Meta Vistara CXL (Re-Using DDR4), OpenAI Reduces Inference Costs, Apple / Qualcomm, & More —(Semis & IT Hardware)

Research evidence excerpt

Meta Vistara CXL (Re-Using DDR4), OpenAI Reduces Inference Costs, Apple / Qualcomm, & More —(Semis & IT Hardware)

Meta Vistara CXL (Re-Using DDR4), OpenAI Reduces Inference Costs, Apple / Qualcomm, & More —(Semis & IT Hardware) Equity Research

• Memory Implications: While we think increasing evidence of expanding CXL adoption could be

viewed as a DRAM industry negative (i.e., refuse of de-commissioned DDR4, higher overall DRAM

utilization across servers, etc.), we continue to think CXL deployments could be equally driven

by the continued need to expand overall memory bandwidth + capacity to support expanding

KVcache management requirements.

• CPU Core Count Constraints: While Meta's Vistara CXL solution addresses memory constraints,

we think is very interesting to highlight that approximately 50% of the company's server fleet is

bottle-necked by CPU cores.

2. An Interesting Read - Three HPC Gurus Ask: Do We Still Need GPUs?

• We think The Next Platform published what we believe is interesting read discussing whether the

evolution of CPUs could take on some of the computational (e.g. math matrix) capabilities of GPUs.

• The article re-highlights the recent Top500 list of supercomputers that showcased the new #1

China LineShine supercomputer based entirely on 304-core Arm-based LX2 CPUs.

• While the conclusion is not that GPUs go away, but rather that CPU architectures will evolve to

take on GPU functionality. The article concludes by noting that it is highly likely that NVDA's future

Arm-based CPUs will likely include tensor cores that run the CUDA-X software stack.

3. PC & Server Specific DRAM Contract Pricing (InSpectrum).

• This morning we would highlight the InSpectrum data on DRAM contract pricing specific to PC

DRAM and Server DRAM.