|
shrike82 posted:yeah going to disagree with the walls of texts - LLMs have entrenched the dominance of Nvidia hardware if anything. there are a lot of toy repos out there that cut down LLMs to fit and be compatible with CPUs including Apple Silicon but it's not really useful beyond students and execs wanting something to run locally on their laptops. a) it's lossy and b) why try to penny pinch on on-prem or private cloud hardware - businesses are happy to throw money at racks of nvidia cards to demonstrate having an AI strategy Eh, I don’t know. We’re still relatively early in the development of machine learning and it’s hard to say where things are going. Nvidia has the best support and most developed software ecosystem for sure, but ultimately most DL algorithms just need to do as many matrix multiplications as possible. A simpler architecture without all the GPU baggage designed solely to feed billions of MADDs could end up being the most cost effective approach as models continue to grow. Plenty of companies are experimenting with such designs. I wouldn’t be surprised if we see a bunch more competing products, as Alphabet, Amazon, Meta, Microsoft, and others develop in house, single purpose hardware that is cheaper to rent if you’re already in their cloud ecosystem.
|
# ? Aug 10, 2023 06:58 |
|
|
# ? Jun 3, 2024 21:26 |
|
shrike82 posted:yeah going to disagree with the walls of texts - LLMs have entrenched the dominance of Nvidia hardware if anything. there are a lot of toy repos out there that cut down LLMs to fit and be compatible with CPUs including Apple Silicon but it's not really useful beyond students and execs wanting something to run locally on their laptops. a) it's lossy and b) why try to penny pinch on on-prem or private cloud hardware - businesses are happy to throw money at racks of nvidia cards to demonstrate having an AI strategy ...by people talking about using macs for llms, I thought it was pretty clear for local, personal usage. That is also where the action is since the Facebook leak and due to lora. Nvidia hardware is entrenched, but there is momentum going against CUDA to platform agnostic approaches, which is what we were talking about. And I very, very much doubt OpenAI, Facebook, Microsoft et al. are happy being beholden to nvidia. Nvidia's market dominance is bad for them. I also mention this above.
|
# ? Aug 10, 2023 07:37 |
|
Yeah that's why they shouldn't use GDDR7, AI bros can get hosed
|
# ? Aug 10, 2023 08:21 |
|
I'm sure this is a dumb question, but why do gpus have embedded memory anyway? Couldn't they just have like.. ram slots or whatever on them since they're getting so big? Or is it that the memory has to be so fast for a GPU that either it's not practical for a consumer to buy or it requires a special design to accommodate?
|
# ? Aug 10, 2023 12:34 |
|
Weird Pumpkin posted:I'm sure this is a dumb question, but why do gpus have embedded memory anyway? Yes. Making slotted VRAM would increase the trace lengths/distance to the core, which would be detrimental to signal integrity and make it harder to reach the speeds that VRAM is aiming for, and there wouldn't be that many advantages to it.
|
# ? Aug 10, 2023 12:38 |
|
gradenko_2000 posted:Yes. Making slotted VRAM would increase the trace lengths/distance to the core, which would be detrimental to signal integrity and make it harder to reach the speeds that VRAM is aiming for, and there wouldn't be that many advantages to it. Ah ok, thanks! I figured it would be something like that Kind of a shame though because it would be nice to be able to just pack as much as you'd like, or upgrade it later as requirements get more ridiculous
|
# ? Aug 10, 2023 12:40 |
|
used to be able to jam ram chips into gpus back before 3d acceleration existed as a thing, so you could run a higher resolution desktop gui lol e: pic not mine but i had one of these packed full of those chips, coupled with a voodoo1 in like 2001 lol Truga fucked around with this message at 12:47 on Aug 10, 2023 |
# ? Aug 10, 2023 12:40 |
|
The earliest video cards actually did have RAM slots, but they very quickly moved away from that. Lots of random expansion cards had socketed memory or cache actually, but everyone decided it was best to just solder as much memory as you need for performance and guaranteed compatibility reasons. Hell, even CPU cache used to be slotted into your motherboard, but then they started putting it directly on the cpu die instead, the benefits of which should be obvious. That's a really good example actually because the reasons for using soldered memory are similar to the reasons for moving to on-die cache. Really, everyone's trying to move all forms of memory as close to the processing units as feasible. The highest-performance GPUs put HBM memory on the same package as the GPU, for instance. They'd probably put it all in one big chip together if such a thing were possible. GDDR exists in the middle ground where it's soldered onto the board near the package.
