SanDisk's High-Performance SSDs for AI Applications

Often Asked Questions

What specific products and technologies does SanDisk offer for the data center market?
High-performance eSSD DC SN861 series: The series is based on PCIe Gen5 and TLC NAND. It is designed for high-performance computing applications. This includes AI model training.
High-capacity eSSD DC SN670 series: This series is based on PCIe Gen5 and UltraQLC NAND. It offers up to 128TB capacity. It is suitable for data lakes and AI inference applications.
1PB eSSD platform: Plans to launch a 1PB capacity eSSD to meet the growing capacity demands of AI.
UltraQLC technology: This technology enhances performance, density, and energy efficiency of QLC flash. It achieves this through custom controllers, dedicated hardware accelerators, and advanced system design.

What is SanDisk’s High Bandwidth Flash (HBF) technology? What is its goal?
Match HBM bandwidth: Provides bandwidth equivalent to HBM in workloads like AI inference.
Higher capacity: Offers 8–16 times the capacity of HBM at the same cost.
Simplification for AI applications: Enables larger AI models to run on edge devices, for example, running LLMs on smartphones.
Open standard: SanDisk plans to promote an open HBF standard. They will set up a technical advisory board to foster ecosystem development.
Improved energy efficiency: Future generations of HBF will further improve energy efficiency.

How is SanDisk addressing the “memory wall” issue?

The “memory wall” refers to the widening gap between computing speed and memory capacity/bandwidth.

High Bandwidth Flash (HBF): SanDisk is developing this technology in collaboration with IMEC to tackle the memory wall.

Matrix Memory: A new type of non-volatile memory designed to deliver DRAM-level performance at a lower cost and higher capacity. SanDisk is working with IMEC to develop this technology and has secured a related contract with the U.S. Department of Defense.

What are the key improvements in SanDisk’s next-generation NAND technology (e.g., BiCS8)?
CMOS Bonded Array (CBA) technology: Improves performance, energy efficiency, and density by bonding CMOS circuits directly to the memory array.
Higher storage density: More than 50% increase in storage density compared to the earlier generation.
Higher bandwidth: Programming bandwidth increased by 35%, read bandwidth by 26%.
Faster transfer speeds: Transfer speed improved by more than 80%.
2Tb QLC chip: BiCS8 is used to produce 2Tb QLC chips. These chips are claimed to be the highest-capacity NAND chips presently in production.

What is SanDisk’s outlook on future NAND flash market supply and demand?
SanDisk expects a supply shortage in the NAND flash market by the second half of 2025. This is due to slowed supply growth.
They believe that growth in AI, video, and autonomous driving will drive increasing demand for storage.

How does SanDisk control costs and improve capital efficiency in NAND flash manufacturing?
Joint venture with Kioxia: Leverages large-scale production and advanced technology from the joint venture to reduce costs.
Combination of vertical and horizontal scaling: Utilizes multi-dimensional scaling techniques. These include vertical layer stacking and increased horizontal cell density. Logical scaling and architectural scaling are also used.
Node design improvement: Optimizes node design based on different market applications to balance performance, capacity, and power consumption.
AI-assisted manufacturing: Uses AI technology to enhance manufacturing efficiency and yield.
Innovative testing technologies: Reduces testing costs through innovative testing techniques.
Extending node lifespan: Extends the lifecycle of existing nodes as much as possible.
Advanced packaging technologies: Applies advanced packaging techniques like chip bonding to shorten production cycles and reduce costs.

What are the potential risks and challenges of SanDisk’s High Bandwidth Flash (HBF) technology?
Durability issues: NAND flash has limited write endurance; frequent read/write operations can lead to premature wear and data corruption.
The durability of NAND must be addressed, or it will lead to high error rates and data loss.
Granularity mismatch: NAND flash has a larger read/write granularity (4–16KB), while HBM operates at finer granularity (32 bytes). This can lead to inefficiencies and need modifications to existing software.
Standard ecosystem: An open HBF ecosystem must be established to guarantee compatibility with other devices. If other vendors do not adopt HBF, its application prospects will be limited.
Real-world performance: Whether HBF can truly match HBM in terms of performance still needs to be verified. In particular, NAND flash not match DRAM in latency.

