Here's a recent paper/study (10/22/2025) published just a week ago on the efficiency (token/watt) of Q's new AI chip offerings vs NVDA:
(added this comment after I posted: I also just realized that the chip news today refers to the AI200 and AI250, not the AI100, so this study is somewhat dated in terms of technology, perhaps - making today's announcement very significant. Still, an interesting read)
Serving LLMs in HPC Clusters: A Comparative Study of Qualcomm Cloud AI 100 Ultra and NVIDIA Data Center GPUs
arxiv.org
Can't speak much on the topic myself, but QHTrain likely can help make sense of it. Seems like Q's solutions are quite competitive right now, based on the application with some of the older NVDA chips. Someone on the Q Yahoo board (Skuldur) posted this (sort of) rebuttal to today's news:
"You can't compete in Nvidia's arena by going cheap. AMD could give their chips away for free and it would still make more sense to go with Nvidia. It's about tokens per watt.
I don't know enough about Qualcomm to even guess whether their new chips are any good, but with release dates in 2026 they're going up against Rubin, and in 2027 it's Rubin Ultra. Good luck."
That prompted me to search for any comparisons on token/watt. Don't know anything about NDVA's roadmap, but at first glance, this study is interesting on what QCOM can provide out of the gate.
Upon further review - NVDA is already 2 generations beyond the A100's used to compare in this study, per google AI overview:
While the A100 GPU is available through many cloud providers and data center vendors, it is not NVIDIA's latest offering; the company's newer and more performant GPUs include the H100 and H200. The A100 is based on the NVIDIA Ampere architecture and remains widely available for AI, data analytics, and high-performance computing workloads. However, NVIDIA announced an End-of-Life (EOL) for the A100 in January 2024.
Hopefully the power savings makes Q's option as, or more attractive for many more basic applications. Baby steps can lead to sprints eventually! It might take some time to gain a lot of traction vs NVDA, but optimistic.
Not trying to throw water on the fire here - just wishing to understand the (extent of the) significance of today's announcement.
regards,
metman |
