[β]
β
bitchan
[?] help
π
host1.bitchan.one
π
@Hb9FTvfi
β±
3d seen
β
0
π
0
β§
1
π hide
send βΏitcoin
[!] follow
[@] review
posts 1
/ 1
[β₯] sort oldest
β±
filter [-]
languages:
on
other
/technology
I was kind of suprised at some of the responses I got in this general (don;t believe you post pics etc etc) as if it was sometjhing special when I said I was using stuff like sxm2 V100s and sxm4 A100s and I was wondering why because while there are undoubtedly mouthbreathers and gaymers posting there are also tech obsessives. I realised why and it is quite interesting. I am in Europe. Now I did not realise what has happened to the US but I was talking with a tech old colleague in texas and understand better. I can get v100s here locally for about 95 euro and external nvlink four way boards for 300 euro shipped form china. I can get a100s 40GB for about 2000 euro (the first one I had was given to me by someone I was working for). I can get A100 boards from china for about 500 euro. I am working on a 4xA100 80GB build that would give me parity with frontier rn. It appears that US customs will aggressively stop any AI related material flowing between the US and China and that a lot of his stuff he tried to import got stuck in a customs hell. I guess this is why you anons are not doing this so much now even though this absolutely crushes consumer garbage like 5090s or mac minis at pretty low cost. It seems to me that one of you could make great money cloning the chinese boards and making them in the USA. Anyhow fun fact the V100s stripped form the US supercomputer Summit (OLCF-4) all went in containers to China for resale (I get mine from local datacentre decomissioning). If you are looking for an experimental entry point at bottom dollar probably the cheapest for you would be a PCI adaptor anything even an old 1x16 riser and a couple oif high rpm case fans and a v100 16GB with the 9cm heatsink. Just stick the two fans on it like you would a thermalright CPU cooller and turn it on connected to whatever like a 1x PCIe slot and locad a model that fits and leaves context, after it has loaded it will run perfoecly fine and output plenty of T/S. 300$ project
β±
3d
β©