At yesterday's NVIDIA GTC 2024 conference, Samsung unveiled memory chips based on the next-generation GDDR7 memory standard, which will also serve as the primary standard for memory in next-generation gaming GPUs. Samsung announced its development plan as early as 2023. The showcased individual GDDR7 chip boasts a capacity of 16 Gb, translating to 2 GB, with a blazing speed of up to 32 Gbps!
According to reports, while the bandwidth of this GDDR7 chip can reach 32 Gbps (PAM3), it will initially be provided at 28 Gbps to improve overall performance in terms of initial yield and unit cost. Samsung also stated that its GDDR7 memory operates at just 1.1V (lower than the 1.2V standard), resulting in a 20% improvement in efficiency. Other features include reduced heat output during high-speed operations due to updated packaging materials and optimized circuit designs, resulting in a 70% decrease in thermal resistance compared to GDDR6.
The JEDEC standard organization recently released GDDR7 memory specifications, with major highlights including:
· Core-independent LFSR (Linear Feedback Shift Register) training mode with eye diagram masking and error counters to improve training accuracy while reducing training time.
· Doubling the number of independent channels from 2 in GDDR6 to 4 in GDDR7.
· Support for densities ranging from 16 Gbit to 32 Gbit, including support for 2-channel mode to double system capacity.
· Meeting market demands for RAS (Reliability, Availability, Serviceability) by integrating the latest data integrity features, including on-die ECC (ODECC) with real-time reporting, data poisoning, error checking and scrubbing, and command address parity with command blocking.
Although specifications for the higher bandwidth of GDDR7 are still under development, they are not expected to be finalized until 2025 to 2026. Therefore, the initially offered 28 Gbps GDDR7 memory chips from Samsung are considered the preferred choice for NVIDIA's next-generation GeForce RTX 50 "Blackwell" gaming GPU.