:apa: スプリットショックウイルス † on Nostr: Mondobizarrro Nvidia is the Rolex of GPUs, they like to gatekeep people from buying ...
Mondobizarrro (npub1pq0…cs9j) Nvidia is the Rolex of GPUs, they like to gatekeep people from buying their cards, either by price hikes or by not selling to them.
The devs in my company like to purchase shit we don't need and then never do anything with it because they can't figure it out or "insert-excuse".
AMD, Advanced Micro Devices sent us two expensive 2U Epyc servers free of charge to review, and guess what we did with them? NOTHING, we didn't review them we didn't do shit, we haven't even sent them back because then never asked. Probably because AMD is making so much money from actual datacenters that they don't give a shit about slippage. Meanwhile our CPUs underperform in everything except SQL it's pretty embarrassing.
So what does my company want to do now instead of fixing broken their product? They want to add GPU compute to it. But guess what Nvidia actually has QoS policies in regards to reselling their cards which means your service must meet certain performance requirements and benchmarks and your staff must be competent to their (Nvidia's) training certifications otherwise you are not allowed to resell their cards as a service. You can buy and use as many cards as you want and use them in your own internel service, but you are not allowed to virtualize them and split pass through GPUs to sell to people if you don't know how to configure them correctly.
Because if someone is paying $400+ per month for a GPU VM, that VM better be the best VM on the planet at that price. And when that VM runs bad because the VPS didn't configure the hypervisor and hardware correctly the customer is not going to think, "hmm maybe the VPS host configured the VM wrong" no they are going to think that Nvidia VMs are bad. So instead of having to deal with incompetent middlemen they'll just revoke the license for the company that just sucks at running VMs.
These cards cost a lot of money, think like $40,000. For $40,000 you could double our saleries, or hire more people, or send someone to college for four years. Instead of getting training I have a feeling they are just going to dick around with the card trying to get competent, but ultimately they won't and then Nvidia is going to prevent them from selling a product because the devs can't figure out language models, ML and whatever else you use GPUs in datacenters for.
The devs in my company like to purchase shit we don't need and then never do anything with it because they can't figure it out or "insert-excuse".
AMD, Advanced Micro Devices sent us two expensive 2U Epyc servers free of charge to review, and guess what we did with them? NOTHING, we didn't review them we didn't do shit, we haven't even sent them back because then never asked. Probably because AMD is making so much money from actual datacenters that they don't give a shit about slippage. Meanwhile our CPUs underperform in everything except SQL it's pretty embarrassing.
So what does my company want to do now instead of fixing broken their product? They want to add GPU compute to it. But guess what Nvidia actually has QoS policies in regards to reselling their cards which means your service must meet certain performance requirements and benchmarks and your staff must be competent to their (Nvidia's) training certifications otherwise you are not allowed to resell their cards as a service. You can buy and use as many cards as you want and use them in your own internel service, but you are not allowed to virtualize them and split pass through GPUs to sell to people if you don't know how to configure them correctly.
Because if someone is paying $400+ per month for a GPU VM, that VM better be the best VM on the planet at that price. And when that VM runs bad because the VPS didn't configure the hypervisor and hardware correctly the customer is not going to think, "hmm maybe the VPS host configured the VM wrong" no they are going to think that Nvidia VMs are bad. So instead of having to deal with incompetent middlemen they'll just revoke the license for the company that just sucks at running VMs.
These cards cost a lot of money, think like $40,000. For $40,000 you could double our saleries, or hire more people, or send someone to college for four years. Instead of getting training I have a feeling they are just going to dick around with the card trying to get competent, but ultimately they won't and then Nvidia is going to prevent them from selling a product because the devs can't figure out language models, ML and whatever else you use GPUs in datacenters for.