VoidTheWarranty

On 23/02/2024 at 13:02, UltraMega said:

Nvidia is selling massive GPUs and NPUs to data centers. They have products that sell for a quarter million each in the data center market. 

 

They are also the leader in a market that is growing rapidly. 

 

Nvidia H100: This is the chip behind AI's supersonic stock rally (theedgesingapore.com)

  Reveal hidden contents

Computer components are not usually expected to transform entire businesses and industries, but a graphics processing unit Nvidia Corp. released in 2023 has done just that. The H100 data centre chip has added more than US$1 trillion to Nvidia’s value and turned the company into an AI kingmaker overnight. It’s shown investors that the buzz around generative artificial intelligence is translating into real revenue, at least for Nvidia and its most essential suppliers. Demand for the H100 is so great that some customers are having to wait as long as six months to receive it.

1. What is Nvidia’s H100 chip?
The H100, whose name is a nod to computer science pioneer Grace Hopper, is a graphics processor. It’s a beefier version of a type of chip that normally lives in PCs and helps gamers get the most realistic visual experience. But it’s been optimized to process vast volumes of data and computation at high speeds, making it a perfect fit for the power-intensive task of training AI models. Nvidia, founded in 1993, pioneered this market with investments dating back almost two decades, when it bet that the ability to do work in parallel would one day make its chips valuable in applications outside of gaming.

2. Why is the H100 so special?
Generative AI platforms learn to complete tasks such as translating text, summarizing reports and synthesizing images by training on huge tomes of preexisting material. The more they see, the better they become at things like recognizing human speech or writing job cover letters. They develop through trial and error, making billions of attempts to achieve proficiency and sucking up huge amounts of computing power in the process. Nvidia says the H100 is four times faster than the chip’s predecessor, the A100, at training these so-called large language models, or LLMs, and is 30 times faster at replying to user prompts. For companies racing to train LLMs to perform new tasks, that performance edge can be critical.

3. How did Nvidia become a leader in AI?
The Santa Clara, California company is the world leader in graphics chips, the bits of a computer that generate the images you see on the screen. The most powerful of those are built with hundreds of processing cores that perform multiple simultaneous threads of computation, modelling complex physics like shadows and reflections. Nvidia’s engineers realized in the early 2000s that they could retool graphics accelerators for other applications, by dividing tasks up into smaller lumps and then working on them at the same time. Just over a decade ago, AI researchers discovered that their work could finally be made practical by using this type of chip.

4. Does Nvidia have any real competitors?
Nvidia controls about 80% of the market for accelerators in the AI data centres operated by Amazon.com Inc’s AWS, Alphabet Inc’s Google Cloud and Microsoft Corp’s Azure. Those companies’ in-house efforts to build their own chips, and rival products from chipmakers such as Advanced Micro Devices Inc. and Intel Corp., haven’t made much of an impression on the AI accelerator market so far.

 

Thank you!

Can someone please explain to me what Nvidia is actually selling besides shares? I'm a distributor in real life I represent 40+ companies in sales for me I like having factual tangible equipment to sell and being able to research it.  Few days back unfortunately I don't have it (I'll post if I find it again) but It was showing these insane sales from Nvidia that were trouncing their GPU sales. Okay of what and to who? Seriously can someone name a part number or products and in what scale are they being sold and to who? I seriously want to know cause I must have missed something here. The weird thing is I have asked numerous times around Reddit especially in investment subs and nobody can seem to give me an answer. 

It kinda seems like I'm watching this.
CEO to IT should we get some of that AI? Yeah for sure.
Okay here's a check for $10,000,000. Make it happen IT.
IT sooo ummm lets just put the money into Nvidia stock and show the CEO ChatGPT. 
Bonuses for the IT department this year. 

So whenever I hear AI Nvidia pops up. I feel like Nvidia has an incredible Marketing team they sell the crap feature for a premium to idiots. While that's not bad when it comes to sales it's another when it comes to a long term investment into a specific technology. I worry Nvidia may still be overpriced from Crypto + pandemic however are they seriously that much of a dominant fixture when it comes to AI? Who are other serious players considering TESLA and NVIDIA stock seemed to be linked for whatever crazy reasons right now.  

