Edwin Mascarenhas

I like building fast computing systems. Deep Learning Architect @nvidia.

[Edwin Mascarenhas @Kooledd]

I have a joke on a GPU but you CUDAn't get it.

[Edwin Mascarenhas @Kooledd]

🚀🚀🚀

[Elon Musk @elonmusk]

SpaceX has committed to using Nvidia GPUs exclusively because they are the best

[Edwin Mascarenhas @Kooledd]

@CNNnews18 @SmritiMishra22 Very 1984: Two minutes hate from the ministry of truth.

[Edwin Mascarenhas @Kooledd]

@ShivAroor The dog whistle has been sounded and all of you have assembled for the witch-hunt. When are you going to pivot to real journalism for a change? What a sellout.

[Edwin Mascarenhas @Kooledd]

@TimesNow Priorities lol

[Edwin Mascarenhas @Kooledd]

@dpradhanbjp @RahulGandhi @INCIndia It’s better use of everyone’s time and tokens if you type your resignation instead of tweets and whataboutery.

[Edwin Mascarenhas @Kooledd]

“My CPU is a neural-net processor.”

This almost aged well. Just missed the part where we’d run them on GPUs.

[Edwin Mascarenhas @Kooledd]

@PMOIndia

[Sarthak Sidhant @sidhant_sarthak]

@cbseindia29 @EduMinOfIndia @PIB_Edu @PIB_India @AkashvaniAIR @airnewsalerts @DDNewslive @dpradhanbjp @sanjayjavin Can u please extend the deadline please because i was busy this week exposing you

[Edwin Mascarenhas @Kooledd]

@petergyang Created symlink. My Claude.md is just a symlink to agents.md

[Edwin Mascarenhas @Kooledd]

@anjanaomkashyap Kids do better journalism than you, catch up.

[Edwin Mascarenhas @Kooledd]

@cbseindia29 Finally!!! Took you long enough and multiple pwnd instances to admit this.

[Elon Musk @elonmusk]

The GB300 is the best AI computer

[Ben Pouladian @benitoz]

Wow @SemiAnalysis just dropped day-zero numbers on @deepseek_ai V4 Pro

Blackwell (B300): 8,075 tok/s/GPU
AMD MI355X: 6.99 tok/s/GPU
Hopper (H200): 186 tok/s/GPU at higher interactivity

Same interactivity. ~1,000× the throughput per GPU vs MI355X. Blackwell also keeps serving all the way out to 80 tok/s/user, more than 2× the range where AMD flatlines

That number isn’t a typo!

But the throughput isn’t the story

The story is the words “Day Zero”

DeepSeek V4 Pro is a 1.6T-parameter Chinese frontier model. It dropped Friday. Shortly after, the open source community released day-0 deployment recipes, with NVFP4 checkpoints, TensorRT-LLM + WideEP, Dynamo, optimized V4 kernels in the pipeline

That ecosystem doesn’t exist for the alternatives. That’s the real moat

If you’re a neocloud operator, an enterprise AI buyer, or a hyperscaler sizing your next infrastructure build, this is what you’re actually buying when you buy NVIDIA

You’re not buying a chip. You’re buying the guarantee that when the next frontier model drops, Chinese or American, open or closed, reasoning or multimodal, your fleet gets strong day-0 performance with continuous improvements over time

I wrote about this four weeks ago in 27× On The Same Iron. MLPerf v6.0 showed GB300 delivering 2.7× more throughput than its v5.1 debut. Same silicon. Pure software. 60%+ cost-per-token reduction. Every operator already running Blackwell got that gain for free

That was the proof point on installed-base compounding. This is the proof point on day-zero readiness for models that didn’t exist when you bought the iron

Now look at the alternative

You buy MI355X. The next frontier model drops. Now you wait. For AMD’s kernel team. For SGLang or vLLM upstream. For the open-source community to prioritize a stack that ships on a fraction of the world’s deployed compute. Weeks. Sometimes months. Sometimes never at peak

Same story for any TPU buyer who isn’t Google. Same story for anyone betting on chip startups that haven’t earned a place in the open-source optimization queue

Hopper, the last generation, is beating brand-new AMD silicon by a wide margin on a model that didn’t exist a week ago

That’s not a benchmark. That’s a procurement decision

Day zero is the moat

$NVDA $AMD $GOOGL

[Edwin Mascarenhas @Kooledd]

@LayoffAI You’re clearly double counting people over and over and artificially inflating the number for optics. Hypothetically - If a person switches employers ie transfer every year over 10 years you’re increasing the cumulative count by 10. It’s just the same 1 person 🤷

[Edwin Mascarenhas @Kooledd]

@bawejakunal Where is this?

[Edwin Mascarenhas @Kooledd]

I feel bad for the LLMs that will be trained on all the accumulated AI slop in 10 years.

[Edwin Mascarenhas @Kooledd]

Want a break-in have a Kit Kat

[KITKAT @KITKAT]

Regarding recent press coverage

[Edwin Mascarenhas @Kooledd]

I’m the largest LLM bottleneck I can’t prompt fast enough

[Edwin Mascarenhas @Kooledd]

@MuzafferKal_ Haha true that let’s save those token costs

[Edwin Mascarenhas @Kooledd]

I keep typing :q at the end of my conversation with ChatGPT. I need vim bindings asap

[Edwin Mascarenhas @Kooledd]

@VipulDivyanshu 8.3% util isn’t great.
The numbers don’t add up, 8.3% peak util = 0.083*19 peak fp16 ANE TFLOPs = 1.57 delivered TFLOPs. At 1.2W from your screenshot you’re more like at 1.3 TFLOPs/W. The 80x number is very exaggerated. A100 is 312 peak fp16 TFLOPs at 400W.