Видео

@@kefferaberin1437 In NG+1 after passing the game once I decided to NG... then realized I lost my whole map, so worked my way to the palace and maxed out the last boss sword and farmed as a hobby to max out. I was planning to get back into the game after school, but hasn't happened yet...

I took a look at the S300-Mini-ITX case and the image they show on Amazon is weird where the cable doesn't look quite right. I think they edited the image or something and not accurate. I took a look at a couple videos and see the kind of cable that is being used, and it looks like the double reverse cable that I used in my build. Check Amazon and even the linkup.one site and they sell the cables there for cheaper, think on newegg some people are jacking up the price of the cable, then order from one of these places when you order on newegg and send to you. Have to be careful with newegg and make sure it is newegg sending you the part. As for size if they recommend 20 cm adding 1 or 2 cm wont hurt you, and give you a little more play so you aren't pinching the cable as much. I think for me it was listed to use a 18 cm cable, the cable fit, but was tight, so next go around went with 19 cm and it was perfect.

Forgot to add some links for the AVA5 PCIE 5.0 Riser Cable for 21 cm and 22 cm - linkup.one/linkup-ava5-pcie-5-0-riser-cable-future-proof-for-gen-5-gpu-vertical-mount-x16-128gb-s-speed-compatible-with-pcie-4-0-designed-for-itx-double-reverse-ver-2-black-21cm/ - linkup.one/linkup-ava5-pcie-5-0-riser-cable-future-proof-for-gen-5-gpu-vertical-mount-x16-128gb-s-speed-compatible-with-pcie-4-0-designed-for-itx-double-reverse-ver-2-black-22cm/

I have a playlist for this build, many different parts and things I tried - ruclips.net/p/PLqL965J4xElLav8sEGE0hr6EsUhbjDfzP - I have another video in that list where I chose to use the Crucial Ballistix White DDR4 3600 16GB memory. This is a low profile memory kit and I also grabbed the Crucial Pro DDR4 3200MHz 64GB kit and may swap out soon because I found that 16GB at times is constraining while working on the computer. I like the white kit and if I was using this more for gaming instead of every day I think the kit works well and is speedy for my casual gaming. Crucial Ballistix White height is 39.17mm and Patriot Viper Steel height is 44.4mm.

Yes, the ROI if you can make money on this compared to an enterprise setup. For me this is implicitly making me money in that it helped me get my doctorate which should increase my income by way of Generative AI and Cybersecurity knowledge. Here is the playlist for my Hephaestus Build ruclips.net/p/PLqL965J4xElJek_JlCG60EOddZZykeb1y - I have another video - ruclips.net/video/r_48PaGLMnA/видео.html and this shows the CPU, motherboard, and memory. AMD Ryzen Threadripper Pro 5975WX, Asus Pro WS WRX80E-SAGE SE WIFI II, and 512GB of Micron 64GB DDR4-3200 RDIMM 2Rx4. The Optane is supposed to be great due to it's IOPS speed, so latency is really low making this a fast drive for repetitive read/writes. The durability is supposed to be insane with the Optane, but not sure only time will tell. Again, check out the playlist and as I work on stuff and have time I planned to make videos and updates. My rigs over time get upgrades as I mess with stuff.

The NVLink so far hasn't technically speed up inference speed because most times the LLM inference I've used will fit on one GPU. If you can run everything in one GPU the response is best, I haven't yet tried to see if I can make an inference span multiple GPUs. Something I may have to dig into as LLMs have been getting larger and larger. I took a quick look at Aphrodite and think there will be many frameworks like this to better scale on GPUs. I noticed I was able to span and leverage a mix of cards with two 3090s, two 3090 ti, and 3 4090s with PyTorch.

I tried the Noctua NH-L12 Ghost S1 Edition and in another video the Jonsbo HX6200D. With the case and setup really like the Noctua cooler better and it seems to be better overall while being very quiet.

Yeah, exchanged several emails with them. Haven't tried many more cables, but have bought seven more of the ADA5 LinkUp cables for my Hephaestus AI Build. The newer cables work well and very stable with multiple 4090s at x16. The 4060 itself has a limitation of only x8 so was really weird what I was experiencing. I can't wait for the new Blackwell 50 series, and hope to get my hands on one when they come out. Ha e to see if they will be as hard to get as the previous 30 and 40 series GPUs.

@@nispoe Sorry man, Hephaestus AI Build what mobo processor and RTX are you using?

