🚀 Scaling LLMs is no longer just about more GPUs—it's about the geometry of the cluster-Stanford's CS336 Lecture 8 episode artwork

EPISODE · May 9, 2026 · 7 MIN

🚀 Scaling LLMs is no longer just about more GPUs—it's about the geometry of the cluster-Stanford's CS336 Lecture 8

from Steven AI Talk · host Steven

🚀 Scaling LLMs is no longer just about more GPUs—it's about the geometry of the cluster.In Stanford's CS336 Lecture 8, we dive deep into the parallelization strategies that make training trillion-parameter models possible. From Zero Redundancy Optimizers (ZeRO) to 4D parallelism, the complexity is staggering.Key Takeaways: 🔹 ZeRO-3 (FSDP) allows sharding parameters "almost for free" on high-speed networks. 🔹 Tensor Parallelism is mandatory for intra-node scaling but relies on massive bandwidth. 🔹 Pipeline Parallelism is the bridge for cross-node training, now improved with "Zero-Bubble" techniques. 🔹 Expert Parallelism (MoE) decouples MLP layers for sparse routing efficiency.The golden rule? Use all sharding methods until the model fits in memory, then scale with Data Parallelism.Check out the full technical summary and transcripts in our repo! All my links: https://linktr.ee/learnbydoingwithsteven#learnbydoingwithsteven #LLM #AIResearch #DistributedComputing #MachineLearning #DeepLearning #StanfordCS336 #GPU #TPU #ModelParallelism #DataParallelism

🚀 Scaling LLMs is no longer just about more GPUs—it's about the geometry of the cluster.In Stanford's CS336 Lecture 8, we dive deep into the parallelization strategies that make training trillion-parameter models possible. From Zero Redundancy Optimizers (ZeRO) to 4D parallelism, the complexity is staggering.Key Takeaways: 🔹 ZeRO-3 (FSDP) allows sharding parameters "almost for free" on high-speed networks. 🔹 Tensor Parallelism is mandatory for intra-node scaling but relies on massive bandwidth. 🔹 Pipeline Parallelism is the bridge for cross-node training, now improved with "Zero-Bubble" techniques. 🔹 Expert Parallelism (MoE) decouples MLP layers for sparse routing efficiency.The golden rule? Use all sharding methods until the model fits in memory, then scale with Data Parallelism.Check out the full technical summary and transcripts in our repo! All my links: https://linktr.ee/learnbydoingwithsteven#learnbydoingwithsteven #LLM #AIResearch #DistributedComputing #MachineLearning #DeepLearning #StanfordCS336 #GPU #TPU #ModelParallelism #DataParallelism

NOW PLAYING

🚀 Scaling LLMs is no longer just about more GPUs—it's about the geometry of the cluster-Stanford's CS336 Lecture 8

0:00 7:08

No transcript for this episode yet

We transcribe on demand. Request one and we'll notify you when it's ready — usually under 10 minutes.

No similar episodes found.

No similar podcasts found.

Frequently Asked Questions

How long is this episode of Steven AI Talk?

This episode is 7 minutes long.

When was this Steven AI Talk episode published?

This episode was published on May 9, 2026.

What is this episode about?

🚀 Scaling LLMs is no longer just about more GPUs—it's about the geometry of the cluster.In Stanford's CS336 Lecture 8, we dive deep into the parallelization strategies that make training trillion-parameter models possible. From Zero Redundancy...

Can I download this Steven AI Talk episode?

Yes, you can download this episode by clicking the download button on the episode player, or subscribe to the podcast in your preferred podcast app for automatic downloads.
URL copied to clipboard!