- Sandisk enlists high computing minds to form flash-based reminiscence various for AI
- HBF reminiscence backed by RISC and GPU leaders guarantees excessive bandwidth and big capability
- Patterson and Koduri be part of Sandisk to information flash reminiscence growth past HBM limits
Sandisk has appointed two main figures in computing to assist form the path of its high-capacity reminiscence tech for AI workloads.
Professor David Patterson and Raja Koduri have joined Sandisk’s new Technical Advisory Board to supply strategic and technical enter on Excessive Bandwidth Flash (HBF), a flash-based various to Excessive Bandwidth Reminiscence (HBM).
Patterson is credited for co-developing Decreased Instruction Set Computing (RISC) and Redundant Array of Cheap Disks (RAID), and can lead the advisory board. Koduri is understood for his management in graphics structure, having overseen GPU designs at AMD and Intel.
Many years of expertise
Collectively, they convey a long time of expertise throughout computing, reminiscence programs, and large-scale structure.
“We’re honored to have two distinguished laptop structure consultants be part of our Technical Advisory Board,” mentioned Alper Ilkbahar, Government Vice President and Chief Know-how Officer at Sandisk.
“Their collective expertise and strategic counsel will probably be instrumental in shaping HBF as the longer term reminiscence customary for the AI trade, and affirming we not solely meet however exceed the expectations of our prospects and companions.”
Patterson mentioned, “HBF exhibits the promise of taking part in an vital position in datacenter AI by delivering unprecedented reminiscence capability at excessive bandwidth, enabling inference workloads to scale far past at present’s constraints. It may drive down prices of latest AI functions which can be at the moment unaffordable.”
Signal as much as the TechRadar Professional e-newsletter to get all the highest information, opinion, options and steerage what you are promoting must succeed!
Koduri added, “HBF is about to revolutionize edge AI by equipping gadgets with reminiscence capability and bandwidth capabilities that may assist subtle fashions operating regionally in actual time. This development will unlock a brand new period of clever edge functions, essentially altering how and the place AI inference is carried out.”
HBF is designed to match the bandwidth of HBM whereas providing as much as 8 instances the capability at an analogous price.
Constructed with BiCS flash, CBA wafer bonding, and proprietary stacking that enables 16 dies per bundle, HBF affords a brand new technique to broaden GPU reminiscence with out relying completely on costly DRAM.
Whereas not a direct substitute for HBM, HBF shares the identical electrical interface and solely requires minimal protocol modifications.
Sandisk beforehand demonstrated how an AI GPU utilizing solely HBM would possibly assist 192GB of reminiscence, however by combining it with HBF, that determine may attain 3TB.
In a configuration utilizing solely HBF, reminiscence capability may scale as much as 4TB.
The expertise was first revealed at Sandisk’s Future FWD 2025 investor occasion again in February 2025, alongside its roadmap for future HBF generations.
These updates present will increase in capability and bandwidth over time, with some tradeoffs in vitality effectivity.
By forming an advisory board and in search of open customary growth, Sandisk is making an attempt to keep away from locking the market into proprietary options.
This will assist it achieve traction towards rivals like Samsung and SK Hynix that are each closely invested within the HBM house.
You may additionally like
- Sandisk may use structure referred to as Stargate to energy its 256TB and 512TB SSDs
- Samsung HBM roadmap exhibits Google will probably be Nvidia's fiercest competitor in AI
- Sandisk quietly places 2TB SD card on sale – Excessive Professional is as quick as a tough drive