Not that this really takes away from the substance of the article, but the first two paragraphs are giving heavy Claude smell. Semicolons, em dashes, "That sequencing matters"... I guess I'm just a little surprised that anyone could be arsed to take on a hardware project like this but can't be arsed to write their own introduction.
sabareesh 1 hours ago [-]
Appreciate the feedback. I have improved the article.
arjie 2 hours ago [-]
Pass the generated text through some kind of quality prompt. It’s got too much filler right now.
The story is interesting but it’s hard to read because it’s hard to tell which parts are meaningful and which parts are filler.
E.g. “we pulled the card cold - straight from the rig to the workbench”. Okay, but why would going straight from the rig to the workbench make it cold? If anything it would be warm. But it turns out the temperature is meaningful in your story.
sabareesh 1 hours ago [-]
Appreciate the feedback. I have improved the article.
cogman10 2 hours ago [-]
Ok, how are people powering these things? 2.4kW is well beyond a standard circuit in the US. Are people having 240V/30A circuits installed? Are they hijacking the dryer plugs? EV charger plugs? Hottub circuits?
ssl-3 1 hours ago [-]
When I dipped my toes into vaguely-serious eth mining with GPUs 5 or 6 years ago, I just installed a dedicated 240v, 30a circuit in the basement. The run of cabling was short, the basement was unfinished utility space, and the parts didn't cost very much. It went together quickly.
The business of protecting individual power cords was handled by an Eaton PDU that had a 30a twist-lock plug on one side and a couple of rows of current-limited IEC C13 sockets on the other side.
Kirby64 2 hours ago [-]
240V-20A circuits will handle 3.8kW continuous. It’s probably a 240V-20A circuit, as that is what the power supplies typically want. Also, easy to convert an outlet to 240V, if the breaker is dedicated to that outlet. Just requires swapping the breaker and the outlet, not the wires.
FireBeyond 28 minutes ago [-]
I don't think that's exactly common other than for outlets that are already wired for those specialized purposes, no?
Certainly on my panel the only "single outlet" breakers are hot water, AC, oven/stove, dryer.
Kirby64 21 minutes ago [-]
I agree it’s not common. Doesn’t mean you couldn’t do that. If every room has its own breaker for the outlets in that room, you could convert that room to 240V, as an example, though.
Also, another place where you might already have this outlet: some older houses that use window AC units that were larger had 240V 20A outlets. Not common these days, but you can still buy these types of window AC units.
hgoel 2 hours ago [-]
Chaining two PSUs on separate circuits is also an option. If they're using the MaxQ versions though, the total GPU power draw is only ~1200W. The bigger question to me is how are they cooling it? Sticking an AC in that room just doubles the power draw issues.
jmalicki 34 minutes ago [-]
If I wanted to use two circuits without running extension cords, every place I've lived would mean getting electrical rewired.
If you're gonna get rewired you may as well install a 240v circuit, and some 120v 20a sockets while you're at it.
hgoel 1 minutes ago [-]
Fair point.
I'm very close to just running a cord over or devising a way to put my machine closer to a second circuit because my rental is horribly setup and both my bedroom AC and living room desktop (that also doubles as a ML training box) end up on the same circuit.
sabareesh 2 hours ago [-]
It is basically on 2 different circuits/breakers. Asus wrx90e supports 2 psu as well. You may need to synchronize both psu and several adapter for this is available in Amazon. Soon planning to upgrade it to 240V
tjwebbnorfolk 2 hours ago [-]
exactly, I had a 220v 30a circuit installed to run a multi-GPU server in my basement.
I'm air cooling so I set -pl 450 so I'm not running them all at the full 600w
tomaytotomato 2 hours ago [-]
What a time to be alive, I remember 10 years ago as a poor student waiting to buy a ATI Radeon X1600 Pro with 256mb, yes 256mb of RAM.
It cost about £190 in 2006.
Now we have GPUs that are in tens of thousands of pounds with insane performance, but what would their price be without the AI and Datacentre squeeze?
a012 2 hours ago [-]
2006 is 20 years ago
tomaytotomato 11 minutes ago [-]
It feels like 10 years ago for me :D
Thanks for making me feel older now
testing22321 2 hours ago [-]
I remember buying the Radeon PCI with 32MB RAM for $650AUD…
rvba 46 minutes ago [-]
I remember buying 128mb that was added on top of 64mb what brought me to the unbalanced 192 mb setup.
Diablo 2 stopped lagging when a necromancer joined the game and summoned all the skeletons...
