←back to thread

Nvidia won, we all lost

(blog.sebin-nyshkim.net)
977 points todsacerdoti | 2 comments | 04 Jul 25 21:58 UTC | HN request time: 0.418s | source
Show context
benreesman ◴[04 Jul 25 23:40 UTC] No.44468901[source]▶
The thing is, company culture is a real thing. And some cultures are invasive/contagious like kudzu both internally to the company and into adjacent companies that they get comped against. The people get to thinking a certain way, they move around between adjacent companies at far higher rates than to more distant parts of their field, the executives start sitting on one another's boards, before you know it a whole segment is enshittified, and customers feel like captives in an exploitation machine instead of parties to a mutually beneficial transaction in which trade increases the wealth of all.

And you can build mythologies around falsehoods to further reinforce it: "I have a legal obligation to maximize shareholder value." No buddy, you have some very specific restrictions on your ability to sell the company to your cousin (ha!) for a handful of glass beads. You have a legal obligation to bin your wafers the way it says on your own box, but that doesn't seem to bother you.

These days I get a machine like the excellent ASUS Proart P16 (grab one of those before they're all gone if you can) with a little 4060 or 4070 in it that can boot up Pytorch and make sure the model will run forwards and backwards at a contrived size, and then go rent a GB200 or whatever from Latitude or someone (seriously check out Latitude, they're great), or maybe one of those wildly competitive L40 series fly machines (fly whips the llama's ass like nothing since Winamp, check them out too). The GMTek EVO-X1 is a pretty capable little ROCm inference machine for under 1000, its big brother is nipping at the heels of a DGX Spark under 2k. There is good stuff out there but its all from non-incumbent angles.

I don't game anymore but if I did I would be paying a lot of attention to ARC, I've heard great things.

Fuck the cloud and their ancient Xeon SKUs for more than Latitude charges for 5Ghz EPYC. Fuck NVIDIA gaming retail rat race, its an electrical as well as moral hazard in 2025.

It's a shame we all have to be tricky to get what used to be a halfway fair deal 5-10 years ago (and 20 years ago they passed a HUGE part of the scaling bonanza down to the consumer), but its possible to compute well in 2025.

replies(2): >>44469534 #>>44470203 #
1. 827a ◴[05 Jul 25 05:06 UTC] No.44470203[source]▶
> Fuck the cloud and their ancient Xeon SKUs

Dude, no one talks about this and it drives me up the wall. The only way to guarantee modern CPUs from any cloud provider is to explicitly provision really new instance types. If you use any higher-level abstracted services (Fargate, Cloud Run, Lambda, whatever) you get salvation army second-hand CPUs from 15 years ago, you're billed by the second so the slower, older CPUs screw you over there, and you pay a 30%+ premium over the lower-level instances because its a "managed service". Its insane and extremely sad that so many customers put up with it.

replies(1): >>44470337 #
2. benreesman ◴[05 Jul 25 05:42 UTC] No.44470337[source]▶
Bare metal is priced like it always was but is mad convenient now. latitude.sh is my favorite, but there are a bunch of providers that are maybe a little less polished.

It's also way faster to deploy and easier to operate now. And mad global, I've needed to do it all over the world (a lot of places the shit works flawlessly and you can get Ryzen SKUs for nothing).

Protip: burn a partition of Ubuntu 24.04 LTS which is the default on everything and use that as "premium IPMI", even if you run Ubuntu. you can always boot into a known perfect thing with all the tools to tweak whatever. If I have to even restart on I just image it, faster than launching a VM on EC2.