Specifications
Manufacturer
SK hynix
Product
GDDR7 SGRAM
Die Density
24Gb (3GB per module)
Pin Speed
48 Gbps
Module Bandwidth
Up to 192 GB/s per GDDR7 module
Bandwidth vs 28Gbps GDDR7
Over 70% improvement
Prior Gen Pin Speed
28 Gbps — 112 GB/s per module
Process Node
1cnm class
Disclosure
IEEE ISSCC 2026, DRAM/SRAM/Non-Volatile Memories session
3GB Module Config
8 x 3GB devices (3GB module capacity)
Target Workload
Mid-range AI inference and graphics
Use Case
AI GPUs for regions moving away from HBM on export grounds
Compatibility
PAM3 signalling, GDDR7 interface
Compression
N/A — standard GDDR7 memory
Supply Class
Announced device — production ramp by vendor schedule
Ecosystem
NVIDIA RTX 50-class and AI inference board designs
Overview
SK hynix's 24Gb GDDR7 die is the 3-gigabyte-per-module entry in the GDDR7 generation. Increasing the die density from 16Gb to 24Gb raises module capacity from 2GB to 3GB while keeping the same package footprint, which is what allows a board to reach 24GB or more of graphics memory without widening the bus or adding device positions.
The headline number is pin speed: 48Gbps, up from the 28Gbps of first-generation GDDR7 parts. That takes a single GDDR7 module from roughly 112GB/s to as much as 192GB/s — over 70% more bandwidth per device. SK hynix positioned the part for mid-range AI inference as well as graphics, which reflects the fact that a growing share of AI accelerator boards, particularly those built for markets where HBM is restricted on export grounds, are turning to GDDR7 as their memory layer.
The device was disclosed at IEEE ISSCC 2026 in the DRAM, SRAM and Non-Volatile Memories session, on a 1cnm-class process. Standard GDDR7 PAM3 signalling is retained, so board designs that already route GDDR7 can adopt the higher-density part without a signalling change.
QS Compute supplies GDDR7 and AI accelerator memory for board-level and system-level integration. Contact us for availability, density options and pricing on 24Gb GDDR7 modules.
Key Benefits
3GB per module raises board memory capacity in the same footprint. 48Gbps pin speed delivers up to 192GB/s per module — over 70% more than 28Gbps GDDR7. Standard GDDR7 PAM3 signalling means no board re-design to adopt it. Targets mid-range AI inference, not just graphics, matching where GDDR7-based accelerators are growing fastest.
Applications
AI inference accelerator boards, graphics and workstation GPUs, RTX-class add-in cards, edge AI compute modules, and video/graphics workloads requiring high memory bandwidth per watt.
Request a Quote — SK HYNIX GDDR7 24GB — 3GB GDDR7 MODULES AT 48GBPS
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →