The first iteration of high-bandwidth memory (HBM) was limited to a maximum speed of 128GB/s per stack. But there's one major caveat: Due to physical limitations, graphics cards using the HBM1 are capped at 4GB of memory.
Over time, HBM manufacturers such as SKHynix and Samsung improved HBM's shortcomings. The first update, HBM2, doubled potential speeds to 256GB/s per stack, bringing the maximum capacity to 8GB. In 2018, HBM2 was slightly updated (HBM2E), the capacity limit was further increased to 24GB, and the speed was also increased again, eventually reaching a peak of 460GB/s per chip.
After the launch of HBM3, the speed doubled again, reaching a maximum speed of 819GB/s per stack. Even more impressively, capacity has nearly tripled, from 24GB to 64GB. Like HBM2E, HBM3 is getting another mid-cycle upgrade, HBM3E, which increases theoretical speeds to 1.2TB/s per stack.
Over time, HBM in consumer graphics cards was gradually replaced by more competitively priced GDDR memory. High-bandwidth memory is becoming standard in data centers, and workplace graphics card manufacturers are choosing to use faster interfaces.
Amid various updates and improvements, HBM retains the same 1024-bit (per stack) interface in all iterations. According to a report from South Korea, this may finally change when HBM4 hits the market. If reports are true, the memory interface will double from 1024 bits to 2048 bits.
In theory, jumping to a 2048-bit interface could double transfer speeds again. Unfortunately, memory manufacturers may not be able to maintain the same transfer rates with HBM4 compared to HBM3E. However, higher memory interfaces allow manufacturers to use less stacking in graphics cards.
For example, NVIDIA's flagship AI graphics card H100 currently uses six 1024-bit known-good stacked dies, enabling a 6144-bit interface. If the memory interface were doubled to 2048 bits, NVIDIA could theoretically halve the number of chips to three and get the same performance. Of course, it's unclear which approach manufacturers will take, as HBM4 will almost certainly not enter production for several years.
Currently, both SK Hynix and Samsung believe that when they start producing HBM4, they will be able to achieve "100% yield." Only time will tell if these reports hold up, so treat this news with caution.
access:
Jingdong Mall