Kioxia prototypes 5 TB, 64 GB/s flash memory module for large-scale Artificial Intelligence

Kioxia has developed a prototype flash memory module with 5 terabytes capacity and 64 gigabytes per second bandwidth targeted at large-scale Artificial Intelligence models as part of NEDO project JPNP20017.

Kioxia Corporation, described in the article as a world leader in memory solutions, announced a successful prototype of a large-capacity, high-bandwidth flash memory module. The module delivers five terabytes of capacity and a sustained bandwidth of 64 gigabytes per second. This development was carried out under the Post-5G Information and Communication Systems Infrastructure Enhancement R&D Project (JPNP20017) commissioned by the New Energy and Industrial Technology Development Organization, referred to as NEDO.

To overcome limits seen with conventional DRAM-based memory modules, Kioxia developed a new module configuration that uses daisy-chained connections with beads of flash memories. The company also developed high-speed transceiver technology capable of 128 gigabits per second and implemented techniques to enhance flash memory performance. According to the article, these technological advances have been applied to both memory controllers and memory modules in the prototype.

The announcement frames the prototype as addressing a long-standing trade-off between capacity and bandwidth in memory design, especially relevant for large-scale Artificial Intelligence models that demand both large capacity and high data throughput. The article does not specify production timelines, costs, target systems, or compatibility details. It limits its claims to the successful prototyping and the technical features listed, including the 5 TB capacity, 64 GB/s bandwidth, daisy-chained flash configuration, the 128 Gbps transceiver capability, and the application of these innovations to controllers and modules.

74

Impact Score

NVIDIA groq 3 LPX targets low-latency Artificial Intelligence inference

NVIDIA positions Groq 3 LPX as an inference accelerator for Vera Rubin built to handle low-latency, large-context workloads for agentic systems. The platform combines Rubin GPUs and LPUs in a co-designed architecture aimed at boosting throughput, token generation, and efficiency at rack scale.

Nvidia sets the stage for GTC 2026 keynote

Nvidia is preparing to outline its next wave of computing, networking, and rendering plans at GTC 2026, with Jensen Huang leading the keynote. The event is expected to focus on next-generation platforms, broader Artificial Intelligence infrastructure, and the company’s expanding partnership with Intel.

Nvidia chief projects chip sales growth

Nvidia’s chief executive is tied to a projection of massive future Artificial Intelligence chip revenue, but the available source material provides no reported details beyond the headline and a brief author description.

Can world models unlock general purpose robotics

World models aim to help robots learn physics from large-scale video instead of relying mainly on hand-built simulators and scarce robot-specific data. Early results are promising, but major questions remain around consistency, tactile sensing, speed, and economics.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.