The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Rate my PC build for large model local LLM
Poast new message in this thread
Date: December 11th, 2024 1:25 AM Author: Costumed milky stain AMD Ryzen 9 9950X GPU1: RTX 4090 GPU2: RTX A6000 128gb RAM 8TB NVMe 4.0 Windows 11 Pro |
 |
Date: December 11th, 2024 2:03 AM Author: Sepia stag film boltzmann why not two 4090 (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48435788) |
|
Date: December 11th, 2024 2:12 AM Author: Costumed milky stain My combo is more optimal for dual GPU since you can't onboard dual 4090s on 1 motherboard and eGPU throttles 50-70% of 4090 performance due to inherent bandwidth limitations with thunderbolt connection vs A6000 is optimized for low bandwidth and has 48gb VRAM (vs 24gb for 4090) to handle massive data sets
|
|
Date: December 11th, 2024 3:34 AM Author: Sepia stag film boltzmann the CPU absolutely does nothing for LLM? (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48435877) |
|
Date: December 11th, 2024 5:14 AM Author: Costumed milky stain CPU matters. That's why I got the best one (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48435907) |
|
Date: December 11th, 2024 8:22 AM Author: cerebral pearly double fault There's something called CPU offloading where some processing is handled by the CPU. But the kind of operations used in LLMs is different than the kind of operations at which CPUs excel. (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48436129) |
|
Date: December 11th, 2024 1:36 PM Author: Costumed milky stain Do you think Apple's future M4 max/ultra cpu will be better at LLM? (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48437276) |
|
Date: December 11th, 2024 2:57 PM Author: cerebral pearly double fault no (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48437673) |
|
Date: January 8th, 2025 9:20 PM Author: canary deep sweet tailpipe striped hyena M1 Ultra is still plenty (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48533861) |
Date: December 11th, 2024 1:34 AM Author: Jet Galvanic Institution i have similar but just for poasting (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48435773) |
|
Date: December 11th, 2024 2:14 AM Author: Costumed milky stain (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48435795) |
|
Date: December 11th, 2024 3:05 PM Author: Yapping Parlour (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48437693) |
|
Date: December 11th, 2024 9:07 PM Author: lascivious carmine area (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48438902) |
|
Date: December 11th, 2024 11:24 AM Author: dun volcanic crater center Where are you finding 4090s for sale that aren't jacked up 2x MSRP by scalpers? (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48436665) |
|
Date: December 11th, 2024 1:33 PM Author: Costumed milky stain I got it months ago and just putting together now.
Are prices on 4090 still that bad? I thought 5090 and 5080 were about to drop any day now |
|
Date: December 11th, 2024 1:34 PM Author: mustard territorial messiness business firm For work? (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48437264) |
|
Date: December 11th, 2024 8:56 PM Author: Costumed milky stain Yep (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48438874) |
|
Date: December 11th, 2024 1:35 PM Author: Hyperventilating idiotic sound barrier Is setting up a local LLM pretty easy? Does it need coding savvy? (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48437271) |
|
Date: December 11th, 2024 1:36 PM Author: mustard territorial messiness business firm really easy with ollama, https://ollama.com/ (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48437274) |
|
Date: December 11th, 2024 1:38 PM Author: Yapping Parlour (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48437286) |
|
Date: December 11th, 2024 3:40 PM Author: Seedy copper factory reset button whats the benefit of running these locally vs just subscribing to claude or whatever
|
|
Date: December 11th, 2024 3:45 PM Author: mustard territorial messiness business firm the ones running locally are not as powerful and don't accept really long context windows. the main benefit is no limits to the number of requests and you can find uncensored models. (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48437832) |
|
Date: December 11th, 2024 3:55 PM Author: Seedy copper factory reset button does the hardware above make it practical to train your own? could you train on just on a giant dump of legal cases and get one superior at legal research at the expense of everything else, or is that even available better commercially at this point?
|
|
Date: December 11th, 2024 4:38 PM Author: Costumed milky stain 1) Yes the above hardware can easily handle large models and is possibly even overkill for CRM, basic legal queries / prompts about client's file, and drafting complex legal briefs
2) Local llm with win10 Pro (eg, ollama, GPT4All, chat rtx) is the ONLY option for attorneys to do this offline given client data sensitivity
3) if you spent $10k on a fully spec'd out Mac Pro/Studio with 192gb unified ram it would not come close to this setup for local LLM in terms of speed and ability to handle large data and complex writing tasks |
|
Date: December 11th, 2024 5:23 PM Author: mustard territorial messiness business firm 0 chance you will be able to train your own LLM, these companies spend tens of millions $$$ on gpu hardware to train an LLM.