On an unrelated note Path of exile 1 still lags even on a 5090
amluto 3 hours ago [-]
I wonder whether those cards ran the model that wrote the nonsense about the forces involved.
Hint: when you have a piece of metal stuck with thermal goop to a lot of components, the force doesn’t “concentrate” on one of them. You need to detach it from each one with however much force is needed to detach it from that component.
sabareesh 3 hours ago [-]
Not sure what really happened but some force or bad solder caused it.
dwroberts 1 hours ago [-]
> Don’t RMA it, and don’t solder it yourself. A local phone-repair chain with a microsolder tech can put a 3 mm SMD part back on a GPU PCB in twenty minutes for the price of dinner. The skill is in your city. You just have to look.
The trouble with this though is, what if that is not the only issue with the card? That’s normally my thought process on reaching for RMA. The unit could be an all-round lemon that should not have passed QA etc. (and as noted in the post itself, working for a week on various tasks is not enough to prove it good)
NwtnsMthd 2 hours ago [-]
It's difficult to speculate as to the exact failure from blurry pictures but the solder on that choke (inductor) looks terrible.
Something went wrong in manufacturing. The solder should have wicked to cover the entire pad, not just a small square, and there should be no (brown) discoloration.
stryakr 2 hours ago [-]
Why does this post sound like it's an AI story based on the inputs from the engineer?
The phrasing is very claude like:
"That cracked joint is the whole story. The card had passed initial bring-up and ran fine at light loads for a week."
"That sequencing matters — it’s why we have a story to tell. The pilot card failed, taught us a lesson, and the lesson is the reason the other three went on without incident."
"Driver swaps, CUDA reinstalls, and inference-engine theories were dead ends I spent hours on. The failure pattern itself told the story — listen to it earlier."
ssl-3 28 minutes ago [-]
Some bot-isms like that are amusing to me.
Stuff like "it's the whole story," "this part matters," and "it's not X" (when X wasn't ever under discussion to begin with).
They're like a bot characterizing to itself what is important, what is unimportant, or sometimes even arguing with itself. Their presentation seems like bits of the internal thinking mechanism leaking into the output queue.
josephg 3 hours ago [-]
Cool post. FYI you might be better off getting one big fan for your "radiator" instead of lots of little fans. Big fans don't need to spin as fast as small fans to push the same amount of air. So they run a lot quieter.
sabareesh 3 hours ago [-]
Sure 140mm fans you may call little but it does need enough static pressure for the radiators. This setup is already several times quieter than stock setup
Those are SM120 so no tmem/tcgen05 and lack of support in main libraries (it's like everybody is focusing on B300/SM100).
For that money I'd buy a single B300, similar total AI TOPS, similar GPU bandwidth aggregated, and only 25% less total memory (probably saved in less implementation complexity), half the energy consumption...
Also by having all SMs local they have the special L1-level interconnect. SMs can collaborate on the same GEMM. And a bunch of other nice features.
Or, you know, rent it.
sabareesh 1 hours ago [-]
Yes this has been on my mind as well. But this was built one at a time but still overall very happy with them
voidUpdate 3 hours ago [-]
Is that little computer training LLMs from scratch all by itself? That must take years to get any kind of progress, given the scale of training other providers do. Where do you get the training data from?
sabareesh 3 hours ago [-]
Most of the training i am working on is with post training. You can do so much with a system that is running 24/7
atemerev 3 hours ago [-]
You can train TinyStories in a few hours on retail hardware, and this is a highly illuminating experience that I can recommend for everyone.
sieabahlpark 3 hours ago [-]
[dead]
robin_reala 3 hours ago [-]
Complete side note, but I can’t work out how the author managed to mistype “at” as “Δt”.
Edit: reading fail on my part, nothing to see here.
xmichael909 3 hours ago [-]
I caught that too, probably a qwen bug (;
OneDeuxTriSeiGo 3 hours ago [-]
huh? the only Δt in the article is used correctly.
> With 18× 140 mm of surface, the fans run quietly and the coolant Δt across the rads stays small
robin_reala 3 hours ago [-]
Hah, wow, I completely misread it. Delta-t makes sense when you get the context right, thanks.
iagooar 2 hours ago [-]
Out of curiosity, what are you training with these cards?
sabareesh 2 hours ago [-]
I am primarily experimenting on post training stack. As of now working on training a model that is natively RLM https://github.com/alexzhang13/rlm
iagooar 39 minutes ago [-]
Nice. Are you working on it for "fun" or as a lab? I am looking into building a semi-professional cluster + setting up a lab, but the investment needed is beyond what I could justify as a hobby.
sabareesh 4 days ago [-]
Converting four RTX PRO 6000 Blackwell cards to waterblocks, finding a VRM choke loose on the workbench, and getting back to 41k tok/s.
sandworm101 2 hours ago [-]
Ditch the tiny DC fans. Build a shroud and switch to a single ac-powered industrial blower / duct fan.
