The evolution of HBM memory does not stop, and SK Hynix has introduced its HBM3e, capable of transferring 1.18 TB/s per stack. This component is responsible for feeding data to the new NVIDIA Blackwell GPUs, a flow of information that allows these cards to process massive workloads without bottlenecks. A key piece in the artificial intelligence ecosystem. 🚀
A data bus that breaks the bottleneck ⚡
The key to this memory lies in its density and bandwidth. With 24 GB per stack and an effective speed of 9.2 Gbps per pin, the HBM3e achieves that throughput of 1.18 TB/s. To achieve this, SK Hynix has refined the manufacturing process with 1b nm technology and advanced TSV packaging that reduces latency. In practice, this allows the Blackwell to move giant language models without waiting for data to arrive from VRAM.
When the bottleneck goes on a diet 😅
With this bandwidth, the bottleneck in AI training seems to have gone on a diet. Now the real problem is not whether the data arrives on time, but whether your wallet can handle the price of a Blackwell GPU equipped with these stacks. Because yes, the memory flies, but your bank account will probably make a hard landing. At least the fan won't have to work as hard.