Sandisk wants to replace HBM with new HBF memory – Raja Koduri enters as a technical counselor

Raja Koduri is now in a new mission: to achieve greater vara capacity in AI GPUs. The announcement was made by him at the X -account, former Twitter, and Sandisk, who formed a technical advisory board for the project. The goal is to eliminate bottlenecks faced by HBM memories.

The storage device manufacturer Now it will also focus on the development of HBF (High Bandwidth Flash) memory. The plan is to provide a high capacity AI -capacity memory solution.

Raja’s collaboration with Sandisk is a significant step due to his experience in GPUS development. Particularly, your contribution is fundamental for the experience in designing computer architectures. She aligns herself perfectly with Sandisk’s mission to produce HBF to achieve greater capacity to.

Related News:

HBM X HBF

Although HBM has advanced quickly, providing high memory capabilities for AI superchips, HBF can dramatically increase memory capacity by taking advantage of Through-Technology Via Via.

A single HBF battery can reach Terabytes of memory capacity. And the integration of eight of these batteries into a system can help IA GPUs to reach up to 4 tb of Vram. And this keeping the high bandwidth that HBM offers.

The result is the overcoming of HBM boundaries, which cannot keep up with the intensive AI demands and needs much more time to move forward.

Raja’s role

It is important to note that Sandisk HBF will not be directly competing with DRAM in critical latency workloads. AI operations, such as AI inference and large -scale models development, require more memory capacity to provide rapid results.

Greater capacity and bandwidth exceed the gross latency requirements in such operations. And Raja’s role will be strategic in developing a high capacity HBF.

After leaving Intel in 2023, he joined Tenetorrent. It is a company that seeks to develop high performance AI and CPU RISC-V solutions capable of challenging Intel itself.

As Sandisk is striving to establish HBF as an open standard ecosystem, This will encourage a wide adoption by the industry. Raja’s vast network and experience in building ecosystems will be excellent for this task. And this will promote even more partnerships with GPU suppliers.

Below is an approximate translation of the executive message on X:

Enthusiastic about the partnership with Sandisk.

When we started the development of HBM, our focus was to improve bandwidth/watt width and bandwidth/mm² (both important restrictions for mobile devices), maintaining competitive capacity with existing solutions. With HBF, the focus is on significantly increasing memory capacity (per dollar, by watt and mm²), while we offer competitive bandwidth.

Calculation (flops) * Memory capacity (bytes) * Band width (bytes/sec) modulates the performance of AI models for training and inference.

Fonte: Sandisk.