Just giving folks a bit of an update here. So I'm locked into these settings for my CPU. I don't know what it is about them but they seem to perform the best no matter what I do. Can I get slightly better multicore yes. Can I get higher single thread performance no. Opting to go that route because this system is for entertainment and gaming and not multi-threaded work. 

PPT: 120
TDC: 65
EDC: 90
Scalar: X1
1.110 SOC Voltage.
XMP ON.
-20 on most cores -16 on my best two. 
Running +25Mhz or +50Mhz it doesn't make much difference performance wise. However increasing it further to +100Mhz there is a performance drop when benching.  

I might mess around with my memory however I don't think I'll be doing much OCing over the summer. I don't have AC in my spare bedroom where my computer is and it would be like comparing Apples to Apple Pie since most of my benching was done in the winter, also with the window cracked. It's been in the high 20's (80's Fahrenheit) this week and watching twitch and web surfing my CPU is in the lower 50C's and gaming I'm in the low to mid 70C's for example. 

I still feel like there's something about the 120-65-90 that just squeezes performance out of the 5600X. I have pages and pages of notes and benchmarks so maybe this will help someone else out if they try it. 

Anyone work in the industry? Seems like there's a lot of hype in AI right now however I don't know if it's basically investment marketing to pump these companies up or actually legit. Obviously AI has merit however is it completely overblown at this point.

Does anyone actually work in the industry or AI R&D on here?

It's like Nvidia every couple months sends me a reminder that I don't need to have any buyers remorse over switching teams. Thanks Nvidia. 

So this weekend I didn't have much time to do anything however I just got a massive single thread performance boost.

Currently I've been rocking -16 2 best cores -20 on rest. I tried -17 and -20 wasn't anything special improved multicore like 15 pts. So then I tried the reverse I went back to -16 on my best and reduced -20 to -19. Boom! Huge uptick in single thread. With -16 and -20 + 50Mhz boost. I was at 632 single thread and 5030 Mult-thread in CPU-Z. With -16 and -19 + 50Mhz Boost I'm at 646 & 4999. It reduced my mult-thread performance granted my temps weren't as good due to the weather being significantly warmer and I toned down my fan curve a bit but I'll take it.

Single thread performance has been the hardest to increase. 639 & 5045 has been my best up until this point but that was with 100Mhz boost. So with 50Mhz boost and my single thread is beating that I'm going to experiment with this. I'm still utilizing my Radioactive Spider Venom:  PPT: 120  TDC: 65  EDC: 90  Scalar: X1

Link: 

AMD Ryzen 5 5600X @ 4623.92 MHz - CPU-Z VALIDATOR

VALID.X86.FR

---

[f560ul] Validated Dump by DESKTOP-GFAEV7D (2023-05-08 21:04:50) - MB: MSI MPG B550 GAMING PLUS (MS-7C56) - RAM: 32768 MB

2 hours ago, kaliz said:

so have you tried to set the timings i showed you? preferably with the subtimings? Setting a manual value Soc 1,1v is good. or leave it on auto which will give about the same. You are on a overclockers forum and you dont want to adjust RAM timings, i cant understand? Maybe because for me its my hobby

Ram overclocking is like dry humping on a date in your 30's...  All you're going to end up with is like a 2 FPS increase and case of blue balls getting her there. 

17 hours ago, kaliz said:

 this will work for you, and if it doesnt, raise tRFC to 600. 

 

leave tRDWR and tWRRD on auto

I might meet you 1/10th of the way by trying your 18-22-22-22-54 in 3733 1.36v. Should I increase my SOC to 1.2 or should I just leave it on auto?  Should I leave XMP on or disable it and have everything at auto except my main timings?

