HBM2 Doubles Bandwidth, Capacity
May 3, 2016
Author: David Kanter
High Bandwidth Memory 2 (HBM2) has several incremental but critical improvements over the previous-generation technology: it increases the memory capacity and bandwidth, adds ECC support, and improves memory-controller efficiency. These enhancements are essential for high-performance computing (HPC) and workstation tasks and will lead to wider industry adoption. Samsungs HBM2 chips are already in production, and the memory will appear in new graphics cards later this year.
Over the last decade, GDDR has been the high-performance graphics memory of choice, balancing bandwidth and cost for low-capacity systems. GDDR5 is the latest incarnation, but it cant scale much past 8Gbps; in practice, no GPU has achieved more than 320GB/s using GDDR5. The industry is currently developing GDDR5X, which will roughly double the bandwidth, but this memory is unlikely to reduce power consumption at higher data rates.
HBM uses 2.5D or 3D packaging to increase bandwidth and power efficiency by implementing a wide slow interface, but the interposer raises the cost. Since modern computing systems are almost always limited by power consumption, HBM gives architects an effective tool to improve system performance. For example, the first GPU to use HBM was AMDs Fiji, which delivered an incredible 512GB/s while using 3x less power than an equivalent GDDR interface. But HBM primarily serves in consumer graphics because of limitations in the first-generation technologylimitations that are alleviated by HBM2.