ajdecon
@ajdecon.org
Building supercomputers! Former materials physicist, recovering SRE, now mostly herding cats. Perpetually a bit confused. He/him.
Currently at NVIDIA, formerly FB and LANL. Opinions mine as always.
🏡: Denver, CO
🔗 : https://www.ajdecon.org
Currently at NVIDIA, formerly FB and LANL. Opinions mine as always.
🏡: Denver, CO
🔗 : https://www.ajdecon.org
Also I listened to this while making some delightful cream biscuits from Smitten Kitchen, pictured here with some homemade lemon curd 😋
November 29, 2025 at 4:33 PM
Also I listened to this while making some delightful cream biscuits from Smitten Kitchen, pictured here with some homemade lemon curd 😋
A bratty cat who was just caught licking the butter on the counter
November 27, 2025 at 4:03 PM
A bratty cat who was just caught licking the butter on the counter
Yes, we put the Christmas tree up before Thanksgiving. We couldn’t help it, Phryne and Percy love it so much.
November 26, 2025 at 3:59 AM
Yes, we put the Christmas tree up before Thanksgiving. We couldn’t help it, Phryne and Percy love it so much.
I stayed up way too late the other night reading the new edition of @qntm.org’s excellent “There Is No Antimemetics Division”.
If you read the original, I highly recommend checking out the new version too, it’s really well-done.
If you read the original, I highly recommend checking out the new version too, it’s really well-done.
November 22, 2025 at 7:15 PM
I stayed up way too late the other night reading the new edition of @qntm.org’s excellent “There Is No Antimemetics Division”.
If you read the original, I highly recommend checking out the new version too, it’s really well-done.
If you read the original, I highly recommend checking out the new version too, it’s really well-done.
As an aside, one of the things I enjoy about the DGX Spark is that they decided to use the same faceplate design as the “grown-up” DGX servers in datacenters.
After building a bunch of large DGX clusters, it tickles me to have a mini-scale version on my desk.
After building a bunch of large DGX clusters, it tickles me to have a mini-scale version on my desk.
November 15, 2025 at 6:21 PM
As an aside, one of the things I enjoy about the DGX Spark is that they decided to use the same faceplate design as the “grown-up” DGX servers in datacenters.
After building a bunch of large DGX clusters, it tickles me to have a mini-scale version on my desk.
After building a bunch of large DGX clusters, it tickles me to have a mini-scale version on my desk.
Bonus cluster! A baby Raspberry Pi cluster I built back in… 2013? The Lego rack was one of the most fun parts
November 15, 2025 at 6:19 PM
Bonus cluster! A baby Raspberry Pi cluster I built back in… 2013? The Lego rack was one of the most fun parts
The most recent good photo I can find is from Tyche, a GB200 cluster. One of the big features of this architecture was multi-node NVLink, which is great for performance, and works pretty well, but was… fun… to debug early on. 😁
November 15, 2025 at 6:16 PM
The most recent good photo I can find is from Tyche, a GB200 cluster. One of the big features of this architecture was multi-node NVLink, which is great for performance, and works pretty well, but was… fun… to debug early on. 😁
Next up is Nyx, a DGX B200 cluster. Including it here mostly because I found a nice (public) photo. Really no complaints about this system, it’s been pretty well-behaved.
November 15, 2025 at 6:16 PM
Next up is Nyx, a DGX B200 cluster. Including it here mostly because I found a nice (public) photo. Really no complaints about this system, it’s been pretty well-behaved.
After a bunch of other A100 SuperPODs for customers, the next big NVIDIA system I worked on was Eos, based on the DGX H100 server and NDR IB.
I recall Eos being a bit more annoying from a sysadmin standpoint than Selene, but that may be recency bias. Still not as bad as Cielo.
I recall Eos being a bit more annoying from a sysadmin standpoint than Selene, but that may be recency bias. Still not as bad as Cielo.
November 15, 2025 at 6:16 PM
After a bunch of other A100 SuperPODs for customers, the next big NVIDIA system I worked on was Eos, based on the DGX H100 server and NDR IB.
I recall Eos being a bit more annoying from a sysadmin standpoint than Selene, but that may be recency bias. Still not as bad as Cielo.
I recall Eos being a bit more annoying from a sysadmin standpoint than Selene, but that may be recency bias. Still not as bad as Cielo.
Another of the DGX A100 SuperPODs was BioHive 1, built for Recursion Pharmaceuticals for drug discovery workloads. I enjoyed how the gold logos on these racks went with the DGX faceplates.
November 15, 2025 at 6:02 PM
Another of the DGX A100 SuperPODs was BioHive 1, built for Recursion Pharmaceuticals for drug discovery workloads. I enjoyed how the gold logos on these racks went with the DGX faceplates.
Selene spawned a host of SuperPODs based on the same design, and I got to help build a bunch of them. IIRC, HiPerGator at the University of Florida was the first one, though I only worked on this one a bit.