Would tightening up my current timings even be worth it? How low do you think I could get my timings without much voltage increase with my current 3600Mhz speed? I don't want to increase voltage too high since heat is a factor in my case and I don't have a fan blowing on them.  Current: 18-22-22-22-42 to 18-20-20-20-40 or something?

Long story short I wasted a month this winter trying to get 2 dual rank modules to work I'm kinda put off with memory tweaking after dealing with that and the return, and having to reinstall everything. Pretty nervous especially if it borks my system since everything is running fine. Also I like to be able to quickly just set everything back to normal after a BIOS update and individual ram timings is not a fun time. 

26 minutes ago, kaliz said:

wouldnt recommend running stock XMP timings. The subtimings are so loose

With a little bit of tuning its worth it 

 

Could you take a Zentimings screenshot?

and what die is it? Hynix? maybe take a Thaiphoon Burner read out so I can help suggest some timings which will work

 

for CPU just Enable PBO , set +200 MHz override, but dont use Curve optimizer leave that disabled. 

 

Thaiphoon Burner won't load for me. Worked for my last ram but I tried 2 versions and it won't work. I'm not sure what I have. I'll try looking for a different number or something but this is the set I'm working with (Without more detail). 
 https://www.corsair.com/ca/en/Categories/Products/Memory/VENGEANCE-LPX/p/CMK16GX4M2D3600C18 CMK16GX4M2D3600C18

5 hours ago, kaliz said:

the EDC bug was with old agesa where if you set 1 EDC on 'locked' motherboards, youd have unlimited EDC. This is fixed years ago already as far as i know (i asked on OC.net) 

 

voltages of 1,4 - 1,45v in idle are very normal for Ryzen. this looks all good.

 

for some extra performance set your SOC voltage to 1,2v and RAM to 3733 or 3800 if possible.

 

and +200 instead of +100, with auto Vcore voltage. And disabled PBO.

 

maybe for powerlimits 90 75 125 and start running Ycruncher and adjust powerlimits in windows with so you can find the sweetspot (maximum)

I feel like I have tried those settings before but I'll check them out. 

I went through a month of frustration with some ram so I don't know that I will spend much time on it. I have my 3600Mhz kits running stock right now. I might try my stock settings with 3733 but I'm not sure I want to push my luck. My memory usually only uses around 1.34v not even the 1.35v should I still increase the SOC voltage?

I'll be honest with you dudes running a TDC of 75 and EDC 125 previously I was performing worse in benchmarks. I know with a TDC of 75 anyway. I might try it again but I'll look at my notes. I really hit a wall with certain settings that's why I am using what I am now. I started going through too many post-it notes so moved to notebook. 

What I find odd is that in CPU-Z if I have my boost set to +100 I'm hitting my stock boost. However if I have it set to something less or off then it won't be listed in CPU-Z.  

2 hours ago, Sir Beregond said:

I am just talking about boosts where it hits higher voltages for micro seconds/nano seconds/whatever. 

 

It probably doesn't matter, just something I thought of.

You know part of the reason I went with a lower EDC was because I heard there was some sort of EDC bug and I couldn't find out if it was fixed or not so exactly so I figured run with 90 see what happens. I seem to get better performance with lower voltages than higher ones. 

4 hours ago, Sir Beregond said:

What agesa are you running? Wondering if you are topping out at 1.42, 1.43v on boosts due to that EDC bug (or some would say intentional limit) with newer agesa. I think you are stuck at 90A for EDC on a 5600X.

I have 1.2.0.8 for my MSI. But I mean I'm trying to avoid being in the 1.4XX I didn't think I should spend a lot of time there. I guess it shouldn't matter unless I was there 24/7 but I honestly don't know that the maximum the 5600X should be hitting in Hwinfo64 I've been trying to keep things in the 1.2X-1.3X for fan noise.
 