@@DrMartinC33 oh, some other videos I uploaded for another rig. That machine is using a 32 core Threadripper Pro, has 7 GPUs and 3 of the GPUs are on the last slots. In that build I needed to also upgrade the riser cables. I was using older LinkUp PCIe 4.0 Ultra cables, but found the motherboard has re-driver chips to help boost the signals because of the extra length. The LinkUp ADA5 PCIe 5.0 cables work without having to boost the signal. The mobo is the Asus Pro WS WRX80E-SAGE SE WIFI II.

You have the right lines of thinking... actually I'm a bit slow and haven't updated my progress with new videos. I've actually figured out the problem and yes there is a part with signaling and power with the PCIe riser cables. I found that the last slots furthest from the CPU have signal re-driver chips to take signals and boost them so communications from the GPUs can reach the CPU without error. There is a line of 75W of power also that communicates to the GPUs and when a GPU doesn't have any power from a PSU the GPU won't function. No need for extra hardware, adjusting the signal through the BIOS works with the PCIe 4.0 cables I have. I got new PCIe 5.0 cables and found the extra shielding with the cable does the trick as well. There is some extra signaling feature of the PCIe 5.0 cables which adds communication signaling and I think better power control. I have run weeks and even months straight during my doctoral research and was able to successfully and with stability run with this setup. Let me work on getting a new video to show what I learned... just need to find a moment...

Short answer is 96GB, and not a "pool" as you think. I was thinking the NVLink would act as a way to pool VRAM between cards, but it only seems to help with processing time between data that can transfer between cards. I was messing with various frameworks while training and tried HuggingFace Accelerate, Meta PyTorch Distributed Data Parallel (DDP), and at the end used Low-Rank Adaptation (LoRA) which leverages I believe Pytorch DDP. What happened was that adding more cards enabled me to work with larger models because I could distribute load across the 4 GPUS with 24GB of VRAM, and later this would be the 96GB of VRAM for the four cards. Later I added the final 3 GPUs and able to run with a combined 168GB of VRAM. I was thinking (and maybe I didn't find the configuration for this) that the NVLink would give you 48GB with two cards in some way, and it does not. The NVLink helps to speed up any data exchanges between the two cards linked instead of routing through PCIe lanes to the CPU. To fine-tune larger models or to run a larger LLM instance, VRAM is the major constraint. I was ultimately able to train a model that was 140 billion parameters. I think I could go larger but only went so far. I was utilizing roughly 60 or 70% of the memory and ran some fine tuning that went for around 15 days. I also had deadlines for my doctoral research so had to cut myself from going too far. As a hobby and time willing I planned on digging into this more to get a deeper understanding. Planned on making more videos to show this, but life... it happens...

@@PeGy Wave of Gold it comes on the sword. The sword is the final sword you can get after beating the last boss in the game.

There may be synchronization issues between motherboard and GPUs, but what I have done it all seems to work fine. Just need to make sure to power on the GPUs before the motherboard is turned on. This may may the fans on 3090s spin at full speed initially. The 4090s seem to control this better. If you work with LN2 and overclocking GPUs this is something people do all the time. Also, if you have the PCIe cables attached and the GPUs not powered on, you'll see red lights on the power being supplied and nothing will happen with the motherboard and OS. It will not be seen as installed. I should make a video about this, there are some things to know about this that many probably don't know.

@@videocruzer yes, if you can run on one 4090 or 5090 and the VRAM is enough that is what you want to do to save time. I found that my constraints were more VRAM, so multiple cards for VRAM more important than GPU processing power. There are times I hit limits training purely because I didn't have enough VRAM.

@@nispoe I get the ramming thing, Me I dont plan on using the AI stuff until I can just punch in a whole script and or screen play as i have had way too much of my property stolen over the decades. I am guess that the software access has some super confusing end user agreement that will stiff the average user for their screen play at the end of the day.

@@videocruzer I think my situation is different, I used these GPUs to learn more about setting up the hardware and software for my doctorate. I had to run for months to get the results I needed to show results on my dissertation's hypotheses. I think the end user agreement is what you would expect for hardware and the software frameworks like CUDA and Meta/Facebook PyTorch, and etc. I think having cards locally and not having to pay the thousands for using a cloud GPU environment has the perks that I can continue to make mistakes and continue to use the hardware after the initial cost of buying the hardware. In one of my articles It sounded like one research team spend $25,000 for training, and who knows how much on mistakes. Sorry to hear you have had too much property stolen over the decades... maybe time to be somewhere else and have your hardware somewhere else?