|
# ? Aug 10, 2023 12:47 |
|
Dr. Video Games 0031 posted:Hell, even CPU cache used to be slotted into your motherboard
|
# ? Aug 10, 2023 12:54 |
|
I forget if the MMX/K6 era had cache on the CPU but motherboards still had additional cache slots to augment it. I still think those are cool.
|
# ? Aug 10, 2023 13:02 |
|
Probably won't be long until CPU RAM gets soldered in too.
|
# ? Aug 10, 2023 14:08 |
|
The one true way is slot CPUs with cache slots on them. I will not be taking questions at this time.
|
# ? Aug 10, 2023 14:22 |
|
Nfcknblvbl posted:Probably won't be long until CPU RAM gets soldered in too. that's already mostly the case in laptops, as Dr. Video Games 0031 alluded to
|
# ? Aug 10, 2023 14:24 |
|
and consoles, and macs it's only a matter of time before desktops get it, perhaps with an intermediate step where motherboards still have RAM slots but you can opt for a CPU with embedded memory and leave the slots empty if you want to
|
# ? Aug 10, 2023 14:25 |
|
repiv posted:and consoles, and macs Yeah, this is what I meant. Upgrading RAM is becoming less of a thing. A meme will truly die.
|
# ? Aug 10, 2023 14:27 |
|
That sounds like it's going to get expensive for the home market. Since then RAM will be a market segmentation tool in both GPUs and CPUs i spose
|
# ? Aug 10, 2023 14:29 |
|
they can take away our ram slots but they can never take away https://downloadmoreram.com
|
# ? Aug 10, 2023 14:31 |
|
repiv posted:and consoles, and macs Stack everything. A compute die on top of a cache die on top of a PHY die on top of a HBM die. Let the 3D revolution begin!
|
# ? Aug 10, 2023 14:40 |
|
ConanTheLibrarian posted:Stack everything. A compute die on top of a cache die on top of a PHY die on top of a HBM die. Let the 3D revolution begin! This is the way
|
# ? Aug 10, 2023 15:22 |
|
Good luck with cooling that stacked die! Chips will have tiny cooling channels between each stack.
|
# ? Aug 10, 2023 15:53 |
|
Nfcknblvbl posted:Probably won't be long until CPU RAM gets soldered in too. no way ram getting soldered survives in the server market, and desktops are really just an extension of that
|
# ? Aug 10, 2023 16:08 |
|
Truga posted:no way ram getting soldered survives in the server market, and desktops are really just an extension of that Performance gains have to come from somewhere.
|
# ? Aug 10, 2023 16:21 |
|
And hell, it'll possibly further reduce replacement cycles, so they'll be making even more money off the customers!
|
# ? Aug 10, 2023 16:27 |
|
Yo Dawg, I heard you like stacked dies, so I put a stacked die in your stacked die, so you can stack dies while you die.
|
# ? Aug 10, 2023 16:30 |
|
Truga posted:no way ram getting soldered survives in the server market, and desktops are really just an extension of that I don't see why soldered RAM in the server market couldn't work, especially with the way that there's building momentum for single-socket servers in some type of higher density configuration over the traditional 2P configuration.
|
# ? Aug 10, 2023 16:43 |
Nfcknblvbl posted:Performance gains have to come from somewhere. Yeah, make software efficient again instead of throwing everything inside electron and brute forcing passable performance by throwing hardware at the problem.