NAND Market Outlook

Overall Trends:
The storage capacity of NAND flash is expected to continue growing. The projected compound annual growth rate (CAGR) across all end markets (excluding enterprise SSDs) is 17%.
SanDisk expects a rebound in smartphone and PC shipments. They also foresee accelerated content growth. These factors will drive the development of the NAND market.

Key Growth Drivers:
AI smartphones and edge inference are expected to drive a CAGR of 20%.
Generative AI and LLM training are projected to contribute to a CAGR of 29%.
Demand for high-performance and high-capacity eSSDs is increasing due to AI inference and AI infrastructure.
Automotive embedded storage (EB) shipments are expected to grow at a CAGR of +35% through 2028.

Supply and Demand Dynamics:
NAND flash supply growth is slowing down. A supply shortage is anticipated by the second half of 2025.
Vendors are now digesting excess inventory, and future demand is expected to outpace industry supply.
Micron and Samsung are both reducing NAND wafer production.

Competitive Landscape:
The “layer race” in NAND flash has come to an end, and industry capital expenditures are decreasing.
SanDisk considers its testing technology a competitive advantage, leveraging AI-assisted innovations in manufacturing and testing.

BiCS8 3D NAND

Key Features:

The latest generation of 3D NAND technology with 218-layer stacking.
Produces 2Tb QLC dies, now the highest-capacity NAND dies in production.
Industry-leading in performance and power efficiency.
Optimized for AI workloads, enabling high-speed data transfer and accelerated data processing.

Technological Innovations and Improvements:

Storage density increased by over 50%.
Layer density improved by more than 12%.
Programming bandwidth improved by over 35%.
Read bandwidth improved by more than 26%.
Data transfer speed improved by over 80%.
Utilizes CMOS directly bonded to array (CBA) technology to enhance cell and I/O performance and shorten manufacturing cycle time.
Compared to competitors, offers lower read latency, higher write efficiency, and faster I/O speed.
Accelerates time-to-market and optimizes total cost of ownership (TCO) for NAND.
Designed for AI workloads, offering high-speed transfer rates.
Ideal for high-performance, high-capacity storage solutions.

Applications and Markets:

Client SSD market, especially suited for QLC SSDs.
Enhanced data center storage solutions, particularly UltraQLC SSDs.
Automotive applications, supporting autonomous driving and advanced driver-assistance systems (ADAS).

Future Development:

Expected to account for 10% of customer shipments by the end of fiscal year 2025.
Projected to reach 40%–50% by the end of fiscal year 2026.
Anticipated to become the mainstream node by late 2026 or early 2027.
Future BiCS technology will exceed 300 layers, supporting the production of 1Tb TLC dies.

UltraQLC

Core Concept
UltraQLC is an innovative technology developed by SanDisk. It aims to achieve the perfect combination of density, performance, and power efficiency. It does so without any compromise.
Its core goal is to disrupt the cloud storage market through BiCS8 UltraQLC™. It also aims to extend compute coverage to PCIe Gen5 and Gen6.
The technology is specifically optimized for AI workloads, delivering high-speed transfer rates and accelerated data processing capabilities.

Technical Details and Innovations
Custom Controller:

Equipped with dedicated hardware accelerators to achieve peak performance.
Dynamically adjusts power based on workload requirements.
Integrates advanced Toggle mode bus multiplexing control.
Supports scaling up to 64 dies/channels.

Advanced System Design:

Optimized SanDisk flash system fully leverages leading NAND nodes and controllers.
Application-specific features including machine learning-based voltage sensing and data retention (DR) recovery.

BiCS8 NAND:

Utilizes BiCS8 NAND’s high-speed transfer rates and accelerated data processing, providing an ideal solution for AI workloads.

UltraQLC™ DC SN670 NVMe™ SSD
Performance Breakthrough:
A PCIe® Gen5 QLC SSD based on BiCS8 NAND technology.

Performance Improvements (Compared to Leading Gen5 128TB QLC SSD):

Sequential read speed: +7%
Random read speed: +68%
Sequential write speed: +27%
Random write speed: +55%

Launch Timeline and Specifications:

Planned release: Q3 2025
Available in two capacities:
- 64TB (61.44TB usable capacity)
- 128TB (122.88TB usable capacity)

Application Scenarios
AI Data Lifecycle:
Suitable for all six phases of the AI data lifecycle, including:

Interface and prompt
Raw data archiving and content storage
Data preparation and ingestion
AI model training
AI inference engine
New content generation

High-Performance Computing & Data Lakes:

Build high-performance computing eSSDs for LLM training.
Build high-capacity storage eSSDs for fast data lakes.

