized SSD for LLM Training 24/7 Model Training Demands 4x 6000 ADA 4x W7900 Pro 8x 6000 ADA 8x W7900 Pro GPU (Workstation) (Workstation) (Server) Pro (Server) 24/7 Training 15.71 16.96 31.42 33.92 DWPD* Enterprise 8x H100 8x MI300 8x Gaudi2 8x Gaudi3 GPU (Server) (Server) (Serve

  16. Slide 16: Ease of Use with aiDAPTIV+ Pro Suite

    Ease of Use with aiDAPTIV+ Pro Suite Optimized End-to-End User Experience Spanning Data Ingest to Inference Pro Suite UI Data Fine RAG Monitor Validate Inference Ingest Tune CUSTOMER PRESENTATION GTC 2025

  17. Slide 17: Workstation Scaling

    Workstation Scaling M ak es Po s si bl e F i ne - T un i ng B ey o n d VR AM C a paci t y Single node 4x GPU configuration comparing GPU only and GPU with 5 with 4.5 4.41 GPU Card Only 4 3.5 3 2.5 2 1.5 FAIL FAIL 1 0.74 0.5 0.42 0.42 0 7B 7B Parameters 13B 13B Parameters 70B 70B Parameters Fixed Tra

  18. Slide 18: Server Scaling

    Server Scaling M ak es Po s si bl e F i ne - T un i ng B ey o n d VR AM C a paci t y Single node 8x GPU configuration comparing GPU and GPU with 12 with 10.73 GPU Card Only TRAINING TIME (HOURS) 10 8 6 4 2.47 FAIL FAIL 2 0.40 0.42 0 Llama-2 Llama-2 13b 13B Llama-2 70b 70B 180b Falcon 180B Model Size

  19. Slide 19: Improves Inference Recall Performance with aiDAPTIV+

    Improves Inference Recall Performance with Jetson only 2.5 with Reduced TTFT by 14X 2 TTFT(s) 1.5 1 Extend token length by 8X 0.5 0 8 16 32 64 128 256 512 1K 2K 4K 8K Token Length Note: 1. Estimated figures are based on Llama 3.1-8B 2. System : NVIDIA Jetson Orin Nano Super 8GB NVIDIA Jetson Soluti

  20. Slide 20: "Teaching" PC for LLM Domain Training

    "Teaching" PC for LLM Domain Training AI Teaching PC Universities, Researchers, Students, Developers and Enthusiasts • Universities need available platforms for teaching access “Learn how to train an LLM • Researchers and Software Developers need to keep up with the latest AI technology beyond just

  21. Slide 21: Train LLMs at Any Budget with aiDAPTIV+

    Train LLMs at Any Budget with IoT Device Laptop PC Desktop PC Workstation PC Server Up to 64B Parameter Up to 8B Parameter Up to 13B Parameter Up to 100B Parameter Up to 405B Parameter LoRA Model Training Full Model Training Full Model Training Full Model Training Full Model Training $500-1,000 $2,0

  22. Slide 22: The Affordable Path to Trillion Parameter Training

    The Affordable Path to 1 Trillion Parameter Training The largest size model training at 1/30 (<4%) of the cost $3M 1 Trillion 671 Billion 2U AI Training Appliance 405 Billion • NVIDIA GH200 Superchip x 1 $100K 69 Billion • 480GB DRAM Llama Llama DeepSeek • Phison aiDAPTIV+ MW/SW & TBD 2 3.1 R3 aiDAP

  23. Slide 23: Two Ways to Fine-Tune AI Models

    Two Ways to Fine-Tune AI models On-Prem/Edge Cloud-Based Challenges Challenges Current All GPU/VRAM AI Equipment Data Privacy Too Expensive Unpredictable/Unlimited Costs CUSTOMER PRESENTATION GTC 2025

  24. Slide 24: Phison aiDAPTIV+ Delivers Affordable Memory Scaling

    Phison aiDAPTIV+ Delivers Affordable Memory Scaling Current Phison’s aiDAPTIV+ AI Algorithm Architecture AI Algorithm Architecture GPU + HBM/GDDR Scaling Limits Expands Memory Scaling with NAND Flash Lots of GPUs to Run Mainstream LLM 8-10x Cost Savings AI Application PyTorch Extend GPU RAM AI Frame

  25. Slide 25: LLM Domain Training Cost: On-Premises vs Cloud

    LLM Domain Training Cost: On-Premises vs Cloud Node @ Session 5 Year Training Cost Name GPU Duration tk/sec $/min $/Mtk3 (sorted by cost) Cost/hr1 (hour) Cost 15/month 30/month 120/month Phison: aiDAPTIV+ v1.03 4x 6000ada 1 4.34 $6.942 640 $0.03 $0.69 $50,000 Generic: GPU Only 4x 6000ada 8 0.54 $6.9

  26. Slide 26: Thank You!

    Thank You! CUSTOMER PRESENTATION GTC 2025

  27. Slide 27: Enterprise SSD

    GTC 2025 Enterprise SSD CONFIDENTIAL

  28. Slide 28: Phison Enterprise SSD Product Line-up

    Phison Enterprise SSD Product Line-up High Performance Data Center SATA Boot Drive Artificial Intelligence X-Series D-Series S-Series B-Series AI-Series PCIe 4.0/5.0 PCIe 4.0/5.0 SATA III PCIe 4.0 / SATA III PCIe 4.0 Up to 30.72TB Up to 122.88TB Up to 15.36TB Up to 960GB Up to 8TB 1/3/60 DWPD 0.3/1/

  29. Slide 29: D200V/D205V Series Product Overview

    D200V/D205V Series Product Overview D200V/D205V QLC SSD Phison’s D200V/D205V series utilize QLC technology to meet the high storage density demand driven by AI. With an impressive 122TB per SSD, D205V will help drive the trend of efficient data storage with improved space utilization and reduced pow

  30. Slide 30: X200 Series Product Overview

    X200 Series Product Overview X200 Series SSD The best PCIe Gen5 performance, features, endurance, and economics for enterprise applications. The X200 shows Phison's dedication to developing advanced SSD technology to lead the industry in density, performance, and power efficiency for all mass-capaci

  31. Slide 31: X100 Series Product Overview

    X100 Series Product Overview X100 Series SSD Phison’s X100 series SSD is highly customizable and marks the spot for enterprises demanding faster and smarter global infrastructures. Featuring best-in-class performance, the X100 enables enterprises to reduce the total cost of ownership through higher

  32. Slide 32: Industry's Most Advanced PCIe 5.0 Enterprise SSD

    Industry's Most Advanced PCIe 5.0 Enterprise SSD (7.68TB, 1 DWPD) X200P PM1743 CM7-R PS1010* Sequential (MB/s) 4K Random (KIOPS) 14,800 3,200 8,700 460 Read Write Read Write X200P PM1743 CM7-R PS1010 X200P PM1743 CM7-R PS1010 *PS1010 performance is not based on 7.68TB GTC 2025

Phison Electronics — NVIDIA GTC 2025 on 6ix