Faster and denser GDDR7 memory from Sk Hynix
SK hynix is preparing to showcase 48 Gb/s 24 Gb GDDR7 DRAM at ISSCC 2026, targeting mid-range inference AI workloads. The conference program lists a paper titled “A 48Gb/s 24Gb GDDR7 DRAM for Mid-Range Inference AI with Symmetric 2CH-Mode Operation, Clock-Path Optimization, and RAS Features,” which specifies both the speed and capacity goals for this new chip. These modules would deliver more than 70% higher per-pin bandwidth than the 28 Gbps GDDR7 used on NVIDIA’s RTX 5090 cards, while also moving from 2GB to 3GB per package.
The 3GB memory modules have only started to appear in first products. Officially, RTX PRO 6000 and RTX 5090 Laptop among the only products using this type of memory. It allows GPU vendors to utilize larger capacity on narrower buses.
The 48 Gb/s target pushes past previous high-speed GDDR7 announcements, such as Samsung’s 24 Gb chips rated up to 42.5 Gbps (discussed at the same event last year) and Micron’s 32 Gb/s parts now sampling for gaming and AI hardware. On paper, SK hynix is now setting the highest speed goal among the three major GDDR7 suppliers.
Source: ISSCC
At 24 Gb density, each module carries 3 GB of memory. Eight chips would yield 24GB, twelve would give 36GB, and sixteen would reach 48GB on a single GPU. With a 32-bit interface per chip, one device at 48 Gb/s translates to 192 GB/s of bandwidth; a 256-bit GPU memory bus built from these parts could reach about 1.5 TB/s, while a 512-bit configuration would approach 3 TB/s.
In reality, GPU vendors rarely use the highest speeds for its products, usually to preserve good yields the memory speed is lowered to meet the spec requirement. Only XX80 class GPUs from NVIDIA shipped with faster memory at launch, but it was never the full speed for GDDR6X nor GDDR7.
As we mentioned, Samsung’s 42.5 Gbps paper was presented last year and there was no sign of such memory coming to the market. In other words, this is a discussion on an R&D paper rather than a product launch. ISSCC presentations often precede real hardware by a year or more, so such chips may arrive first in datacenter or AI boards before they appear in gaming GPUs, if at all.
Source: ISSCC
