🔱 Triton: GPU Programming for Deep Neural Networks
Free news digest about the most important things that happen in the ML world
For decades, the software industry has evolved, removing the dependencies in hardware architectures. Developers don’t spend any cycles thinking about hardware infrastructures when they develop web apps or APIs. The rise of deep learning seems to have brought us all the way back. Optimizations for GPU architectures are a common state in the lifecycle of deep learning models. Data science teams are often puzzled by the differences that GPU topologies can induce in the execution of neural networks. Optimizing for data partitioning, memory allocation, computation distributions, and other aspects are typically beyond the skill set of most data scientists.
The hardware dependencies in deep learning solutions are both a blessing and a curse. In some sense, these dependencies have unleashed a renaissance in AI hardware innovations. On the other hand, most machine learning teams struggle when it comes to GPU optimizations. This week, AI household OpenAI unveiled Triton, a new domain-specific language that abstracts the complexities of GPU optimizations. Triton removes many of the top challenges of GPU optimizations, such as memory coalescing, memory management, and computation scheduling. The core idea is that a machine learning engineer with no GPU experience can perform quite sophisticated optimizations without diving into the underpinning of the hardware. Triton represents one of the most exciting developments towards removing the GPU dependencies existing in deep learning systems today.
🔺🔻TheSequence Scope – our Sunday edition with the industry’s development overview – is free. To receive high-quality content about the most relevant developments in the ML world every Tuesday and Thursday, please subscribe to TheSequence Edge 🔺🔻
🗓 Next week in TheSequence Edge:
Edge#111: what is Attention; what is the biggest transformer model ever built; why is Hugging Face the most popular library for building and using transformer models.
Edge#112: how DeepMind’s compressive transformer improves long-term memory in transformer architectures.
Now, let’s review the most important developments in the AI industry this week
🔎 ML Research
Creating Generally Capable Agents by Playing
DeepMind published a fascinating research paper detailing the models for deep learning agents that can master many different games without human intervention ->read more on DeepMind blog
Facebook AI Research (FAIR) published a paper proposing a technique that combines few-shot-learning and neural architecture search (NAS) to automate the creation of neural networks with minimum computational cost ->read more on FAIR blog
AI and Common Sense
Researchers from MIT and IBM published a paper proposing a benchmark to evaluate signs of common sense reasoning in neural networks ->read more in the original research paper
🛠 Real World ML
Data Movement Architecture at Netflix Studio
The Netflix engineering team published a blog post detailing the architecture behind the ETL pipelines in Netflix Studio ->read more on Netflix technology blog
Containers and Hadoop at Uber
The Uber engineering team published a blog post about their journey to transfer their Hadoop infrastructure to Docker containers ->read more on Uber engineering blog
Sentiment Models at Airbnb
The Airbnb engineering team published a blog post detailing the use of sentiment models to assess customer service quality ->read more on Airbnb tech blog
🤖 Cool AI Tech Releases
OpenAI open-sourced Triton, a Python-like programming language that abstracts the complexities of GPU programming ->read more on OpenAI blog
Google Research published a detailed blog post discussing recent advancements to the TF-Ranking framework for scalable ranking models ->read more on Google Research blog
💎 We recommend
Tecton’s conferences and webinars are very well curated and super practical. Register today, it’s free. And you can win a swag bag ;)
💸 Money in AI
Congratulations to enterprise AI development platform DataRobot with a massive $300 million Series G funding round led by Altimeter Capital and Tiger Global and an acquisition of MLOps startup Algorithmia. 245 job positions are open across the globe.
Conversational customer engagement software developer Dixa raised $105 million in a Series C funding round led by General Atlantic. A lot of job openings across Europe, the US, Israel, and remote.
AI-powered sales support platform Orum raised a $25 million Series A funding round led by Craft Ventures.