We are looking for talented engineers to work on ClusterMAX 3.0 and related consulting projects. If you like ClusterMAX, InferenceMAX, or any of our other research on ML Systems and Performance, please consider applying at the link below!

went to a friend's wedding last weekend and there was quite a bit of contrast framing in the pastor's homily

my friend Tanj posting stuff like this into the LinkedIn ether is such a shame

Interesting to see given the administration’s recent trip to China

Even if Unitree doesn't dominate global robotics, American startups still depend on the Chinese hardware supply chain Unitree’s Verticalization Is Remarkable - Even In China "Unitree self-develops BLDC motors, planetary gearboxes, LiDARs, and depth cameras, each typically (or even previously for Unitree) outsourced by other Chinese humanoid OEMs. Instead, Unitree’s self-produced motors can run as low as 30-40% of equivalent Western motors, and they now make some of the cheapest humanoid gearboxes in the world."

People ask this question all the time and then don’t accept valid answers, likely because there are many overlapping reasons to make your own chip. Different chips have different perf/$ on different workloads. With a custom chip you can: - save money on a given workload - improve performance on a given workload (same thing) - have a more predictable supply chain - gain control over software and IP - learn a lot - potentially sell it to other companies The reasons that different companies make custom chips are specific to them. TPU and Trainium are competitive on perf/$ and as a result are largely sold out. But depending on who you are, perf and $ are different. For a lab building a 10k cluster, stacking Broadcom/Marvell GCP/AWS margins is not always better than NVIDIA neocloud. But if you’re GDM, it almost always is because no GCP margin. Also the NVIDIA sw stack is the default. Why take that risk? For MTIA or MAIA or AMD GPUs or Groq or Cerebras or whatever else, the perf/$ on the most important workloads is just not competitive today. So MTIA is used for recsys, and the others are used for single node, small model inference only. And then they don’t produce enough of them for it to matter anyway. For OpenAI and Anthropic, if you’re spending 100B on compute, and a chip program costs 1B for a tapeout why not give it a go? Pick a workload you care about, make some customizations for it (which are all just tradeoffs anyway), establish a new supply chain, and learn some things make some friends along the way.

What are they asking for? I'll do it

We’re the first cloud to bring up an NVIDIA Vera Rubin NVL72 system for validation, another big step in building the next generation of AI infrastructure with NVIDIA.

We have the first @DellTech @nvidia Vera Rubin NVL72 @CoreWeave. Here we go! 🚀