@@nispoe Naa, I am descendant of Polish Grandparents that survived their Notzy Concentration camp experiences and British Grandmother that was a spy for the Allied side during the 2nd great war ( Gwenn Fussy was assaulted dropped from her overhead transfer lift 5 times in a row, almost every Saturday shift in a row until she was dead from her injuries), not to mention my Canadian Grandfather that survived the war against the Germans. Surviving in a City after being hit with a hard ball and stuffed into a deep freezer because My Polish Grandfather stood up to three of his neighbor's that were threating his family members with a bat about 1 year before the hit on his grandson happens just makes the story that much funnier. Being forced to live in one of the most corrupted cities in Canada, you end up connecting to the crowd that flush them out. It has been interesting to say the least. you know the saying, there are always survivors that live long enough to tell the tail. BTW one learns a valuable skills since mid 70's. Look for the Book, movie and video game deal Tagged. "Mares Leg" ;)

@videocruzer sounds interesting, my daughter is a quarter Polish. I think many families have some sort of history they have lived through in their life. I feel lucky having gone through my own hardships and able to tinkier with all this tech. Just recently had a new son, and never thought I would have another kid after my daughter. I want to try to give them what I never had, and not have them experience the bad things I have. Just like what I hope any parent would.

I was going to make a video to show this, but my research and deadlines approached fast. I found when using BERT I was able to get about 20% uplift using the NVLink. I wrote new code to work with llama and didn't perform this same test. PyTorch does some loading underweight the hoods, so not exactly sure if it helped greatly or not. When I have a moment will try and confirm, make a video also.

@@nispoe Thanks so much for the response! that 20% uplift with BERT is really interesting. I understand how research deadlines can be, so no worries! if you do get a chance to test it with Llama or any other models in the future, I'd love to hear about it. looking forward to that video whenever you have time to make it-I'm sure it'll b insightful. Best of luck with your research!

@@alirezashekari7674 I finished my research, now a doc. I did run Llama 2 and 3. I found that at times depending on what I was training the three 4090s were so much more powerful than the four 3090s.

Their instructions said they wanted them destroyed like this. They will not replace since they've decided they won't make these anymore. I think they had too many failures with 40 series cards that they were probably losing money since they advertised they would repair the video cards if something went wrong while using these adapters.

I tried this early on when setting up windows, but since then setup Linux. What I remember was not really thrilling results, think the idea is that a certain kind of workload would be needed to generate the right kind of IOPS results. Next time I can reformat this disk I'll probably try to really suss out it's performance and show/document.

@@nispoe remember the IOPS are going to be limited predominantly by two things. First if your interface is limiting; either by power over distance, or successful interaction with full bandwidth via the lanes. And second… if you meet the first requirement. The CPU itself will limit the IOPS. You know that your hard drive is amazing when the bottleneck of your ssd is the CPU. 😂

@@drewphilip The interface and distance was definitely something I noticed when using an adapter card and the motherboard has issues with signaling between slot pcie connections to the CPU. At slots 5, 6, 7 the motherboard even has re-driver chips to improve signaling. I chose the 32 core Threadripper PRO, and not sure if I'm limiting the IOPS, I need to run through some different benchmarking tools to try to push the IOPS to see if I can get to what is spec.

@@manuellopez22 this may be more practical, the idea is that you could take items and max out without maxing out key or all stats

All tarnished should max out elden ring characters, you have eternity...

whoohoo max cape ftw!

Glad the video helped, I found the AVA5 worked better on another rig I built. Haven't put a video out yet, but turns out on this other board there are redrivers on slots 5, 6, and 7. When I boost or adjust the signal the PCIe 4.0 cables work. But if I change out to the AVA PCIe 5.0 cables no need to boost the signal by adjusting the redriver chips.

@@nispoe is that a hedt board? I started on a gigabyte b550 ITX and use an Asus x570 ITX now, both seem to have the pcie signaling issues. Just glad to have found a solution

@@MrStoffzorYes, I have some videos of my Hephaestus build with ASUS Pro WS WRX80E-SAGE SE WiFi II AMD WRX80. This motherboard had the redrivers and I think some motherboards will have this problem if there isn't a redriver to continue or strengthen weak signals.

@@nispoe a redriver would be a cool addon module!

ha, that's a new one, think Musk would do something like this?

