POPA Recon🔗메모리 다음은 연결성 source
🔗메모리 다음은 연결성 source 1. AI bottleneck is on the move - GPU → Power and Cooling → Through HBM, data movement is now an issue - Slow connection between chips and racks reduces computational efficiency even with more GPUs - The larger the AI cluster, the greater the importance of networks and interconnects 2. Connectivity Beneficiaries Are Not One Basket - Network architecture: AVGO, ANET - Rack internal connection: ALAB, CRDO - Optical communication/parts: LITE, COHR - Same AI connectivity theme has significantly different real-world benefit intensity 3. The real bottleneck is in margins - Si…
The excerpt is quoted from a public Telegram channel, which owns its text.