Tiny Corp Details More Of Their Planned Tinybox System Specs

[varikonniemi]

i appreciate tiny for their effort to bring forward AMD:s lousy state of things. Hopefully the news around this will result in better drivers or more open firmware that the community can make better.

[skeevy420]

Proper names in English are capitalized. Tiny is free to use lowercase all they want for branding of course but general literature should probably follow general rules instead of the stylistic rule of the moment of a specific company.

You'll be very popular in the systemd threads. While the proper name rule is true, there can be an exception in regards to following branding or artistic choices like using all caps, lowercase, first letter only, camel case, etc and general literature can follow whichever rule. That's not true with the news where people are expecting to get the facts and to see things written out in the technically correct manner, not in a "correct manner" from blindly following the rules of English. Everything on the tinygrad website is lowercase. Intentionally writing it as Tinygrad, TinyGrad, Tinybox, or TinyBox is like writing it as SystemD or Systemd and can be seen as disrespectful or is intentionally written that way to be disrespectful.

FWIW, I prefer the follow the rules way or camel case. I'd prefer SystemD but I force myself to use systemd. I do the same with tiny stuff. I'm defending something I don't agree with doing because this is a news site and technically correct means something with the news.

[Quackdoc]

They are trying to enter a market space Bitcoin miners perfected 5 years ago. Give them time, they surely will get "AI" folks served at some point when the miners already have their datacenters up and running for the past year.

Bitcoin miners actually do share a lot, but not enough. Now if you had those "Asics" like the antminer x3 that you could flash (it's actually just 3 SG2042s) that could be pretty neat.

[jeisom]

Seems a bit much for the Nvidia model. Napkin math of $950 for an XTX and the $10,000 difference in price has the 4090s at about $2600 each. Newegg has them starting at $1,800 right now. Obviously it isn't that simple, but still.

[Quackdoc]

Seems a bit much for the Nvidia model. Napkin math of $950 for an XTX and the $10,000 difference in price has the 4090s at about $2600 each. Newegg has them starting at $1,800 right now. Obviously it isn't that simple, but still.

tinycorp has always been honest about selling computers for more then they make at least

[loganj]

driver quality | mediocre | great

[theriddick]

Probably the Mediocre rating is because AMD isn't the best for workstation work, no CUDA and slower performance in ML/AI stuff.

[stargeizer]

So many things wrong:
1) They sell a computer they call "tinybox", which is 12U in height, or 21 inches, hardly tiny.

Compared with the usual requirements for "BIG" AI applications, is tiny. Real tiny. ChatGPT will never run on it, but is not designed for these use cases, but for smaller use cases. Basically, they are trying to "Comoditize". At least is known that can run Stable Diffusion .

2) Their website looks like it was designed by a 10 year old back in the early 90's. I tend to favor minimalism, but this is ridiculous.

Hardware people doing software. Like AMD, usually Another example: https://www.mov-axbx.com/wopr/wopr_concept.html

3) They list 144gb vram, how do they get to that? 6 cards that have 24gb vram each? If so, then why 128gb system ram?

Nowadays, modern GPU doesn't need ram to load data, they can just fetch from the SSD/hard drives directly. Both 7900XTX and RTX4090 can do GPUdirect Storage. The main difference is that the NVIDIA implementation is not buggy at level of firmware, like AMD is. (Again, hardware people doing software, rather than software people directing the hardware people)

I can't imagine someone spending $25 to buy something from them, much less $25,000.

No comment here, since i don't really care about IA, but the circus here is too fun to let it pass...

The make System76 look good in comparison.

Both look good.

[pong]

Yeah that's what I was thinking, too, since 128G is already what I'd spec. for a nominal last-generation (DDR4) general purpose "prosumer" desktop "workstation", and it's easy to want 2x or 4x that.

OTOH IDK how the memory channels are done on this device, I can imagine it has 4-channels and one 32GBy DIMM per channel or something like that. Anyway the point (if any) being I wonder if to expand it to 256GBy they'd have to use
slower RAM speeds or something like that based on the available RAM and the CPU RAM channel capabilities.
It might be the case that there could be some performance hit which may or may not be worth the doubling (or more) of the RAM available.

Of course another consideration is that if "serious users" need more RAM then it's at least a legitimate question as to whether many such users should really just buy ANOTHER entire box to double everything and have a little "cluster" to divide work and compute over. Though the small RAM to VRAM ratio still does seem non-ideal, I'd assume having at least 1.25x the amount of VRAM as RAM would be pretty minimal if one wanted to be able to run the app / OS and also mirror the GPU data on the host RAM if needing to swap in / out new data to the GPUs etc. somewhat efficiently.

128GB RAM is very underwhelming for a system like this. These days that's like the minimum for a data science workstation with just 1 maybe 2 GPUs. The fact that they list more GPU RAM basically makes this point. You have to load models into CPU memory first before you can transfer it to GPU memory.

[pong]

So many things wrong:

1) They sell a computer they call "tinybox", which is 12U in height, or 21 inches, hardly tiny.

2) Their website looks like it was designed by a 10 year old back in the early 90's. I tend to favor minimalism, but this is ridiculous.

3) They list 144gb vram, how do they get to that? 6 cards that have 24gb vram each? If so, then why 128gb system ram?

I can't imagine someone spending $25 to buy something from them, much less $25,000.

The make System76 look good in comparison.

That's a bit harsh. I don't know I mean 6 top tier "consumer" GPUs plus a respectable basic server motherboard plus dual 1600W PSUs plus cooling and making things maintainable etc. seems reasonably appropriate to stick in 12U.
6U would not be that "roomy" compared to a consumer full sized ATX chassis's side to side width.

The 12U one might run quieter or be easier to maintain or whatever and if this is going to applications where they're not just ordering whole racks of servers by the job lot but rather one, two, or a few tiny boxes in a small office / server room I'm not sure the size matters much.