GOD BOX 2.0

[venturi]

So I received the modDIY power cables for my 4090s.

configured with 600w full power each, the corner / bend of the cable happens easily just from gravity. Cables (without sleeving) feel (compared to sleeving) light as hair and just as flexible. No stress on the 4090 sockets.

I was going to get the right angle adapter from CableMod, but now that the cables are here from nodDIY I no longer see the need for the right angle adapter from Cablemod.

Here are some pics, seriously, cables just fall into place like a small pony tail. Made cable management easier in a SFF case:

 

[Chicken Patty]

I have the cable for my 3090Ti from them and it's fantastic. Good choice.

 

[AnotherReader]

Amazing build! After the CPUs, the RAM is the most expensive part. Why did you opt for so much RAM? Are you using the CPUs for Training as well? How hot do the 4090s get when you run something like Blender or Folding at home on them? Do you intend to replace your CPUs with a single upcoming EPYC? If you don't mind sharing, what is your day job?

 

[qubit]

Well, it's great that they can be used together in other applications.

@W1zzard any chance you could write an article showing what two 4090s in this kludged "SLI" can do in games?

Just a few benchies for this I believe would be sufficient.

 

[venturi]

Amazing build! After the CPUs, the RAM is the most expensive part. Why did you opt for so much RAM? Are you using the CPUs for Training as well? How hot do the 4090s get when you run something like Blender or Folding at home on them? Do you intend to replace your CPUs with a single upcoming EPYC? If you don't mind sharing, what is your day job?

it d does get used for gaming, but that's the smallest part of its function, I use iit for deep learning and part of my dissertation, and believe it or not dual 4090s, 1.5TB, and 112 threads gets used up almost immediately on some algorithms.

I explored the epyc motherboards and unfortunately, so farm there aren't any suitable choices, so far.

.my day job?.... if it was only that simple....
Full time doctoral student, two full time jobs in healthcare, owner nuclear pharmacy, owner IT company, part time physicist, DoD contractor, etc.

lucky me, there are 36 hours in a day....

 

[claes]

Ngl, the thing I dislike most about TPU is how often people post without reading the thread. @venturi posted what he does with two 490’s a number of times, and people are still asking as if the information isn’t in front of them. Even have people speculating that dual-GPUs can be used liked dual-CPUs when the OP clearly shows a dual-CPU setup…

Congrats on the build @venturi, I’m still impressed with the open chassis results, let alone your work hours, cleaning regimen, and having a family to boot. Hope your dissertation comes together as you hope

 

[Lei]

@venturi
I wonder which part of Nvidia architecture is utilised while deep learning. Is it RT cores, tensor cores or CUDA cores. And why doesn't Nvidia make cards that are designed for AI only? Using all the die size for machine learning....

Why wouldn't you buy A6000 for example?

If you had 2xA6000 cards, you'd be having 96gb vram. Since you have 2x3090ti and 2x4090 sitting there. That would use a lot less power and is way thinner.

Would you please tell me how faster each epoch gets trained on 4090 comparing to 3090ti?

Either get A6000 or watercool your 4090. In your box, one is blowing hot air into the other. Perhaps you're throttling at least 10% of performance due to heat even if you bear the noise and dust. Liquid cooling those stuff would cost less than 10% of the money you got them for.

 

[venturi]

@venturi
I wonder which part of Nvidia architecture is utilised while deep learning. Is it RT cores, tensor cores or CUDA cores. And why doesn't Nvidia make cards that are designed for AI only? Using all the die size for machine learning....

Why wouldn't you buy A6000 for example?

If you had 2xA6000 cards, you'd be having 96gb vram. Since you have 2x3090ti and 2x4090 sitting there. That would use a lot less power and is way thinner.

Would you please tell me how faster each epoch gets trained on 4090 comparing to 3090ti?

Either get A6000 or watercool your 4090. In your box, one is blowing hot air into the other. Perhaps you're throttling at least 10% of performance due to heat even if you bear the noise and dust. Liquid cooling those stuff would cost less than 10% of the money you got them for.

Well

there a lot of questions in there

so, my goal is to keep the build compact and quiet and able to run on a single household socket.

so while tempting to put in 4 Mai water cooled cards, the power (and space for the coolers) would lose the finesse

secondly, I would like the PC to do more than just my dissertation and just be normal PC in terms of turn it on do email, word docs, excel, remote in to work, surf the net and play games.

third, when I need more horsepower I group the wifes’s PC into a node and while not as efficient, I get it done - that unit is just 10-12 feet away (see pic) with a similar set up

As for the 6000 series cards, - I have one of those builds at work (6 cards in one rack unit) and the 6000 for the money, isn’t justifiable an expense

Last, I need a fast on board data set, so that is why I have the 50Tb array and the 50 GB/s write speed, all in the same PC

yes, the a 6000 is a good choice, but price point vs performance was not there, I pay for all my hardware.

