Joydeep

Senior Staff Engineer at Samsung SSIR. Primary interests: machine learning & NLP. Author of books on fasttext and ML with Rust. AI to make peoples lives easier.

[Joydeep @alt227Joydeep]

One of the first steps in any NLP pipeline is tokenization. It is the process of splitting a given text into a sequence of units that are related to the vocabulary of the corpus. Thus we get a representation of a given text. ==thread==

[Joydeep @alt227Joydeep]

TL;DR of a Blackwell GPU:

160 SMs → each with 128 CUDA + 4 Tensor cores
→ Warps of 32 threads (SIMT)
→ Latency hiding via warp switching
→ Fast local memory → L2 → HBM

Optimize for keeping data close to compute. Everything else follows. 🔥

[Joydeep @alt227Joydeep]

GPUs are behind the rise of AI and the current change in the world order. But what's actually inside these GPUs? 🧵

[Joydeep @alt227Joydeep]

The golden rule of GPU performance:

Keep data as HIGH in the memory hierarchy as possible.

Constantly reading/writing from HBM = slow cores, wasted compute.

Data in registers or shared memory = blazing throughput. This is the whole game.

[Joydeep @alt227Joydeep]

With AI you move from debugging code to debugging your intent.

[Joydeep @alt227Joydeep]

🔗 https://t.co/B2A5WVyskG

Check out my topmate page: https://t.co/jOzxLC0NUD

[Joydeep @alt227Joydeep]

1/ NVIDIA just dropped the Vera Rubin technical report.

I spent last couple of days going through it.

Here's what it tells every AI/ML developer about where to invest their learning time 🧵

[Joydeep @alt227Joydeep]

11/The takeaway: NVIDIA's hardware bets are a signal about what software will matter in the next 2-3 years.

They don't build hardware hoping developers catch up.

They build hardware for where they know developers are going.

Read the report. It's a roadmap.