5 hours ago, kaliz said:

idling in low 50's then there is something wrong with your cooler either the mounting or too little thermal paste, i would look at that first because that is not normal for a 5600X? 

if you have HWinfo64 tab open and run Ycruncher pi-1b, what is your temp and voltage? if you dont have that yet its in Benchmate (benchmate.org) 

thats the only way i can give you proper advice on your system  

I mean I wasn't quite idling I was downloading Call of Duty MW 2019 so I was being a bit dramatic. But my max temps in Hwinfo 64 when not benching would hit 1.4X and it made me nervous. I think my lows were 38-45C somewhere in there. Factor in this is in a NZXT 510 case not exactly a 1337 cooling rig. 

You can see my temps and voltages from the CPU-Z run below. Keep in mind I had my window open and it was like 10C or 50 degrees outside. Also you can see the Cinebench score above which would likely be a better guide for temps and such.

So first of all my offsets for these runs are not fine tuned so I could probably do better. 
The +100Mhz has a substantial impact on single core scores.
The following PPT-TDC-EDC will however provide you with substantial performance especially in regards to Multicore. I have pages and pages of different settings benchmarked however I seem to be narrowing down on these for stability in benching/gaming. I didn't have any games or benchmarks crash previously but some games didn't feel quite right..... Just not dialed in.
Radioactive Spider Venom: 
PPT: 120
TDC: 65
EDC: 90
Scalar: X1

Score I get without the +100Mhz  (Peter Parker)
Single Thread: 624   <-- Average 5600X thread performance but CPU vcore and temps reduced
Multi-Thread: 4914  <-- Multicore still excellent with 120-65-90
-20 allcores except 2 best cores -15
https://valid.x86.fr/k9j6k3

Score I get with +100Mhz (Venom)
Single Thread: 639     2nd test cause I didn't believe it: 639   <-- Do you even lift bro? 
Multi-Thread: 5045    2nd test: 5011 <-- Still above 5000.
-20 allcores except 2 best cores -15
https://valid.x86.fr/m959w2

Tombraider Benchmark Score Avg FPS: 170  Frames Rendered: 26645 GPU BOUND: 38%  <-Best FPS/Most Frames I've rendered before.
Cinbench R23 Score: 11689

Having a beast of a multicore score is great however boosting the single core has been the hardest part. Increasing the Mhz on the chip has definitely helped. I'm not too sure how at this point I can reduce my idleish - low load temps / voltage other than changing coolers obviously. CPU CORE voltage (SVI2 TFN) Maximum on HWINFO64 hit around 1.42-1.43 or something like that. It might spike to that at desktop here and there as well. 

I'm going to run Ryzen Master to see what it tells me for max offsets per core sometime and try those values out.

Idling in the low 50's isn't going to kill my chip nor will the voltage likely kill it during the time I plan on using it from what I understand from Der8auer. He ran his like full bore 24/7 for a year and required a 20mv increase on one of 2 I think . I'm not worried about dipping into the 1.4v range during Max voltage range in HWINFO64 for benchmarks but I'm not sure about it spiking that high here and there at the desktop and being around the 1.35X territory frequently. You guys tell me from your own experience. 
 
***→ Does anyone know how sketchy or dangerous it would be to run a CPU Offset Voltage of say -.1000V with core offsets and PBO on? I'm not sure if you should do a cpu voltage offset while PBO/core offsets are on. I worry about stability especially at the lower power end. I suspect it's keeping my voltage high to keep it stable.  

CPU-Z Multi-Thread

CPU-Z Single Thread

Does Cinebench not show you your previous runs anymore? I can't figure out how to get it to show that now.

I hope at this point nobody is following this thread in hopes of learning something specific. I think we've gone off the rails a couple pages back. 

My weekend is cut in half this weekend (inventory at work) unfortunately so if you're here for the amusement I might only have a few settings to mess around with.  

I think I might try running Ryzen Master at some point and see what it tells me for individual core offsets. The thing is I read a lot of people saying things like "Master tells me that I can run -30 however it's unstable". I haven't used Ryzen master before, does anyone use it or the auto OC feature? Not sure if I should just dial whatever it tells me per core offset back 20% per core for less demanding software stability. 