AnthonBerg 32 minutes ago [-]
dB?
atemerev 3 hours ago [-]
If you want ready, well engineered, water-cooled multi-GPU research workstations, my colleagues at https://comino.com build and sell them. Or you can purchase fitted waterblocks from them for many GPUs, and build your own.
lightedman 1 hours ago [-]
You can tell this is AI slop by the horrible soldering descriptions - anyone with experience can look at that VRM and go "Oh, the solder is still in its stencil-applied state and has not flowed across the contacts at all, on either component or board. This is a reflow-in-oven issue from the manufacturer." This wasn't a cracking joint this was a poorly-done joint.
The story is interesting but it’s hard to read because it’s hard to tell which parts are meaningful and which parts are filler.
E.g. “we pulled the card cold - straight from the rig to the workbench”. Okay, but why would going straight from the rig to the workbench make it cold? If anything it would be warm. But it turns out the temperature is meaningful in your story.
The business of protecting individual power cords was handled by an Eaton PDU that had a 30a twist-lock plug on one side and a couple of rows of current-limited IEC C13 sockets on the other side.
Certainly on my panel the only "single outlet" breakers are hot water, AC, oven/stove, dryer.
Also, another place where you might already have this outlet: some older houses that use window AC units that were larger had 240V 20A outlets. Not common these days, but you can still buy these types of window AC units.
If you're gonna get rewired you may as well install a 240v circuit, and some 120v 20a sockets while you're at it.
I'm very close to just running a cord over or devising a way to put my machine closer to a second circuit because my rental is horribly setup and both my bedroom AC and living room desktop (that also doubles as a ML training box) end up on the same circuit.
I'm air cooling so I set -pl 450 so I'm not running them all at the full 600w
It cost about £190 in 2006.
Now we have GPUs that are in tens of thousands of pounds with insane performance, but what would their price be without the AI and Datacentre squeeze?
Thanks for making me feel older now
Diablo 2 stopped lagging when a necromancer joined the game and summoned all the skeletons...
On an unrelated note Path of exile 1 still lags even on a 5090
Hint: when you have a piece of metal stuck with thermal goop to a lot of components, the force doesn’t “concentrate” on one of them. You need to detach it from each one with however much force is needed to detach it from that component.
The trouble with this though is, what if that is not the only issue with the card? That’s normally my thought process on reaching for RMA. The unit could be an all-round lemon that should not have passed QA etc. (and as noted in the post itself, working for a week on various tasks is not enough to prove it good)
Something went wrong in manufacturing. The solder should have wicked to cover the entire pad, not just a small square, and there should be no (brown) discoloration.
The phrasing is very claude like:
"That cracked joint is the whole story. The card had passed initial bring-up and ran fine at light loads for a week."
"That sequencing matters — it’s why we have a story to tell. The pilot card failed, taught us a lesson, and the lesson is the reason the other three went on without incident."
"Driver swaps, CUDA reinstalls, and inference-engine theories were dead ends I spent hours on. The failure pattern itself told the story — listen to it earlier."
Stuff like "it's the whole story," "this part matters," and "it's not X" (when X wasn't ever under discussion to begin with).
They're like a bot characterizing to itself what is important, what is unimportant, or sometimes even arguing with itself. Their presentation seems like bits of the internal thinking mechanism leaking into the output queue.
Those are SM120 so no tmem/tcgen05 and lack of support in main libraries (it's like everybody is focusing on B300/SM100).
For that money I'd buy a single B300, similar total AI TOPS, similar GPU bandwidth aggregated, and only 25% less total memory (probably saved in less implementation complexity), half the energy consumption...
Also by having all SMs local they have the special L1-level interconnect. SMs can collaborate on the same GEMM. And a bunch of other nice features.
Or, you know, rent it.
Edit: reading fail on my part, nothing to see here.
> With 18× 140 mm of surface, the fans run quietly and the coolant Δt across the rads stays small
Signed, IPC-610 certified tech.
I picked up on it too, this wouldn't have been something difficult to share but it's far too verbose to be a real person's words in this way.