Data Lake Storage Requirements:
Meets demands for high capacity and high density. It requires low TCO and high performance. The storage should deliver consistent IOPS per TB and support read-intensive workloads. Additionally, it should offer high energy efficiency and low idle power consumption.

Technical Features and Advantages Summary

Perfect combination of density, performance, and power efficiency.
Custom controller with dedicated hardware accelerators for high performance and low power.
Advanced system design and SanDisk flash system improvement to enhance overall performance.
BiCS8 NAND provides fast transfer rates and accelerated data processing to support AI workloads.
Scalable to 64 dies/channels, with integrated advanced Toggle mode bus Mux control to improve data transfer efficiency.
Machine learning voltage sensing and data retention recovery technology delivers up to 33% improvement.

Market Positioning and Outlook

Primarily targeted at the cloud market, aiming to disrupt traditional storage solutions.
Especially well-suited for AI workloads, fulfilling storage needs for large-scale, high-performance datasets.
SanDisk anticipates UltraQLC will have a significant impact on the storage sector and drive advancements in storage technologies.
SanDisk is developing a 1PB eSSD drive platform to meet the growing demand for high capacity in AI applications.

3D Matrix Memory

Core Technology and Advantages

Technology Overview:
SanDisk, in collaboration with IMEC, is developing the 3D Matrix Memory technology. It is aimed at breaking through the “storage wall” bottleneck. This bottleneck is between computing power, memory capacity, and bandwidth. This technology is specifically designed for next-generation Strategic Radiation Hardened (SRH) memories. It meets the stringent demands of the Air Force Research Laboratory (AFRL). With the innovative 3D Matrix Memory architecture and advanced materials, SanDisk successfully won the ANGSTRM contract from the U.S. Department of Defense (DoD).

Technology Details:

Core Differentiation:

A new memory cell design that achieves unprecedented performance breakthroughs.

Architecture:

Uses a dense array architecture, significantly improving memory density.

Compatibility:

Fully compatible with existing open industry standards (e.g., CXL), ensuring easy integration.

System Media Management:

Leverages SanDisk’s deep knowledge in system media management to enhance performance.

Performance Goals:

Aiming to achieve DRAM-like performance levels.

Capacity and Cost:

Memory capacity is up to 4 times greater than DRAM, with bit cost reduced by 50%.

Innovation Highlights:
Scalable 3D Architecture:

Introduces an innovative scalable 3D architecture to offer a solid foundation for high-density storage.

Advantages Summary:

Cost Advantage: As the technology progresses, the cost advantage over DRAM will become increasingly significant.
Capacity Advantage: The SRH memory has a capacity of up to 4Gbit. This shows a 16-fold leap compared to the current 256Mbit products.

Application Areas

Air Force Research Laboratory (AFRL): Core supplier for next-generation Strategic Radiation Hardened (SRH) memories.
U.S. Department of Defense (DoD): Successfully awarded the ANGSTRM contract due to superior technological advantages.

Development Roadmap

Isolation device development is progressing smoothly.
Successfully produced Gen1 media samples.
Custom tools developed for IMEC 4-8Gbit CMOS.
Gen1 quality/small batch production system-level validation completed.
Gen1 media capacity successfully scaled up to 32-64Gbit.
Passive array technology progressing to the 100b-1kbit range.
Manufacturing facility transitioned smoothly from 150mm WD research plant to 300mm IMEC facility.

Key Personnel Changes:

Max Mirgoli transitioned to EVP of Global Strategic Partnerships at IMEC.

This translation keeps the original meaning and technical details intact. Let me know if you need any further adjustments!

HIGH BANDWIDTH FLASH（HBF）

1. Technical Architecture and Specifications

Core Concept:
HBF is an innovative storage architecture. It combines the high capacity characteristics of 3D NAND flash with the high bandwidth advantages of High Bandwidth Memory (HBM). It is designed to supply solutions for applications requiring high bandwidth and high capacity, like AI inference.

Technical Details:

Stacking and Interconnect:

Uses Through-Silicon Via (TSV) technology to stack 16 3D NAND BiCS8 chips.
Parallel access to memory sub-arrays is achieved through a logic layer.

