Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by andersa on Tuesday August 27, 2024 @08:49AM
(#41366978)
You can't fine tune a 70B model with this. It barely even fits the weights before it runs out of vram. Need a bigger machine.
[ Parent
| on HN ]
Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by StrangeDoctor on Tuesday August 27, 2024 @10:09AM
(#41367711)
I think the idea is you network them together if you need more and most models can be split nicely.
Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by anon389r58r58 on Tuesday August 27, 2024 @10:18AM
(#41367826)
For that you'd probably be better off removing one of the GPUs, and replacing it with a networking card.
The problem of the form factor will remain. The tinybox is 15U big for compute that you'd normally expect to find in a 4U form factor.
Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by andersa on Tuesday August 27, 2024 @10:27AM
(#41367923)
I don't think they're intended for rack usage like that. More like for people to put under their desks... there would be no reason to build the giant case with fancy silent-ish cooling if you're going to put them next to your other jet engines.
Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by anon389r58r58 on Tuesday August 27, 2024 @10:59AM
(#41368300)
Fully agree, and I think the tinybox is great if you put only one of them somewhere in your local office.
I just don't think it makes sense to connect multiple of them into a "cluster" to work with bigger models, as the networking bandwidth isn't good enough and you'd have to fit multiple of these big boxes into your local space. Then I might as well put up a rack in a separate room.
Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by StrangeDoctor on Tuesday August 27, 2024 @11:27AM
(#41368631)
there's an ocp 3.0 mezzanine, so no need to remove a card and you'd get 200gbps, unless I've missed something about needing to remove a card to access it. But yeah stacking these or racking them seems less than ideal.
Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by omikun on Tuesday August 27, 2024 @11:28AM
(#41368639)
3kW under your desk... no need to turn on the heat in the winter!
Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by andersa on Tuesday August 27, 2024 @10:21AM
(#41367857)
Most models actually can't be split nicely by 6. There's a reason nvidia builds nodes with 4 and 8 GPUs.
Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by StrangeDoctor on Tuesday August 27, 2024 @11:31AM
(#41368669)
I don't see why 6 is inherently worse than 4 or 8, not all of the layers are exactly equal or a power of 2 in count. 2^2, 2^3, vs 2^1*3^1 might give you more options.
The main issue I run into mainly is flops vs ram in any given card/model.
Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by andersa on Tuesday August 27, 2024 @12:01PM
(#41368983)
Usually you want to split each layer to run with tensor parallelism, which works optimally if you can assign each kv head to a specific GPU. All currently popular models have a power of 2 number of kv heads.
Re:Tinybox of tinygrad by George Hotz is finally entering pr...
by StrangeDoctor on Tuesday August 27, 2024 @01:02PM
(#41369911)
interesting, thank you for the pointers.