AI accelerators will get terabytes of memory: SanDisk plans to replace HBM with ultra-fast HBF flash memory

SanDisk will soon become independent from Western Digital, and it has big plans for the future. In particular, the company intends to develop HBF (High Bandwidth Flash) technology, which will provide AI accelerators with significantly more memory than the HBM (High Bandwidth Memory) used today.

Image source: SanDisk

The idea may seem absurd, because NAND flash memory is traditionally much slower than DRAM, which underlies HBM. HBF is designed to solve this problem and become the optimal solution for tasks related to running already trained models (inference), rather than training neural networks. With the change of HBM generations, the available memory capacity also increases – today, accelerators from AMD and Nvidia offer 192 GB. SanDisk assures that HBF will help increase this figure by eight or even 16 times at a comparable price.

The company offers two scenarios compared to the traditional configuration with eight HBM stacks for a total of 192 GB. The first option is six HBF chips and two HBM. In this case, the total memory capacity increases to 3120 GB, or about 3 TB. The second scenario considers completely replacing HBM with HBF, which will give the accelerator 4096 GB (4 TB) of memory. This volume can completely fit a large Frontier language model with 1.8 trillion parameters and a size of 3.6 TB. In the examples given, the capacity of one HBM stack is taken as 24 GB, and HBF – 512 GB (more than 21 times higher), because NAND flash memory has a higher density.

In the HBF design, NAND flash dies are stacked on top of each other and placed above the logic die. The HBF stack, like HBM, sits on an interposer next to the GPU, CPU, or TPU, depending on where the memory is needed; the interface will require only “minor protocol changes.” The key is that HBF offers the same throughput as HBM, according to SanDisk.

In the traditional version, NAND really approaches DRAM in terms of bandwidth, but desperately loses in terms of access time. SanDisk solved this problem with a special NAND architecture, dividing the component into several areas with a greater number of data lines, which means faster access and higher performance.

The HBF architecture was developed by the company last year under the “influence of major players in the AI ​​field.” Next, a technical advisory board will be formed, which will include the company’s partners and industry leaders; then an open standard will appear. In 2019, Japanese scientists proposed the HBN (High Bandwidth NAND) concept, and it is possible that HBF has something in common with it. Meanwhile, SanDisk has already prepared a “roadmap” describing new generations of this type of memory — both capacity and performance will increase.

admin

Share
Published by
admin

Recent Posts

Zuckerberg Announces AI That Will Revolutionize Digital Advertising Industry — It Will Be a “Results Machine”

Mark Zuckerberg said that Meta✴ is developing an AI system capable of automating the entire…

2 minutes ago

Linux kernel to lose support for Intel 486 and early 586 processors

The Linux kernel 6.15 is in the works for a stable release. A preliminary version…

2 minutes ago

Modders Launch Classic Doom Inside Zelda 64: Recompiled, Unofficial PC Port of The Legend of Zelda: Majora’s Mask

Image Source: Gamesradar By pressing the L and R buttons, a table appears in front…

2 minutes ago

Amazon Zoox Recalls Robotaxi Software After Las Vegas Crash

Amazon-owned Zoox has announced a software recall for all 270 of its vehicles after an…

2 minutes ago

Samsung has banned the creation of more than a hundred alarm clocks on its smartphones

Smartphones have replaced many previously familiar separate devices, including alarm clocks. Smart gadgets allow you…

2 minutes ago