But I want to try +100 and +200 Mhz with 10-60-115 and 110-65-115. I feel like I might have to increase my TDC/EDC a bit I'm not sure. At least I'll know where my temps land if I can even get +200 stable unless all it's going to be is clock stretching which I don't quite understand.  


**COD 2022 was free last week for trying. I was using it to test stability. Does anyone have a spare code for the game? I have some game codes I'd be willing to trade for it. Shoot me a message if you're interested.  

On 25/04/2023 at 14:50, neurotix said:

That's weird because with my 5900x just setting max boost clock override to +200 and PBO_Fmax Enhancer (Asus exclusive) would get me to 5150MHz in games.

What does pbo_Fmax enhancer do?

On 26/04/2023 at 16:53, neurotix said:

I wouldn't recommend it on the Wraith Prism cooler, I don't think it could handle it. I'd suggest motherboard limits and +150. You might see really high temps. If you don't, then try 200.

 

I used a Wraith Cooler for a while on my 3900x when I first got it in 2019- looks pretty but on a 3900x I basically couldn't overclock at all, temps were too high.

 

I'd suggest getting a better cooler if you have the money. A H100i or Liquid Freezer II 240mm would definitely allow you to leave CO alone, set motherboard power limits to the max, and do +200 boost clock.

 

Though I digress, under a gaming load doing +200 and mobo limits might be ok in games because you're really probably only going to see 1-2 threads hitting 4800MHz. If you don't, the chip might not have enough voltage on auto and you might need to apply a positive voltage offset of 0.025v to see it hit those speeds. Until the last bios update, this is what I had to do to see boosts to 5650MHz on CCD2 on my chip.

 

Hope this is helpful. Just remember, set your co stuff back to 0 if you try this route.

I've been lingering around 110-60-115 to 110-65-115 as of late. When I ran -20 allcore and +50 boost I got some of my best benchmark scores this was after getting bored of tweaking each core. But it felt a bit off in games. I might try the +100 and hope it doesn't crash with 110-65-115 but not sure about an offset. This is a bronze chip I'm working with. 

17 hours ago, neurotix said:

I wouldn't recommend it on the Wraith Prism cooler, I don't think it could handle it. I'd suggest motherboard limits and +150. You might see really high temps. If you don't, then try 200.

 

I used a Wraith Cooler for a while on my 3900x when I first got it in 2019- looks pretty but on a 3900x I basically couldn't overclock at all, temps were too high.

 

I'd suggest getting a better cooler if you have the money. A H100i or Liquid Freezer II 240mm would definitely allow you to leave CO alone, set motherboard power limits to the max, and do +200 boost clock.

 

Though I digress, under a gaming load doing +200 and mobo limits might be ok in games because you're really probably only going to see 1-2 threads hitting 4800MHz. If you don't, the chip might not have enough voltage on auto and you might need to apply a positive voltage offset of 0.025v to see it hit those speeds. Until the last bios update, this is what I had to do to see boosts to 5650MHz on CCD2 on my chip.

 

Hope this is helpful. Just remember, set your co stuff back to 0 if you try this route.

Cooler Master who made these coolers to their credit actually did a good job AMD who cheaped out on the stock coolers for the 5600X I feel for anyone using the stock cooler. It's funny though cause my 3800X I feel like this cooler wasn't super adequate but for my 5600X it's just laughing. I can play COD with tweaked settings for a couple hours in my NZXT H510 case (not exactly know for great cooling) the hottest I'd hit is like 67C and that's not my average that just max temp. If CM offered an upgrade kit for a better fan I'd jump on it cause I love the cool of the cooler. Balancing the noise is the biggest hurdle. 

20 minutes ago, neurotix said:

What cooler do you have? Personally I just set motherboard limits and leave those alone, so I set current capacity to 140% for both the CPU and IO die and change other stuff in the digi+vrm section. I basically nax out the motherboard limits so it makes those irrelevant, then just set max boost clock override to +200 and enable PBO and I boost high and never crash. Leave voltage on auto.