This card was such luck... it was maybe 3 or 4 am when I saw it online and in-stock at the Micro Center. I have a bad hobby collecting these kinds of things...

@@nispoe How much did you get for this one?

@@Oakyman $2,300 USD

Yes, I'm sure it is possible to get max level, and I also play on PC. TBH I sort of want to do this and should have done this not in NG. It's one regret I have after spending all the time on this. I did this after flipping into NG+1 and realizing everything I mapped was reset.

​@@nispoealright thanks bro. Then I will also farm like this. And try to get max level. Then I can also make a good build.

@@toonnut1 you know ... I was thinking about that, but it was late/early and I thought, "this'll get someone 80/90% of the way and that's at least a B/A and passing for sure.

@@nispoe yes its definitely good enough m8 👍

@@wSo1x NG+1

@@nispoe damn I'm on NG+3 and Im far from level 713 (I'm level 274)

@@wSo1x I was somewhere mid 250 when I passed the game, and then wondered how max level would be...

@@nispoe how many days it took you?

@@wSo1x around 511 hours or around 21 days. I did this slowly over the course of more than a year.

@@nonjadammit he probably has life all figured out, I tend to live life learning things the hard way... I'm old and still figuring out life...

These take what power they need when running and the way my AI jobs run will not max out on power most of the time. Yes, would probably love to have efficient data center or even workstation cards, but this is what I have and was far cheaper. I have a bad hobby of collecting hardware, so there's more to why I started with these cards and actually I do have an EVGA 3090 XC3 that is water cooled. I was initially going to start with some water cooled cards but due to time switched to air cooled for quick setup.

There's is a difference with fine-tuning versus running an inference. The largest I could tune was 140 billion parameters, could maybe do more, but stopped there for my research. I did try running larger models for inference and could do much larger, but again was focused on my dissertation, played with Falcon which I think was 180 billion parameters. The one thing to note about the NVLink is it only helps with sending data between the cards, I was not able to figure out how to pool memory between the cards. I do know that for example running 2 cards definitely helps run larger models, but the difference is the time savings when you NVLink them. So just having 48 GB you can run larger models. If you use quantization and other techniques then you can get into the 7 billion models. I should make some videos to show what I mean...

None that I've seen, I also have the 1000W version that's not white. Use it with my 4090 for AI work. Have run them both pretty continuous for months without issues. Have to see if over time this happens.

@@nispoe people who had problems had them emidiatly I think the PSU is fine Most of peoples problems are user error anyways

@@nispoe just a small update I finally got the white 850 Loki and what can I say it was the best decision I made so far It's absolutely silent no coil whine and the fan hardly ramps up or starts running with my 7800x3d and 4070 all maxed out And yes there is a time where the fans starts running and it's not the most silent but also I can't hear it over my arctic case and radiator fans The PSU looks absolutely gorgeous So who ever thinks to buy it just go for it don't listen to those 2-3 guys on the Internet who complain on everything It's a fantastic PSU Just a little bit expensive but still worth it

@@JJ-gk5tn nice, I use this PSU as part of this build as my daily driver, so far it's doing well

Have, not has. Does it have.

nvm on the pcie lanes, the cpu has 128 pcie lanes, so that's more then plenty lol

@@Jackel1337 all good suggestions, I did eventually figure out that on the motherboard PCIe slots 5, 6, and 7 all have redrivers. So if you populate all slots you need to go into the BIOS and and adjust this for better signaling. I later upgraded to better riser cables and found this helped a lot also. Lol, yeah I got this Treadrilper pro so that I can have full x16 usage on all slots.

@@blxsr it's called Wave of Gold and already on the sword Ancient Relic Sword

@@nispoe Alr thx 🙏

The ring is 20 mm in diameter. This fits on my index (pointing) or middle finger. Too large for my ring or pinky.

I had some heavy gaming sessions with this rig playing Helldivers 2 and Diablo IV. I also just ran the 3DMark Steel Nomad Light benchmark. The system stays pretty quiet with the Noctua fan, but I do hear the video card spin up and make more noise. Since the 4060 has those 3 small fans the pitch is higher when gaming, but not overpowering and do hear after it spins down when the inside of the case becomes warmer. What I noticed is that in SFF cases the heat builds up and wants to escape upwards or out the side. Have you tried to take the side or top panel off to see if the PSU runs cooler and doesn't spin up? My thinking is if you have a really beefy GPU this may heat up the case, and since a longer GPU will radiate heat on the side of the PSU this would make the PSU run hot needing the fans to kick in. If you look at the other video with this build and the 4060, the back of the GPU is radiating most heat around the back of the motherboard since it's a shorter card. Unfortunately, the way longer GPUs are designed they are trying to draw the heat away from the GPU chip and towards the end of the GPU (which is where the PSU is).

