Wusste gar nicht, dass Nvidia auch Netzwerkadapter herstellt...
#Nvidia #ConnectX, #BlueField: Angreifer können Daten manipulieren | Security https://www.heise.de/news/Nvidia-ConnectX-BlueField-Angreifer-koennen-Daten-manipulieren-10001558.html #Patchday #NvidiaConnectX
Inside the World's Largest AI Supercluster xAI Colossus
---
Summary
#Largest AI Supercomputer: The #xAI #Colossus is built with over 100,000 GPUs, massive storage, and #High Speed Networking, designed for #AI projects beyond typical #Chatbot applications.
#Record Breaking Construction: The facility, containing over 100,000 GPUs, was constructed in just 122 days—significantly faster than traditional #Supercomputers that often take years.
#Advanced Liquid Cooling System: The #Data Halls are equipped with #State Of The Art liquid cooling, using separate pipes for hot water and cold water, which efficiently manages #Heat from the #GPU Servers.
#Scalable GPU Racks: Each rack includes multiple #NVIDIA #HGX H100 units, optimized with #Compact and easily serviceable designs, featuring #Cooling Manifolds and advanced #GPU configurations.
#Ethernet Driven Networking: Unlike most #Supercomputers, the cluster uses #Ethernet Networking with #NVIDIA #Bluefield 3 #DPUs and #Spectrum X switches, offering robust 400 Gbps connections for efficient data flow.
Widescreen Wonder: #LasVegasSphere
54,000 m2 (~3.67 acre) interior LED display (16x16K) and an exterior LED display (‘Exosphere’) consisting out of 1.23 million LED ‘pucks’. Driving all these pixels are around 150 #NVidia RTX #A6000 #GPU, installed in computer systems which are networked using NVidia #BlueField data processing units (#DPU) and NVidia #ConnectX6 NICs (up to 400 Gb/s), with visual content transferred from Sphere Studios in Cali. All this hardware uses 45kW.
https://blogs.nvidia.com/blog/sphere-las-vegas/
@karppinen Mellanox/NVIDIA has been trying to shove #BlueField into any customer box they can for years. They're even mandatory in some configurations (e.g. DGX) and there's no shortage of stock.
The "Self-Hosted DPU Controller" mode mentioned in the video has been officially supported with BSP 4.5.0 since December 2023, but customers like #Netflix and us got access to that long before.
Probably Netflix is actually running this right now at 100 Watts, but we have no confirmation.
@karppinen According to the video stream for that talk, this refered to a prototype that wasn't ready or in use at the moment he talked about it and consumed at least 125 watts when last measured:
https://m.youtube.com/watch?v=q4TZxj-Dq7s
Around 23:00
So while the idea is nothing new and it's quite possible (people have been running Yocto Linux with nginx and Offload directly on #nvidia #Bluefield for a while), those slides do not prove #netflix "can now" do it.
What better #introduction to #fosstodon than a look at what I think was my first #foss contribution. I’d like to apologize now for the unbounded memory allocation bug I introduced. Oops.
https://github.com/php/php-src/commit/72f675be5005f09c048155b85b5d6efd7341f6e0
Since then I’ve done a bunch of #sysadmin work (mostly #Solaris) followed by OS (#solaris #illumos #smartos #zfs) development and more recently various storage stuff (#nvme #nvmeof #roce #spdk #bluefield #dpu #smb).
I love #hiking through the #wilderness and other natural areas.
