Micron now shipping 11Gbps HBM4 to first customers, future 40 Gbps GDDR7 mentioned

Micron has confirmed that its 12-Hi HBM4 DRAM stacks are now shipping to customers, delivering over 11 Gb/s per-pin speeds and bandwidth above 2.8 TB/s . This is well beyond the JEDEC baseline for HBM4, which defines 8 Gb/s transfer rates and 2 TB/s bandwidth across a 2048-bit interface. The company says the higher-speed design was developed to meet rising demands from accelerator vendors such as NVIDIA and AMD.

According to Micron CEO Sanjay Mehrotra, the company’s HBM4 design outperforms competing products in both performance and power efficiency. Key enablers include Micron’s 1γ DRAM, advanced CMOS base die, and packaging innovations. Customer samples are already in the field, and Micron expects broader platform ramps in 2026.

Source: Micron

Micron also outlined its HBM4E roadmap. Unlike HBM4, which uses a fully in-house base die, HBM4E will involve a partnership with TSMC for manufacturing the logic base die. The HBM4E is set to debut around 2027.

In close collaboration with NVIDIA, Micron has pioneered the adoption of LPDRAM for servers, and since NVIDIA’s launch of LPDRAM in their GB product family, Micron has been the sole supplier of LPDRAM in the data center. In addition to our leadership in HBM and LP5, Micron is also well-positioned with our GDDR7 products, which are designed to deliver ultra-fast performance with pin speeds exceeding 40 Gbps, along with best-in-class power efficiency to address needs of certain future AI systems.

— Sanjay Mehrotra, Micron President & CEO

Micron also reiterated its plans for GDDR7. The company is currently shipping 32 Gb/s parts and is preparing faster variants exceeding 40 Gb/s, which could deliver more than 1.5 TB/s of bandwidth. NVIDIA is already the lead adopter of Micron’s GDDR7 for upcoming GPUs.

Source: Micron

The wording from Micron’s CEO is careful and does not mention any upcoming products directly. As we recall, it was Micron that leaked the NVIDIA RTX 3090 series specs . The RTX 50 SUPER series is not expected to feature faster memory, but it will rely on 3 GB modules. No one should expect 40 Gbps in next-gen GPUs either, as actual cards usually ship with lower frequencies for stability or to maintain higher power efficiency.

Source: Micron