So does this mean u dont have the Same Problem? Im asking cuz i realy wanna buy this PSU but Dome people claiming there ist a Problem with the Fan So is there a Problem or Not and IS the PSU silent ??

​@@QQq-h6bI don't have the same problem. I think this is because I'm not taxing the PSU as much. In my setup my wattage usage maybe much lower so the PSU fan doesn't trigger much. My SFF CPU used 65W, GPU max could be 75+150W, and mono/SSD maybe 100W. The GPU has a max of 850W and for best efficiency you try to be somewhere in the middle so around 400ish watts.

@@Dangz unfortunately the heatsink has a cutout on one side for the RAM, and the cutout is low so only lower profile RAM will work.

@@nispoe buts it`s not possible to turn around this cutout to motherboard shield? the other side hit the ram?

I tried that also and it doesn't work. It was made to fit only a certain way for this motherboard. I think the IO side was too tall, maybe if I remove the shroud, but at that point the low profile memory made more sense.

@@nispoe so sad bro, but thanks for your response, becausei i have a id cooling 67 xt, i think dont have this problem at this cooler

@@Dangz I checked out your cooler and see how if you turn it you can make the heat pipes face the ram and the fins clear the I/O shroud. Makes sense your ID-COOLING IS-67-XT clears and keeps low at 67mm.

@@luiscasillas9853 depends, if you do as I did a little at a time would take more than a year

Thanks, I used it for AI work and for my doctoral research. Could have used cloud services, but this was more fun.

@@nispoe haha definitely mate 👍

Next time I mess with it I'll keep this in mind to show you how it is. The Cobolt has no issues running x16 and PCIe 4.0, very good cable but a bit messy given the individual wires.

@@Ghamba261 lol, yeah... faced the rune counter cap

I'm at NG+1 and just using the Golden Scarab item. Each run is 96,151 runes and can be more if you are lucky and have a glowy eye mob which gives you an additional 29,433 runes just for one kill. Each run took roughly 20 seconds.

Yeah, it is somewhat of a pain but if you've done it enough times it's good experience and good to know. I have some instructions I put on GitHub and some lessons learned are in my research paper. This is the GitHub location github.com/nispoe/kuk-praxis-hephaestus/blob/main/ai-machine-setup.md and here is my published dissertation scholar.google.com/scholar?hl=en&as_sdt=0%2C14&q=Detecting+Machine-Generated+News+Using+Fine-Tuned+Transformers&btnG=

@@nispoe Awesome, thank you! Much appreciated!

@@datascienceharp I have other videos with the machine completed, but haven't posted yet. After building I had to get started into running my experiments. Now that I'm done I hope to work on a video to explain what I had to do to add 7 cards and stuff I learned about the PCIe redrivers on the last slots because they are farther from the CPU.

@@nispoe looking forward to those!

@@woodstockings6545 Should be, I had a plan to play through and explore after maxing out and having more time, would be neat to try and see how I match up at this level

@@c4n0 ha, I earned every rune, no save editor

Anyone who does this straight is more dedicated than I. They deserve to be more tarnished than other tarnished.

@@lms3078 I think I was around 250 when I beat the game the first time. I was working on my doctorate and haven't even touched the DLC yet, but very soon, trying to regain my life.

@@samarrathore3716 I'm using the scarab item that gives extra runes. I'm also NG+1 and think that gives some extra runes as well.

@@nispoe ah can you name the item please so I can just look it up

@@nispoe also one last question will I haven't started the dlc will it be affected with the ng+ even if I haven't started it?

@@samarrathore3716 gprivate.com/6cqlz

@@samarrathore3716 I haven't started the DLC either, I think the NG+ will just up the difficulty. I heard the DLC is challenging, and the final boss very hard.

I bought a Hydro Copper Kingpin in great shape, and think people avoid because of all the waterworks. I sort of collected these over time slowly, and oddly people weren't buying the 3090 Ti on eBay when I was looking. Just have to check for the model number and make sure what you are getting is the correct model. I once bought a 3090 thinking it was a Ti and it wasn't so returned it.

@@jayvongibbs3750 check the doobly doo :D