UCY-CompSci (@ucompsci) 's Twitter Profile
UCY-CompSci

@ucompsci

High-performance computing with a mission: contribute to a better world.

ID: 1349669026193141760

calendar_today14-01-2021 10:47:07

36 Tweet

16 Followers

8 Following

Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

Happy to share โ€œMLX Guruโ€ ๐Ÿš€. It knows everything about MLX & helps w/ coding the framework. Not perfect but speeds up development especially if you are new to MLX. It is free but needs ChatGPT Plus. As a thanks to the MLX team Awni Hannun Available at: chat.openai.com/g/g-7NeyFkq2e-โ€ฆ

Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

On more thing ๐Ÿค”: while better to avoid long chats, if youโ€™ve just started a chat, itโ€™s always good to ask โ€˜MLX Guruโ€™ to read again itโ€™s knowledge base when thereโ€™s an update. MLXG updates will reflect MLX versions. To confirm ask MLXG what MLX version it knows about Awni Hannun

Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

Healthy adults face varied risks from exposure to atmospheric dust since the ratio of peripheral (alveolar) to large conducting airway deposition of particles can vary by 8X across individuals. Weโ€™ve developed an AI-based method to non-invasively assess personal risk UCY-CompSci

Healthy adults face varied risks from exposure to atmospheric dust since the ratio of peripheral (alveolar) to large conducting airway deposition of particles can vary by 8X across individuals. Weโ€™ve developed an AI-based method to non-invasively assess personal risk <a href="/UCompsci/">UCY-CompSci</a>
Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

๐Ÿš€๐Ÿš€๐Ÿš€๐Ÿš€ Happy to report that linear scaling achieved with 4 Mac Studio nodes, which is the max we can have without using a TB hub. Speedup: 4 nodes 4.08 x faster than single node ๐Ÿš€๐Ÿš€๐Ÿš€๐Ÿš€ We plan to share setup instructions on Medium soon. Awni Hannun Angelos Katharopoulos

๐Ÿš€๐Ÿš€๐Ÿš€๐Ÿš€ Happy to report that linear scaling achieved with 4 Mac Studio nodes, which is the max we can have without using a TB hub. 

Speedup: 4 nodes 4.08 x  faster than single node ๐Ÿš€๐Ÿš€๐Ÿš€๐Ÿš€

We plan to share setup instructions on Medium soon. 

<a href="/awnihannun/">Awni Hannun</a> <a href="/angeloskath/">Angelos Katharopoulos</a>
Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

Some like CalDigit Thunderbolt 4 Element Hub do offer 40 Gb/s on all TB ports. But as correctly pointed out by elegyal ห† โ€ขอˆ ยทฬซ โ€ขอˆ ห† โ™ช you can connect up to 7 MacStudio M2 Ultra using either an iPad as monitor over WiFi or other non TB monitor (over usb or hdmi)

Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

Rinat Shay. In the lab we saw a trend of improving speedup ratio as we went from 1 to 2 to 3 to 4 nodes, Iโ€™m suspecting it has to do with more efficient use of lazy evaluation in each node. To be fair we should average performance over many more tries for each setup & compare averages

AppleInsider (@appleinsider) 's Twitter Profile Photo

Xgrid lives -- a new project resurrects the promise of Apple's dead clustering software appleinsider.com/articles/24/06โ€ฆ #Apple

Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

๐Ÿš€๐Ÿš€๐Ÿš€๐Ÿข We verified that if you mix an M1 Max w/ several M2 Ultra you can almost recover full performance of using all M2U by proportionally reducing batch size on M1, increasing it on M2, keeping total # of epochs the same. Might depend on case/size. Awni Hannun Angelos Katharopoulos

๐Ÿš€๐Ÿš€๐Ÿš€๐Ÿข We verified that if you mix an M1 Max w/ several M2 Ultra you can almost recover full performance of using all M2U by proportionally reducing batch size on M1, increasing it on M2,  keeping total # of epochs the same. Might depend on case/size. <a href="/awnihannun/">Awni Hannun</a> <a href="/angeloskath/">Angelos Katharopoulos</a>
Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

Thermo Therapist with CoolProp is being released in the GPT store for my Engineering Thermo students. It explains concepts, uses a custom API to access CoolProp, solves problems & will convince you that you should โค๏ธ Thermo. Not responsible for any Transference issues ๐Ÿ˜‚ OpenAI

Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

๐Ÿซ๐ŸŒŠ Collaborating with Josuรฉ Sznitman on this Annual Review was fun & rewarding. It is finally live in early release Annual Reviews : Multiscale Modeling of Respiratory Transport Phenomena and Intersubject Variability | Annual Reviews - go.shr.lc/3Z5KXKj

๐Ÿซ๐ŸŒŠ Collaborating with Josuรฉ Sznitman on this Annual Review was fun &amp; rewarding. It is finally live in early release <a href="/AnnualReviews/">Annual Reviews</a> : Multiscale Modeling of Respiratory Transport Phenomena and Intersubject Variability | Annual Reviews - go.shr.lc/3Z5KXKj
Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

Starting today we initiate a series of science podcasts ๐ŸŽง on X showcasing some of the most exciting research output of UCY-CompSci UCY-RespiHub & collaborators. Thanks to Google for making this possible. These podcasts are themselves a testimony of the real value of AI ๐Ÿš€๐Ÿš€

Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

The second Podcast ๐ŸŽง in our series dives into our efforts UCY-CompSci with collaborators to create a computationally efficient model of the deep lung ๐Ÿซ by taking some clever shortcuts: The cited article is doi.org/10.1016/j.jaerโ€ฆ #AerosolScience #InhalationTherapies #LungHealth

Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

The Deep Dive ๐Ÿš€๐Ÿš€ podcast on our preprint of โ€œBeyond Language: Applying MLX Transformers to Engineering Physicsโ€ doi.org/10.48550/arXivโ€ฆ Code available at github.com/sck-at-ucy/MLXโ€ฆ Awni Hannun Angelos Katharopoulos #AI #ArtificialIntelligence #ML #MLX #NeuralNetworks

Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

Iโ€™m sooo ๐ŸฆŽ๐ŸฆŽhappy ๐ŸฆŽ๐ŸฆŽw/ Kourkoutas-ฮฒ. Completed 1st full test on training a heavy Transformer model as PDE surrogate in data-driven mode. Kourkoutas-ฮฒ blows vanilla Adam out of the desert, especially for smaller training datasets & quantized models where the grads tend to be

Iโ€™m sooo ๐ŸฆŽ๐ŸฆŽhappy ๐ŸฆŽ๐ŸฆŽw/ Kourkoutas-ฮฒ. Completed 1st full test on training a heavy Transformer model as PDE surrogate in data-driven mode. Kourkoutas-ฮฒ blows vanilla Adam out of the desert, especially for smaller training datasets &amp; quantized models where the grads tend to be
Stavros Kassinos (@kassinoss) 's Twitter Profile Photo

๐Ÿฅณ๐Ÿš€ 1st installment of Mac Studios M4 Max for the โ€œHPC+AI Synergiesโ€ teaching lab has arrived. ๐Ÿš€๐Ÿš€ Benefits: ๐Ÿ‘‰๐ŸปUltra low energy footprint ๐Ÿ‘‰๐Ÿป Quiet student-friendly environment for hands on experimentation w/ parallel connectivity topologies ๐Ÿ‘‰๐ŸปStudents get trained on machines

๐Ÿฅณ๐Ÿš€ 1st installment of Mac Studios M4 Max for the โ€œHPC+AI Synergiesโ€ teaching lab has arrived. ๐Ÿš€๐Ÿš€
Benefits:
๐Ÿ‘‰๐ŸปUltra low energy footprint 
๐Ÿ‘‰๐Ÿป Quiet student-friendly environment for hands on experimentation w/ parallel connectivity topologies
๐Ÿ‘‰๐ŸปStudents get trained on machines
U.S. Embassy Cyprus (@usembassycyprus) 's Twitter Profile Photo

CDA Mangis: Pleased to join the Stanford Alumni Association Cyprus (SAAC) dinner. Thanks to cofounders Suzi Abdel-Malak & Effie Kokkinou. SAAC promotes networking among Stanford alumni in Cyprus, strengthening the ๐Ÿ‡บ๐Ÿ‡ธ๐Ÿ‡จ๐Ÿ‡พ connection to the Stanford University family.

CDA Mangis: Pleased to join the Stanford Alumni Association Cyprus (SAAC) dinner. Thanks to cofounders Suzi Abdel-Malak &amp; Effie Kokkinou. SAAC promotes networking among Stanford alumni in Cyprus, strengthening the ๐Ÿ‡บ๐Ÿ‡ธ๐Ÿ‡จ๐Ÿ‡พ connection to the Stanford University family.