 

I couldn't understand anything about the CO stuff and that video that was posted earlier in the thread about how to use it was just baffling to me so I don't even bother with CO.

I use a RGB Wraith Prism so the AMD cooler that comes with 8 core processors and up. I use that from my 3800X for my 5600X which does a good job. My high temps are usually 67-72 degrees. 

I used motherboard limits for a bit but then I was trying to fine tune things. Maybe I'll give mobo limits a go again and try 200. See what happens.

22 hours ago, Sir Beregond said:

Ah ok, that does look a bit different than mine, but ultimately the same thing. Yeah that will just add to the potential upper limit that your maximum boost clocks can reach. 

 

5600X can boost to 4.6GHz stock, so a +200 override should allow for theoretically boosting up to 4.8GHz for bursty single threaded stuff.

 

Curve Optimizer comes in and by undervolting gives you additional headroom to potentially boost even higher is really the best way to explain it. Stock PBO with +200, max I get on the boost with 5900X is 4.95GHz. Add in CO on auto and I can get up to 5.05GHz. Optimize CO manually and I've gotten it up to 5.15GHz boosts.

 

Yeah I'm curious what will happen especially with stability instead of setting it to 50 or 200 but to set it to just positive and auto. I might try it and see what happens. I also worry if I set it to 200 it will just crash. At 5.15Ghz what do you have your PPT-TDC-EDC set to for it to be stable? 

19 hours ago, Sir Beregond said:

Positive, as in a positive Curve Optimizer? Your BIOS maybe looks different then mine. I am not following the ask in terms of Auto + positive.

5 hours ago, damric said:

The hardest part for PBO tuning is idle and light load testing. There's plenty of heavy load stress testing available but that doesn't help if your rig is crashing at idle

 

The curve optimizer is the right idea but it's just not anywhere near capable of certifying that you won't crash under light loads or idle.

 

I miss the old Phenom MSR Tweaker. It was extremely easy to test all P-states with that software.

I'm finding this sorta but I can't get it to crash. I'm trying to tighten the ship up. Right now if I want I can hit some CPU-Z multiplayer performance I'm hitting between 12500 - 12600 level however when gaming ehhhhh I'm not so sure it's 100%. Although it's also not crashing same with R23 or Tombraider benchmark. 

How many of you folks left Max CPU Boost clock override off?


*** Anyone feel like trying something ***
I haven't tried this but has anyone tried to do this I'm curious what will happen but maybe someone with liquid cooler should try it.

Max CPU Boost clock override: Auto + Positive. What does it boost you to? I assume it will just boost your clock high as it can go. I thought about disabling curve Optimizer and letting stock settings boost to wherever it wants based what you set your PPT / TDC /EDC as a safety net. (I could be wrong).

I'm just curious if it will help reduce clock stretching or anything else.

I haven't tried this but I'm kinda curious where it will take you. 

 

On 20/04/2023 at 10:10, Sir Beregond said:

+200 just means you don't have as aggressive a curve optimizer. It's all in what you are looking to get out of it. Optimized CO can sometimes overcome that gap. For example I was boosting higher with per core CO and +125 than with a straight +200. That said, still had some instability on low load/idle, so needs some more work. The one downside to CO, really a PITA to tweak. Though should be easier with a 5600X than a 5900X.

 

Can confirm, for non 3d v-cache Zen chips, specifically tuned memory makes a big difference. Throwing in a generic 3200 CL16 kit and then going to a tuned C14 3600 kit OC'd to max your max FCLK really makes a difference. For Zen 3, it's pretty rare to have an FCLK that hits 1900 or above stable, so usually I'd say a 3600 kit is the way to go. Looks like you have that already though.

Yeah I have the 3600's. Unfortunately Thaiphoon burner isn't loading for them so not 100% sure what they are. I would tighten them up but so far they are stable with XMP enabled in my BIOS. 1800/1800 1:1. 

21 hours ago, Sir Beregond said:

I found this guy's charts in the first part of the video pretty good for understanding this.

 

 

Thanks I'll check it out. Definitely boggling my mind here.