The best you can do locally is "retrieval augmented generation" that is specific to legal cases/research. |
|
Date: December 11th, 2024 1:37 PM Author: Aphrodisiac haunted graveyard lay You don’t need this much power just to take notes/exams at NYU Law tax classes. (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48437283) |
|
Date: December 11th, 2024 8:48 PM Author: Costumed milky stain 163 (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48438847) |
|
Date: December 11th, 2024 10:29 PM Author: lascivious carmine area (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48439148) |
|
Date: December 11th, 2024 10:32 PM Author: sienna internal respiration (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48439153) |
|
Date: December 12th, 2024 12:12 PM Author: lascivious carmine area (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48440746) |
|
Date: December 11th, 2024 4:41 PM Author: Razzle-dazzle hyperactive pisswyrm how much did it all come to? will a local llm be capable of translating from exotic languages? (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48438026) |
|
Date: December 11th, 2024 8:52 PM Author: Costumed milky stain Under $5k but I didn't buy a6000 card yet
Those fuckers are expensive! |
|
Date: December 11th, 2024 5:25 PM Author: rebellious candlestick maker how many watts does the PSU have? (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48438199) |
|
Date: December 11th, 2024 8:50 PM Author: Costumed milky stain 170
Edit: this is the CPU watts |
|
Date: December 11th, 2024 9:06 PM Author: rebellious candlestick maker that's a bit low, I don't think u actually have built a computer. (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48438901) |
|
Date: December 11th, 2024 9:15 PM Author: Costumed milky stain Dark Side of the Moon 1350W PSU, Advanced Liquid Cooled CPU & Clear Side Panel Item number: 321-BIKW |
|
Date: December 12th, 2024 8:50 AM Author: rebellious candlestick maker You are using an OEM Alienware part for you PSU? How does that make sense? (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48440037) |
|
Date: December 12th, 2024 8:54 AM Author: razzmatazz dilemma OP is growing a beard and changing monikres as we speak (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48440043) |
|
Date: December 12th, 2024 10:17 AM Author: Costumed milky stain (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48440299) |
|
Date: December 11th, 2024 8:53 PM Author: Ruby nighttime marketing idea stead I really want to put together my own local LLM setup too. Tagging this thread for later. 180 (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48438860) |
|
Date: December 11th, 2024 8:53 PM Author: Costumed milky stain We should start a discord or something (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48438863) |
|
Date: December 11th, 2024 8:59 PM Author: Ruby nighttime marketing idea stead I'm down, I'm still a total noob at this I haven't even started to do my own research on what hardware I'll need (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48438883) |
|
Date: December 11th, 2024 9:01 PM Author: Costumed milky stain If you have an M2 or later MacBook/mini you can install ollama or lm studio right now on an external ssd and start fucking around (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48438889) |
|
Date: December 11th, 2024 10:31 PM Author: Sepia stag film boltzmann can u show us what u can do with ur setup (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48439150) |
|
Date: December 12th, 2024 10:23 AM Author: Cyan ticket booth pozpig What will you use it for? How is it better than just getting a subscription to Claude/ChatGPT? (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48440310) |
|
Date: December 12th, 2024 10:58 AM Author: Costumed milky stain Re 1st question: scaling and streamlining CRM and customer service; reducing bottlenecks, automating as much of the client lifecycle as possible Re 2nd question: mainly compliance with HIPAA, SCA, ECPA, CPRA etc |
|
Date: December 12th, 2024 11:04 AM Author: Sepia stag film boltzmann link to wat u can get from ur setup (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48440496) |
|
Date: December 12th, 2024 1:06 PM Author: razzmatazz dilemma automating as much of the client lifecycle as possible
tptptptp |
|
Date: January 8th, 2025 8:34 PM Author: dun volcanic crater center Bump for 5090, how many are you getting OP? (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48533708) |
|
Date: January 8th, 2025 9:08 PM Author: Hyperventilating idiotic sound barrier Isn't it just easier to buy a Mac Studio or whatever it's called? Supposedly it's very good for private LLM stuff (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48533823) |
|
Date: January 28th, 2025 3:12 PM Author: Adventurous hateful parlor You can use this for DeepSeek now! (http://www.autoadmit.com/thread.php?thread_id=5647806&forum_id=2:#48598692) |