Suggestions
Yangqing Jia
Vice President at Alibaba Group
Yangqing Jia is an accomplished professional in the field of artificial intelligence and big data with a strong academic background and extensive experience in the industry.
With a Ph.D. in Computer Science from the University of California, Berkeley, where he maintained a perfect 4.0 GPA, Yangqing Jia also holds a Bachelor's degree in Automation from Tsinghua University, graduating summa cum laude. He furthered his education with a Master's degree in Control Science and Engineering, also from Tsinghua University.
Throughout his career, Yangqing Jia has held various prestigious positions at top tech companies. He served as the Former Director of Facebook AI at Facebook, where he contributed to cutting-edge advancements in artificial intelligence. Prior to that, he was the Former VP of AI and Big Data at Alibaba Group, showcasing his expertise in leveraging AI and big data in a business context. He also worked as a Research Scientist at Google Brain and held multiple internship positions at Google.
In addition to his industry roles, Yangqing Jia is affiliated with academia as an Adjunct Professor at Zhejiang University. He has also contributed to the academic community as a Graduate Student Instructor at UC Berkeley. Furthermore, he has a broad international exposure, having been a Visiting Research Engineer at the National University of Singapore.
Passionate about making AI accessible for social good and scientific progress, Yangqing Jia has a keen interest in democratizing artificial intelligence solutions. His experience in building AI infrastructure at Facebook underscores his commitment to advancing the field.
Yangqing Jia's vast array of experiences, from academia to industry giants like Facebook, Google, and Alibaba, position him as a seasoned professional with a unique perspective on AI, big data, and cloud technologies.
Highlights
Over the past year, we've heard from friends and clients about the countless time and effort they've spent trying to make sense of the ever-evolving GPU market. The reality is, not all GPU vendors are the same, and figuring out how to run them reliably at peak performance goes far beyond just comparing prices.
Unfortunately, information on the GPU market used to be very sparse.
At Lepton AI we help our clients to find cost effective, high quality, and performance GPU solutions. We also enable them to create applications otherwise impossible, such as LLM conversation with built-in voice capability with millisecond latency. Part of our job is to be most knowledgeable about all the nitty gritty detail of GPUs - and bring such knowledge to our clients.
We've written up a blog post detailing the many aspects of the GPU market. If you are thinking of either renting or buying GPUs, we hope this would be a very useful read. If you are thinking of investing in GPU related businesses, this could be an introductory article for you.
We really look forward to hearing from you, and don't hesitate to reach out of you need GPUs with a world class platform! DM me on LinkedIn, or shoot an email to [email protected] - and we'll be in touch.
https://lnkd.in/gWNH8PSG
People often ask why prices like $2.8/m token for Llama 405B, while being super fast, are still profitable at Lepton AI. We've even been asked by a leading GPU provider!
So, I figured we should share some technical analysis. This information could benefit the community. We've taken these statistics and analysis for granted, but they might not be obvious to everyone.
-
Big batches: Each request receives an output of ~30 tokens/second. Batching (grouping multiple requests simultaneously) significantly improves total throughput, often 10x or higher than a single request. GPUs are more efficient with larger batches.
-
Dynamic batching: This technique immediately adds a new request to an existing batch instead of making it wait, ensuring the GPU always works at high capacity.
-
Input tokens: The ~30 tokens/second refers to output tokens. Input tokens are processed much faster (known as "prefilling"). Typically, the input length is many times larger than the output (3x to 10x). This increases the total number of tokens processed, explaining why there is often separate billing for input and output.
-
Quantization: Using 8-bit integers or 8-bit floats instead of 16-bit floats reduces memory usage and speeds up processing because the GPU accesses less memory. Newer GPUs also have hardware instructions for lower bit numbers, increasing speed further. For example, the new Nvidia Blackwell GPU supports 4-bit floats (fp4). Quantization also saves memory, allowing even bigger batches from point 1, making it more economic.
-
Speculative decoding: This method uses a smaller model to predict the next token. For example, predicting "you" after "it is good to see" doesn't require a large model. Smaller models make such predictions faster. The Medusa algorithm by Tianle Cai is a specific example of this approach.
-
Prompt caching: LLMs often encounter repeated prefixes, such as "you are a smart AI agent" in system prompts. Caching these prefilled prompts avoids recalculating them, speeding up repeated requests.
-
Optimizing GPU setups: This involves using large GPUs for big models, small GPUs for small models, and matching GPUs to specific tasks—some are better for prefilling, others for decoding. There are many optimization opportunities here.
This is not a complete list. We integrate these methods (and a growing number of more) in our runtime to ensure profitability with reasonable traffic.
Lepton is created by experts who have developed key AI software over the past decade - Caffe, onnx, pytorch - alongside cloud experts like the creator of etcd and core contributors to Kubernetes. We provide not only LLM APIs, but also a full cloud-native experience to help you find, use, and optimize GPUs on our cloud platform.
We love the open-source and open-access community. What AI technical explanation would you like to hear next?
