… It was originally designed with the PCIe form factor in mind. Stuffing two of these chips onto a custom module and dropping it into a 2025 laptop does not change the underlying trade-offs. …
… A single v8t superpod scales to 9,600 chips, holds 2 PB of HBM, and delivers 121 EFLOPS of FP4 compute, nearly 3x the per-pod compute of an Ironwood superpod. v8i pairs 288 GB of HBM with 384 MB of on-chip SRAM 3x Ironwood within a 1,152-chip scale-up domain and claims 80% better performance-per-do… …
… AMD claims the new PRO lineup can run local AI models, including 300B+ parameter LLMs, on an x86 client processor. …