HBF, which was drawing attention as a next-generation memory, is expected to remain a complementary role to HBM rather than replacing it, due to high latency and heat generation issues.
Want the next one from this channel too?
HBF (Hybrid Bonding Fabric) is a memory technology that stacks core dies in 8 or 16 layers using SSD (Solid State Drive) manufacturing methods. This is a differentiated approach from existing memory structures, representing a new attempt to implement high-performance and high-capacity memory. HBF aims to maximize data processing efficiency through this stacking technology.
HBF aims for a bandwidth of 3TB (terabytes) and a capacity of 512GB (gigabytes) in terms of performance. These goals demonstrate a commitment to meeting the demands of artificial intelligence (AI) and high-performance computing (HPC) fields through innovative performance improvements compared to existing memory technologies. High bandwidth improves data processing speed, and large capacity allows more data to be processed at once.
HBF boasts overwhelming capacity and bandwidth compared to HBM (High Bandwidth Memory), which is currently the mainstream in high-performance memory. While one HBM has a capacity of approximately 36GB, HBF offers a much larger capacity of 512GB. Furthermore, in terms of bandwidth, it targets 3TB, which is higher than HBM4's target of 2.5TB, demonstrating its potential as a next-generation memory. This can be a significant advantage in AI accelerators and other applications that require rapid processing of large amounts of data.
Despite its high bandwidth and capacity, HBF has a critical limitation: latency. HBF increases bandwidth through parallelization, like sending multiple slow trucks with heavy loads, but it cannot overcome the fundamental delay in data reaching its destination. As a result, HBF is likely to serve as a complementary role rather than a complete replacement for HBM.
HBF's bandwidth characteristics show significant reduction depending on the physical distance from the GPU. For example, when the gap between HBF and the GPU is 2mm, bandwidth close to 3TB can be achieved, but when the gap increases to 2.5cm, the bandwidth drops to only 0.4TB. This means that HBF must be placed as close as possible to the GPU to achieve maximum performance, which can be a major design constraint.
Due to the distance sensitivity of its bandwidth, HBF is unlikely to be placed directly on the interposer next to the GPU. Instead, it is expected to be placed on the substrate underneath the interposer. This is analyzed as a realistic placement solution that accounts for complex stacking structures and heat generation issues, even if it means sacrificing some bandwidth by increasing the distance from the GPU.
Nvidia, a leader in the global AI semiconductor market, showed a lukewarm attitude towards the HBF technology consortium and did not participate. Nvidia is known to prefer the CMX (Compute Express Link) method, which solves problems at the rack level across the entire data center, rather than stacking methods like HBF. This could be a significant variable for the commercialization and spread of HBF technology.
The biggest obstacle to the commercialization of HBF technology is heat generation. HBF itself generates a lot of heat, but it is also very vulnerable to external heat, reacting sensitively to changes in ambient temperature. It is pointed out that the heat generation problem is so severe that if placed too close to the GPU, it could lead to system malfunction rather than just performance degradation. Therefore, the successful introduction of HBF depends on securing heat control technology.
The market introduction of HBF technology is expected to be delayed beyond initial predictions. Last year, mass production was confidently anticipated for this year, but recent reports indicate that actual mass production is now expected around the second half of next year. This suggests that more time will be needed to overcome technical challenges, especially heat generation issues, and for major companies to make adoption decisions.
Analysis suggests that the success of HBF technology hinges on Google's adoption of TPU (Tensor Processing Unit). How actively Google, aiming to grow its TPU against Nvidia, adopts HBF will determine HBF's fate. If Google extensively incorporates HBF into its TPUs, it could be a decisive turning point for the market spread of HBF technology.
Answers come from the transcript, with the exact spot cited.
Want the next one from this channel too?
When 인포마켓 강용운 publishes, we'll write it up like the one you just read and email it to you.
{channel} published 2 in the last 7 days.
We skip Shorts. You can unfollow any time.
