据推特爆料,英伟达在GIDS方案中考虑使用高可靠性、高速度的NAND(如200M IOPS),而非HBF。HBF因写入寿命短不适合KV缓存等重写任务,可能仅用于存储权重。若采用,架构或为8 HBM堆栈加4 HBF堆栈。
Yup, using HBF for GIDS doesn't even make sense
HBF won't be used for any write-heavy tasks like KV cache storage (as it will kill the NAND in 2 years)
You can use it for storing weights
If it's ever adopted, the structure will be something like 8 HBM stacks + 4 HBF stacks
"NVIDIA is looking to use high-reliability, high-speed NAND, such as 200M IOPS NAND, for GIDS. It is not considering HBF."
likes: 90 | retweets: 8 | replies: 10 | views: 25959