 

[phill]

I suddenly feel like I'm not quite as overkill as I thought I was...

Amazing setups and thread, a lot of love and appreciation here from me!! Outstanding!!

 

[P4-630]

What brand/model of desk chairs are those?

 

[caroline!]

Sort of looks like my room, I like the white walls and the light fixture. Keep it simple.

One thing I'd never have though is a glass desk, one punch when playing Touhou and my whole setup is gone. Plus I work on my desk so there's always tools, circuit boards, wires and parts on top of it. I'm messy.

 

[venturi]

Hi and thank you

those are “iron horse 12 hour chairs”. They are closer to a car seat than a chair

the desks glass is tempered 2cm glass, also know as skyscraper glass, I doubt it would break under “normal” conditions

 

[Lei]

Hi and thank you

those are “iron horse 12 hour chairs”. They are closer to a car seat than a chair

the desks glass is tempered 2cm glass, also know as skyscraper glass, I doubt it would break under “normal” conditions

Show us some pics of radiology images before and after cleaning up.

Here I auto annotated some orchid flowers:

 

[venturi]

Wrong quote resending

 

[P4-630]

Nice taxonomy, I assume it is for the training set or its post validation?
I assume its at pixel level, colour algorithm?

As to return your query, here is an example of anonymized pre image set with multiple issues (motion, ringing, shadows, echo, Gibson etc) and acquired at 55% of a normal exam time length (resulting in less data) in a 3T MRI.
Its not about the sharpening of the image, its about removing anomalies and artifacts
DLSNR and DLR post algorithms (including K space) were applied for the post image to return the image back to SOC with some improvements without window and level:



PRE
View attachment 270715

and POST
View attachment 270716

I hope that demonstrates better or helps explain

Guess you quoted the wrong person....

 

[venturi]

I did... thanks for catching it

 

[Wirko]

That's a great demonstration, thanks.

My untrained eye actually sees some blurring, not sharpening, in the processed image. Also, the lone white dot in the left-middle of the brain is completely gone. Is there a possibility that the processing actually remove an important detail? Do doctors need access to unprocessed (or less processed) images too in order to see the full picture?

 

[venturi]

That's a great demonstration, thanks.

My untrained eye actually sees some blurring, not sharpening, in the processed image. Also, the lone white dot in the left-middle of the brain is completely gone. Is there a possibility that the processing actually remove an important detail? Do doctors need access to unprocessed (or less processed) images too in order to see the full picture?

A lot of the visible data in the original image is composting from echo shadow overlap moments that enhance on overlap. The image is showing more structures in the postprocess rendering
there is a second step of sharpening, window and level contrast that is performed automatically or by the radiologist, -the POST image has not undergone that process yet (raw), I was just giving an example of the post process. Additionally, the original image is 9% to bright /contrast haze. the post process image may seem dark, but it would normally be viewed on a PACS workstation in a dark room

 

[venturi]

here are the Gravitymark benchmarks to this build.

GravityMark Leaderboard

GravityMark GPU Benchmark

gravitymark.tellusim.com

you can check out the 96% dual GPU load and the corresponding temps

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

and 4k here:

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

 

[Lei]

That's a great demonstration, thanks.

My untrained eye actually sees some blurring, not sharpening, in the processed image. Also, the lone white dot in the left-middle of the brain is completely gone. Is there a possibility that the processing actually remove an important detail? Do doctors need access to unprocessed (or less processed) images too in order to see the full picture?

Excuse me, what image are you referring to? I don't see anything

 

[venturi]

I took the image down because it was a DICOM format and a some folks were getting errors on viewing.

 

[Wirko]

here are the Gravitymark benchmarks to this build.

GravityMark Leaderboard

GravityMark GPU Benchmark

gravitymark.tellusim.com

you can check out the 96% dual GPU load and the corresponding temps

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

and 4k here:

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

I understand you built the God Box 2.0 for your research but would it be fast enough to be usable in a clinical setting?

Or maybe I should ask a more basic question first - how fast would be fast enough? It does post-processing, so there's no need for real time processing, but if it's capable of processing those 1300 images taken during a 30-minute exam in 30 minutes, would that meet the needs of radiologists?

 

[venturi]

I understand you built the God Box 2.0 for your research but would it be fast enough to be usable in a clinical setting?

Or maybe I should ask a more basic question first - how fast would be fast enough? It does post-processing, so there's no need for real time processing, but if it's capable of processing those 1300 images taken during a 30-minute exam in 30 minutes, would that meet the needs of radiologists?

It processes the images for a whole series in less than 1/100th of what the next closest workstation from GE, Canon or Siemens can do, so yea, if anything, complete overkill

4.34 seconds for a full brain mri w/o contrast with DL-SNR and DLR from CNN