With the cluster and storage down for maintenance this week, I thought about doing some #pytorch development locally an my laptop. No dedicated GPU, but a recent #intel Core Ultra 2 CPU. So first I went for the CPU-only packages.

Only yesterday I realized that the Windows task manager shows something called #npu in its charts. Fair enough, let's see whether pytorch can use it ...

I have to say, the #xpu pytorch integration has gone a long way. In my case, I could use it as a drop-in replacement for #cuda / #rocm pytorch, yet with the obvious performance differences between HPC GPUs and a laptop iGPU.

https://docs.pytorch.org/docs/stable/notes/get_start_xpu.html

Always nice to have alternatives. Especially for basic testing I don't need big (expensive) GPUs.

Marvell is set to acquire Celestial AI for $3.5 B, a move that could supercharge XPU memory access fabrics. The deal brings UALink optical interconnect tech into data‑centre AI workloads, promising lower latency and higher bandwidth. Find out how this merger could reshape the AI hardware landscape. #Marvell #CelestialAI #XPU #MemoryFabric

🔗 https://aidailypost.com/news/marvell-buy-celestial-ai-usd-35-b-boosting-xpu-memory-access-fabrics

#AWS is facing a #cloudcrisis in the #GPU #XPU #Cloud era, trailing behind Microsoft Azure and Google Cloud. However, AWS is poised for a resurgence, driven by its #partnership with #Anthropic, a leading #GenAI company. AWS is investing heavily in datacentres and custom Trainium chips to support Anthropic’s aggressive training and inference needs. https://semianalysis.com/2025/09/03/amazons-ai-resurgence-aws-anthropics-multi-gigawatt-trainium-expansion/?eicker.news #tech #media #news

Then to render the custom AOV,
In the Karma Render Settings node, Image Output > AOVs (render vars) > Extra Render Vars > add new var and name it with the custom name defined in the Karma AOV node in the material #houdini #tip #karma #XPU #solaris

#AI #GPU clusters with one million GPUs are planned for #2027
#Broadcom predicts three of its clients, #hyperscalers, will deploy AI #supercomputers with one million #XPU in fiscal 2027. #HockTan, President and CEO, at Q4 2024 earnings call said "In 2027, we believe each of them plans to deploy 1,000,000 XPU clusters across a single fabric." It was rumored #ByteDance and #OpenAI teamed up with Broadcom to develop their AI chips. Broadcom, of course, does not mention names
https://www.tomshardware.com/tech-industry/artificial-intelligence/ai-gpu-clusters-with-one-million-gpus-are-planned-for-2027-broadcom-says-three-ai-supercomputers-are-in-the-works

#Karma #XPU: A #Houdini 20 Preview | Brian Sharpe | #SIGGRAPH HIVE 2023.
https://www.youtube.com/watch?v=yY1m8NeryaI

I can highly recommend this funny and informative #Houdini talk by @tokyomegaplex

https://youtu.be/bYzffKN7CXQ

We're actually using #Karma #XPU in production right now for some water (where we don't need subsurface scattering). It takes just a few minutes to render (on an RTX 3090Ti) compared to 2 hours (on a Threadripper CPU). As long as the scene fits onto your graphics card...

Remember those XPUs Intel was going to challenge AMD and Nvidia with... well forget about em. Intel says the world isn't ready for its datacenter APUs yet. Falcon Shores will just be a GPU now.

https://www.theregister.com/2023/05/22/intel_abandons_xpu/

#tech #ISC23 #XPU #APU #FalconShores #CPU #GPU #HPC