Capacity and Sub-arrays:

The capacity of a single stack can reach up to 512GB. It can be expanded up to 4TB VRAM, which consists of 8 stacks.
Breaks traditional NAND designs by enabling independently accessible memory sub-arrays, each with dedicated read/write paths.
Each HBF core chip is a 256Gb 3D NAND device.

Interface and Protocol:

Uses an electrical interface akin to HBM, though protocol modifications are necessary, making it not fully compatible.

Next-Generation NAND Technology:

The 10th generation NAND products use the Toggle DDR6.0 interface standard.
Uses a separate command-location protocol (SCA).
Adopts power-isolated low-tap termination (PI-LTT) technology.

CBA Design:

This design utilizes CMOS Direct Bonded to Array (CBA). It bonds 3D NAND memory arrays onto I/O chips. These chips are made using logic process technology.

Innovations:

Parallel Access: Significantly enhances bandwidth via sub-array parallel access.
SanDisk Proprietary Stacking Technology: Minimizes warping, enabling 16H stacking.
AI-Assisted Manufacturing Testing: Improves manufacturing and testing efficiency.

2. Performance and Application Goals

Applications:

Primarily targets read-intensive AI inference tasks, like LLM inference.
Suitable for AI model storage and inference, but not ideal for latency-sensitive applications (e.g., gaming).
HBF can store about 1.8 trillion factors (16-bit weights) of an LLM, requiring 3600GB of memory capacity.
MoE models with 64 billion factors can be housed in the HBF chip in a smartphone.
BiCS8 NAND technology, with its fast transfer rate and accelerated data processing capabilities, is ideal for AI workloads.

Performance Characteristics:

High capacity, high bandwidth, but latency is higher than DRAM.
Can offer up to 4TB of VRAM for AI GPUs and can be extended to mobile devices.

Comparison with HBM:

HBF aims to match HBM bandwidth while offering 8 to 16 times the capacity at a similar cost.
HBF can’t match DRAM in terms of bit latency.
Has the same electrical interface as HBM, but requires protocol changes.

3. Challenges and Market Evaluation

Key Defects and Challenges:

NAND Durability, Read/Write Granularity, and Software Adaptation Issues:
- NAND durability leads to higher error rates and silent data corruption.
- Read/write granularity is relatively large, and existing software not effectively handle this.
- AI model update frequencies are not an issue for write durability, but read durability still requires consideration.

NAND Durability Issues:

Error Rates and Data Corruption:
- NAND durability issues lead to higher error rates and data corruption, posing significant challenges for AI research.
- Controllers must track bad blocks and erase them, impacting performance.
- Continuous high-frequency read/write operations can quickly degrade performance.
- Temperature sensitivity requires careful consideration of heat dissipation.

Read/Write Granularity Issues:

Granularity Difference:
- NAND flash reads and writes in large blocks, while HBM/DRAM has fine-grained access, creating a significant difference.
Software Adaptation:
- Existing software needs to be adapted to handle this granularity difference.

Application Scenario Limitations:

Latency-Sensitive Applications:
- HBF is not suitable for all applications, particularly latency-sensitive ones. NAND flash access latency is much higher than DRAM, limiting its use in such cases.

Other Considerations:

Electronic Waste:
- NAND degrades quickly and is difficult to repair or replace, leading to electronic waste concerns.
Heat Dissipation and Durability:
- A large number of NAND stacks placed next to GPU chips need more frequent self-refreshing, impacting durability.

4. Development Roadmap and Ecosystem

Development Phases:

Gen 1: First capacity and bandwidth.
Gen 2: Capacity increased by 1.5x, read bandwidth increased by 1.45x.
Gen 3: Both capacity and read bandwidth doubled compared to Gen 1, with energy efficiency improvements in each generation.

Ecosystem:

Aims to make HBF an open standard and form a technical advisory board.
Focused on achieving seamless system integration of HBF.

Market Trends:

Rapid growth in the AI field is driving the demand for high-bandwidth storage.
The NAND market’s supply-demand dynamics are expected to change.

Other Technologies:

3D Matrix Memory:
- SanDisk is also developing 3D Matrix Memory (DRAM) to solve the storage wall issue.

This translation accurately reflects the original meaning and technical specifics of the content. Let me know if you need any further adjustments!