November 15, 2025 at 6:02 PM
Selene spawned a host of SuperPODs based on the same design, and I got to help build a bunch of them. IIRC, HiPerGator at the University of Florida was the first one, though I only worked on this one a bit.
Some time later, after I joined NVIDIA, I got to help build the Selene cluster!
This was the first DGX A100 SuperPOD, and it was brought up in early 2020. Doing the bulk of the work remotely, and with serious restrictions to in-datacenter work, was a big challenge. But the result was pretty great!
This was the first DGX A100 SuperPOD, and it was brought up in early 2020. Doing the bulk of the work remotely, and with serious restrictions to in-datacenter work, was a big challenge. But the result was pretty great!
November 15, 2025 at 5:55 PM
Some time later, after I joined NVIDIA, I got to help build the Selene cluster!
This was the first DGX A100 SuperPOD, and it was brought up in early 2020. Doing the bulk of the work remotely, and with serious restrictions to in-datacenter work, was a big challenge. But the result was pretty great!
This was the first DGX A100 SuperPOD, and it was brought up in early 2020. Doing the bulk of the work remotely, and with serious restrictions to in-datacenter work, was a big challenge. But the result was pretty great!
Next up is Trinity, another Cray system at LANL. This system was interesting because it was half conventional Intel Haswell CPUs, half self-hosted Knights Landing processors.
If anything this was more of a pain than Cielo from an oncall perspective (sigh) but it was a fun build!
If anything this was more of a pain than Cielo from an oncall perspective (sigh) but it was a fun build!
November 15, 2025 at 5:55 PM
Next up is Trinity, another Cray system at LANL. This system was interesting because it was half conventional Intel Haswell CPUs, half self-hosted Knights Landing processors.
If anything this was more of a pain than Cielo from an oncall perspective (sigh) but it was a fun build!
If anything this was more of a pain than Cielo from an oncall perspective (sigh) but it was a fun build!
Silly thread for a Saturday: some of the #HPC clusters I’ve worked on over the years.
First up is Cielo, a Cray XE6 I worked on at LANL! Which might actually be the prettiest supercomputer I’ve worked on.
First up is Cielo, a Cray XE6 I worked on at LANL! Which might actually be the prettiest supercomputer I’ve worked on.
November 15, 2025 at 5:55 PM
Silly thread for a Saturday: some of the #HPC clusters I’ve worked on over the years.
First up is Cielo, a Cray XE6 I worked on at LANL! Which might actually be the prettiest supercomputer I’ve worked on.
First up is Cielo, a Cray XE6 I worked on at LANL! Which might actually be the prettiest supercomputer I’ve worked on.
Just need to brag for a moment about my team who designed the Theia #HPC cluster, as well as all our colleagues who built it, got it to perf, and keep it happy and running smoothly
November 13, 2025 at 2:59 PM
Just need to brag for a moment about my team who designed the Theia #HPC cluster, as well as all our colleagues who built it, got it to perf, and keep it happy and running smoothly
I recently reread Iain Banks’ “Look to Windward”, and I noticed this exchange between an artist and an AI character.
And I agree that creativity is much more about achievement and experience than the end result.
But also, it’s a lot easier to take that attitude in a post-scarcity utopia!
And I agree that creativity is much more about achievement and experience than the end result.
But also, it’s a lot easier to take that attitude in a post-scarcity utopia!
November 4, 2025 at 5:06 PM
I recently reread Iain Banks’ “Look to Windward”, and I noticed this exchange between an artist and an AI character.
And I agree that creativity is much more about achievement and experience than the end result.
But also, it’s a lot easier to take that attitude in a post-scarcity utopia!
And I agree that creativity is much more about achievement and experience than the end result.
But also, it’s a lot easier to take that attitude in a post-scarcity utopia!
Percy really loves our fireplace
October 27, 2025 at 1:17 AM
Percy really loves our fireplace
The tiny new computer is setting itself up
October 26, 2025 at 10:44 PM
The tiny new computer is setting itself up
AFAICT, the line being drawn here in practice is that the verification in question is only for accessing specific content, not all users.
As well as the practical consideration that they can outsource it to the Epic Games verification service.
Plus that the MS law is more likely to be struck down!
As well as the practical consideration that they can outsource it to the Epic Games verification service.
Plus that the MS law is more likely to be struck down!
September 28, 2025 at 7:20 PM
AFAICT, the line being drawn here in practice is that the verification in question is only for accessing specific content, not all users.
As well as the practical consideration that they can outsource it to the Epic Games verification service.
Plus that the MS law is more likely to be struck down!
As well as the practical consideration that they can outsource it to the Epic Games verification service.
Plus that the MS law is more likely to be struck down!
(phrasing likely not exact, typing while listening on a dog walk)
September 28, 2025 at 5:05 PM
(phrasing likely not exact, typing while listening on a dog walk)
