Dr. Seelam

Scientist at IBM Research. Works in Cloud, AI, and HPC HW and SW. Tweets are my own. FLL Coach, Science Olympiad coach.

[Dr. Seelam @sseelam]

@airindia My son is allergic and I am trying to request nut free meal, after spending a lot of time on the phone with the agent, I gave up. Can someone please help? The PNR is 9QWRCP, Samrudh Seelam, please update his meal preference to Nut free or no-nut veg or non-veg meal. ty

[Dr. Seelam @sseelam]

@airindia This issue is now resolved, thank you @airindia team. I appreciate your prompt response.

[Dr. Seelam @sseelam]

@airindia I have been trying to modify my return journey for the past 2 days with no luck. I cannot reach anyone over the phone. Please help.

[Dr. Seelam @sseelam]

I also think folks in industry should do more to work with Academia. Teach, collaborate, hire interns, provide access to cloud services.

[Torsten Hoefler 🇨🇭 @thoefler]

Academia needs to re-think its strategy and consider stronger industry collaboration!

Authors of notable #AI systems used to be mostly academics in the early 2000s and now it's less than 17% with more than 80% of authors having some industry connection!

https://t.co/TqO4qz1gRl

[Rob Thomas @robdthomas]

Careers are like exclusive nightclubs.

[Dr. Seelam @sseelam]

correct link: https://t.co/rDcVfC5MOF

[Dr. Seelam @sseelam]

@SuperbPhoenix @thoefler There no question that security is critical, that is why our recent panel discussion refers to security 30 times in this context: https://t.co/wg7wFdUEom
However, it is a false choice between security and performance, you can and need to accomplish both for this convergence.

[Dr. Seelam @sseelam]

@SuperbPhoenix @thoefler There no question that security is critical, that is why our recent panel discussion refers to security 30 times in this context: https://t.co/wg7wFdUEom
However, it is a false choice between security and performance, you can and need to accomplish both for this convergence.

[Dr. Seelam @sseelam]

Kubernetes has become that common abstraction layer below containers and above the infrastructure.

What else need to happen for this convergence? We discuss Strategies for Economic Sustainability, Standardization, and Data Accessibility: https://t.co/rDcVfC5MOF

[Torsten Hoefler 🇨🇭 @thoefler]

Performance portability and containers seem in conflict?

This can be resolved with layers of system-optimized libaries linked using #OCI hooks or recompilation!

XaaS explains how such a high-performance container ecosystem may work.

https://t.co/EWC5ijRC2n

arXiv: 2401.04552

[Dr. Seelam @sseelam]

@thoefler congratulations @thoefler!

[Dr. Seelam @sseelam]

@laredoj Lots of fun judging

[Dr. Seelam @sseelam]

It will help our community a lot if Meta can publish Ethernet vs IB results at scale on these two systems and discuss their pros and cons, in performance, operations, and cost. This info can help shape the future of Ethernet via @ultraethernet. @soumithchintala @thoefler

[HPC Guru @HPC_Guru]

.@glennklockwood: With assumptions, each cluster would post over 900 PF FP64 on the #Top500

Network: One is 400G RoCE, the other NDR InfiniBand

FUSE client (Tectonic fs) + pNFS (Hammerspace).

Link to @Meta post:

https://t.co/ySbKzM4TBT

#AI #HPC

[David Siroky @dsiroky]

@glennklockwood Be good to see real world data at scale on IB vs ETH/RDMA at scale for DL. Hope Meta publishes their findings.

[Dr. Seelam @sseelam]

@thoefler Any reference to this 100T model? If there is one, it will be fp16 or fp8, sparse model so DRAM and NVMe storage can be used for parts of the model, doesn’t have to be all in HBM. 1.6T Ethernet is needed for #AI!

[Dr. Seelam @sseelam]

@thoefler @IEEEHOTI amen... Ethernet indeed!

[Dr. Seelam @sseelam]

@thoefler @ronbrightwell @MulticoreWorld Great presentation indeed!!

[HPC Guru @HPC_Guru]

According to Citi's price projections for @AMD's #MI300 #AI accelerators, @Nvidia currently charges up to four times more for its competing #H100 GPUs

AMD's chips cost $10 to $15K apiece

Nvidia's H100 has peaked beyond $40K

https://t.co/D9BQ5acDeu

#HPC #GPU via @tomshardware

[HPC Guru @HPC_Guru]

👇Mesmerizing to watch and a trip down memory lane

[Rob Thomas @robdthomas]

In generative AI, open source and narrower proprietary models (domain specific) are increasingly likely to win.

[Bindu Reddy @bindureddy]

Please go back to your roots and become an open (source) AI company!

If not, change the name of the company so it’s not so hypocritical!

[Dr. Seelam @sseelam]

Please see our latest blog on Vela enhancements here: https://t.co/RCcAivPBSK