Supercomputing 24 (SC24): November 2024: Day 6: Georgia Aquarium

Supercomputing 24 (SC24)

Nov 2024: Day 6: Georgia Aquarium

Competition for snacks seems to be dying down.

That is a lot of network traffic. More than 2.5Tbit/sec aggregrate.

Wifi access points here all fed via 10GbE. Some 800GbE but mostly 100GbE. Nice.

Visited the SK hynix booth. PEB110 shipping now says the staff I talked to.

Valvoline does liquid cooling solutions too.

AIC showing off 2-phase cooling?

Inside the exhibit hall. I did not allocate enough time here.

8.5 amps per fan, holy crap.

Some of these university-based supercomputing groups have serious money.

Delta liquid cooling coldplate/loop for NVIDIA Blackwell (GB200).

NVIDIA GB200 air cooling.

1.5" (38mm) tubing with up to 10 bar (145psi) pressure is no joke.

Google Trillium TPU.

Microsoft ND-MI300X-v5.

Heart of SCinet.

Don't eat conference center vendor food.

Had no time tho, so the disgusting Atlanta Wok it was.

OpenHPC Birds of a Feather.

Costs of deep learning.

From the Sustainability, Carbon-Neutrality, and HPC panel.

Relaibility is hard to achieve.

Design for reliability.

Water is a huge issue.

Inside building B of the Georgia World Congress Center.

Storage Management panel.

CoRD: coordinating RAID-based and delta-based update schemes to reduce network traffic.

HUGE line to get in to the final night's party at the Georgia Aquarium.

20 minutes later, now finally around from the front to the back of the building.

Still going from the front to the back and now back to the front...

And now at the main entrance, still a long line, 25 minutes later...

Georgia Aquarium is worth the wait.

Look up! Shark!

Penguins chillin.

An hour after entering, finally found the main dining room. Buffet food was... so-so.

Georgia Aquarium fully done up for Christmas.

Big male sea turtle (long tail!).

Whale shark showing off.

Georgia Aquarium fully done up for Christmas.

Sessions (panels, birds of a feather, papers, etc.) and exhibit hall for vendor visits highlighted today. I clearly should have visited the exhibit hall earlier and spent more time talk to vendors, but still got some useful info while I was there. SK hynix, AIC, Delta Electronics, Microsoft, Intel, Asus all out in force, did not have time to talk to Samsung, Supermicro, Dell, HPE, or a few others (sigh). Some of the university-baed supercomputing groups have crazy budgets as well, as evidenced by several of them with impressive booths...

Liquid cooling and to a lesser extent immersion cooling were hot items, as well as NVIDIA Blackwell GPUs (or at least preparing for them), and some of the real-time network traffic displays for SCInet were cool.

OpenHPC did a very informative birds of a feather, some of the sustainability stuff was useful to keep abreast of (even if it was too large-scale to be relevant to us), and finally, reliability and its impacts on deep learning was very useful to be aware of. Our little single-GPU jobs clearly will need more consideration towards reliability once we get into multiple-GPU jobs, with some very alarming failure rates to the uninitiated seen. A very oddly delivered but useful talk on data/reliability optimization and its effects on network traffic was also interesting, if too abstract to translate into implementation.

Final night's gala was for technical workshop attendees (thousands of us? Not the full 17,000 in attendance, but still a metric butt-ton) at the Georgia Aquarium. The line to get in went from the front to the side to wrap around the back of the building, then back towards the front (!) and took something like 25 or 30 minutes to get in (!). However, Georgia Aquarium is not only a massive venue that can handle this kind of crowd, but also a world-class aquarium, and well-worth the time, especially the dolphins, penguins, and sharks. The food was decent enough once I finally found it (ugh), but can only do so much for a buffet for a crowd of this size...