|
|
# ? Aug 10, 2023 16:50 |
|
Arrath posted:Yeah, make software efficient again instead of throwing everything inside electron and brute forcing passable performance by throwing hardware at the problem. why not do both things
|
# ? Aug 10, 2023 16:54 |
|
We’ll just use AI software optimization that creates an incoherent codebase that’s “good enough” with uncorrectable bugs at the extreme margins Porbelm solued
|
# ? Aug 10, 2023 17:15 |
|
just lmao if we end up with dumbed-down CPUs to avoid side channel attacks and make it up by going back to VLIW ISAs supported by AI-enhanced optimization in the compiler "where did that bring you? back to me" - Itanium
|
# ? Aug 10, 2023 19:23 |
|
Nfcknblvbl posted:Good luck with cooling that stacked die! Well: https://www.froresystems.com/
|
# ? Aug 10, 2023 19:33 |
|
Well dang, the future is now.
|
# ? Aug 10, 2023 20:12 |
|
Their best model removes 10w of heat... if they jam multiple into some Macbooks that'll be nice, but I don't think they can scale them to laptops with discrete GPUs or anything.
|
# ? Aug 10, 2023 20:27 |
|
Isn't the big asterisk that the cooling module itself will be power hungry, or am I thinking of something else? E: Which I guess wouldn't be an issue on desktops
|
# ? Aug 10, 2023 20:55 |
|
Rinkles posted:Isn't the big asterisk that the cooling module itself will be power hungry, or am I thinking of something else? Their home page states that each module eats 1.75w to cool 5.75(mini) or 10.5 watts of heat. It's only going to cool ultraportables and nucs.
|
# ? Aug 10, 2023 21:08 |
|
1.75 watts to dissipate 10.5 watts is not terribly efficient when compared to a typical laptop fan, but as you say this is more about the form factor such a solution allows more than anything. In larger form factor systems, I don't expect it to offer many advantages. Maybe you can slap some on stuff like SSD controllers or motherboard chipsets so you can give them compact cooling solutions without annoying, whiny fans. edit: The 10.5 watts includes the 1.75 from the airjet itself. The mini dissipates 4.25W of heat from a chip with 1W of power and the pro dissipates 8.75W of heat from a chip with 1.75W of power according to their spec sheets. edit 2: This is mostly tangential but I was curious so I looked into this more. This thing's power efficiency is about 5:1 for watts dissipated to watts used. Going by the max power draw of laptop fans, I'd guess the typical blower-style laptop cooling solution is more like 10:1 to 20:1. And if we were to look at PC fans, the NF-A12x25 is 1.68W at max speed, and on the right heatsink it will be capable of dissipating 200W of heat or more, or about 119:1. The main draw of this device will be its form factor and low noise profile, but the low power efficiency does limit it somewhat in devices where battery life is important, I feel like. Dr. Video Games 0031 fucked around with this message at 00:55 on Aug 11, 2023 |
# ? Aug 10, 2023 21:19 |
|
LTT had a puff piece (PUN INTENDED) on it a few months ago, it does seem pretty cool (PUN INTENDED): https://www.youtube.com/watch?v=vdD0yMS40a0
|
# ? Aug 10, 2023 21:21 |
|
I dunno, the design seems to be full of holes. Their probably just blowing hot air.
|
# ? Aug 10, 2023 21:30 |
|
Dr. Video Games 0031 posted:Maybe you can slap some on stuff like SSD controllers or motherboard chipsets so you can give them compact cooling solutions without annoying, whiny fans. This was just an idle thought in my head, but what do we have here? https://www.tomshardware.com/news/phison-demos-ps5026-e26-max14um-gen5-ssd-with-a-14-gbs-read-speed It makes sense. SSD controllers are getting out of hand with their cooling requirements, and the tiny fans they're putting on the heatsinks for Gen 5 drives are the worst. This seems like prime territory for AirJet cooling.
|
# ? Aug 10, 2023 22:09 |
|
Cygni posted:LTT had a puff piece (PUN INTENDED) on it a few months ago, it does seem pretty cool (PUN INTENDED): Wow, thanks for recapping Thermodynamics 101, Linus!
|
# ? Aug 10, 2023 22:18 |
|
|
# ? Jun 3, 2024 21:26 |
|
I like this. https://youtu.be/_1KjsqaWIU4
|
# ? Aug 10, 2023 23:19 |