Market Strategy and Operational Analysis

Customer Value Creation Core Strategy
Innovation Drive: Promote breakthroughs in density and performance. Improve energy efficiency through technological innovations like BiCS8 UltraQLC™.
Agile Response: Implement flexible supply chain management to accelerate product time-to-market and adapt to rapidly changing market demands.
Brand Strengthening: Enhance the SanDisk brand’s influence through excellent products and strategic partnerships.

Segmentation Market Penetration Strategy
Client Market:

TLC Technology: Strengthen leadership in performance and mainstream market.
QLC Technology: Innovate in the value market.
Expand the client SSD market share.
Launch new PCIe® Gen 4 and Gen 5 QLC and TLC platform products in 2025.
Consumer Market:
Reinforce brand advantages and expand market through excellent products and GTM channels.
Strengthen leadership in portable SSDs, uSD and SD cards, USBs, etc.
Focus on target customers like gamers, camera enthusiasts, and content creators.
Cloud Market:
Disrupt the storage market with BiCS8 UltraQLC™ technology.
Expand computing coverage through Gen5 and Gen6 technologies.
Offer large-capacity eSSDs for AI data lakes and high-performance computing for caching.
Offer customized storage solutions for all six stages of the AI data cycle.
Mobile and IoT Market:
Focus on discrete solutions and improve the iNAND product portfolio.
Give comprehensive storage solutions for high-growth automotive markets like autonomous driving, ADAS, and eCockpit.
Offer advanced technologies like e.MMC 5.1, UFS 3.1, NVMe™ PCIe® Gen 4, and UFS 4.1.

Financial Strategy. Goals.
Revenue Growth: Capture long-term growth trends in data creation and storage. Leverage emerging opportunities in AI, video, and autonomous driving.
Profit Expansion: Improve cost structure, increase flexibility and agility, and leverage the advantages of scalable operations.
Asset Efficiency: Implement proactive supply management, enhance capital expenditures and inventory, and extend node lifespans.
Financial targets include:

$10 billion in revenue with a non-GAAP operating profit margin of approximately 20%.
Free cash flow exceeding $1.2 billion.
Achieving a net cash position and reducing total debt to below $1 billion within the next four quarters.

Supply Chain and Manufacturing Advantages
Flash Memory Joint Venture with Kioxia:

Strategic advantage of scalable operations. There is 80% wafer fab capacity, with each having a 50% share. This achieves top-tier per-bit cost and capital expenditure efficiency. It also provides strong node transition capabilities.
Memory Development Center:
Over 1,000 employees in Japan, responsible for the operation of the memory development center.

Financial Analysis and Future Outlook

Financial Goals and Key Metrics

Revenue Growth: The long-term target is $10 billion. Growth is driven by the long-term trend in data creation and storage. Accelerators like AI, video, and autonomous driving also contribute to growth.
Profitability: Non-GAAP operating margin is expected to be around 20%, and gross margin is expected to be around 35%.
Capital Expenditure: The total capital expenditure-to-revenue ratio is between 10-20%. The strategy aims to improve capital expenditures and inventory. This approach improves asset efficiency.
Cash Flow: Free cash flow is expected to exceed $1.2 billion.
Balance Sheet: It is expected that a net cash position will be achieved in the next four quarters. The total debt is planned to be reduced to below $1 billion.
Joint Venture Advantage: The flash memory joint venture with Kioxia helps lower costs and improve capital expenditure efficiency.

Financial Transformation and Short-Term Outlook (Q3 2025)

Transformation Quarter: Focus on managing the supply chain to match demand, with attention to cost and normal quarterly cost fluctuations.
Financial Forecast:
- Revenue: $1.55 billion to $1.65 billion.
- Non-GAAP gross margin: 21.5% to 23.0%.
- Non-GAAP operating expenses: $395 million to $405 million.
- Non-GAAP earnings per share: -$0.30 to -$0.45.
Market Trends: Stronger demand in the cloud and client sectors, with the average selling price (ASP) continuing to decline.

Future Development Trends

Growth: Continuous improvement, with accelerated growth in the client market.
Pricing: Improvement is expected starting in the next quarter, as supply interventions are implemented.
Costs: Costs are expected to stay relatively stable in the next quarter and then decline according to the model.
BiCS8 Transformation: By the end of FY2025, BiCS8 is expected to represent 10% of customer sales. By the end of 2026, it will account for 40-50%. It is expected to become the dominant node by the end of 2026 or early 2027.