Desktop or Notebook?

[RobertJasiek]

As you know, I have tried to buy an RTX 3080 in vain and now await RTX 4070 to possibly become available at its MSRPs around January 2023. The alternative is a notebook. Desktops and notebooks have their advantages and disadvantages but a major difference is relative speed. Different benchmarks differ, cooling has a great impact etc. To simplify my decision-making, let me state these rough relative benchmarks for 2560x1440 Time Spy Graphics, which might not be the most meaningful benchmark for go AI but is also not the most meaningless:

Code: Select all

169%   RTX 3080 desktop
100%   RTX 3070 TI laptop

So a desktop build might give roughly 69% faster speed. What does this mean in practice when using, say, KataGo? Suppose 3080 desktop enables next move suggestions at above-human level within 0,25s, would 3070 TI laptop then simply mean move suggestions at above-human level within 0,4s and still be more than good enough? Or would move suggestions at above-human level require 2.5s versus 4s for a drastic difference in usability?

AI suggestions at 9p level are insufficient for me - I want clearly above-human level. You know, 9p can make mistakes but it must be hard to detect any AI mistakes.

What do you recommend me? Is a well cooled RTX 3070 TI laptop graphics card good enough for me 5d seeking clearly above-human level AI moves or is at least 3080 desktop mandatory?

For good notebook cooling, imagine MSI Titan GT77 (too loud for my taste), Alienware 17" (too loud and awkward design) or XMG Neo 17 M22 (to be available next month, noise to be tested). Can such notebooks maintain many hours of maximum go AI load without overstraining the cooling system? Or must it be a large desktop tower with exceptional airflow?

I have comtemplated these things for two years but still cannot decide whether notebooks make sense at all for go AI at super-human level. Please advise!

***

On a related question, not every notebook has all navigation keys as dedicated keys. When using go AI software, which navigation keys are needed frequently? Arrows keys? Pos1 and End? PageUp and PageDown? Is a mouse needed or will any touchpad do?

[Gomoto]

https://www.amazon.de/MSI-GeForce-GAMIN ... 603&sr=8-3

3080 are available and your best option in my opinion

[Gomoto]

By the way I use my desktop in silent mode.

I use MSI Afterburner to set power and temp limit for my 3080 and I use Windows to set the CPU max performance level to achieve silent computing on my desktop. I am ready to increase the limits if I would need them, but I never feel the urge even during demanding tasks. I achieve much better performance on my silent desktop than on any of the high quality notebook computers I am using.

I enjoy reviewing games on my desktop much more than reviewing them on a notebook. Good keyboard, mouse and monitor are important to me.

[RobertJasiek]

Almost two years after launch, German 3080 10GB prices are at 120-140% MSRP but should be about 80%. Up to May 2021, I would have paid up to €970. Now, I pay up to €630. I rather wait for RTX 4000 than being ripped off.

So you say 3080 desktop was better for me - what are your reasons, please?

[RobertJasiek]

Which exactly are your settings for undervolting, underclocking and Windows CPU? How much faster is your desktop and what japtops with what GPUs do you compare it to?

[Gomoto]

I am not a numbers guy myself. It is just my opinion. I like using the desktop more. Could be different for you. In all honesty both options are viable.

[Gomoto]

I am using a 3600x at 80% max cpu.
Currently my 3080 is sitting at a power limit of 65% and a temp limit of 70.

I can not be bothered to do any benchmarks. It works just fine.

[RobertJasiek]

Thanks! What is the experience of others?

[kvasir]

I think any laptop with this type of GPU will be loud and hot at times. I have a Lenovo laptop with a 3080 and it is pretty fast, it has shortcut keys to go into quiet mode but it will run hot and loud if you let it. I rarely if ever enable this quiet mode because I have noise cancelling headphones.

Here are some pointers:

• The way that the cards are labeled is that the fastest laptop card and desktop card are both 3080 even though they don't have the same specs. I think I remember that the number of different compute units on the laptop 3080 is similar to the desktop 3060, but I am not going to double check that. I have seen similar marketing ploys for other hardware components too.
• The number of playouts per second seems to vary a lot depending on many things, including the position. I did an analysis in katrain of a game I was watching last night (student honinbo) using the settings I happened to have configured. The game was 251 moves (252 positions) and my settings called for 10k playouts. This took 18 minutes and 5 seconds, which is just over 2300 playouts per second. When I tried 1k playouts per move instead, it took 2 minutes 22 seconds which is just under 1800 playouts per second. That is the performance is not very predictable.
• It is slowest when you are waiting but it is still very fast.
• The keyboard on this laptop is big and not missing the numpad or anything like that, but one could simply use a regular keyboard. It also came with a gaming mouse and has eye tracking hardware that I never tried.

It is probably safe to assume that a desktop with 3080 is more performance per euro than a laptop with 3080. The bottom line is that if you are willing to spend the money then laptop or desktop is probably more a question of if you wish to carry the computer with you or not than it is a question of which is faster.


==EDIT
I realized that I wasn't running in the much coveted high performance mode. When I analyzed the same game in the more coveted mode it took 12 minutes and 54 seconds, that is about 3250 playouts per seconds. It is simply fast.

[RobertJasiek]

I have seen very many notebook tests. The major noise factors are cooling, CPU / GPU power maximums and software. While typical gamming notebooks with some reasonable cooling are around 50dB for about CPU 70W and GPU 120 ~ 150W, very few content creator / workstation notebooks with CPU 45W and GPU 80W (with RTX A2000 Laptop or the like) are in the 39 ~ 43dB range. This could be because power is low enough for the good cooling or because the wrong softwares were tested. There is no reason why gaming notebooks could not be designed modestly alike. But maybe you are right that every dGPU notebook can be loud at times. Desktop is the safer way towards relative silence down to circa 37dB.

[RobertJasiek]

Lenovo laptop with a 3080 [in] high performance mode [...] 3250 playouts per seconds

goame reported 2x RTX2080 TI KataGo 40-block 384 channel ~1667 playouts per second. To compare this to your number, which network (blocks, channels) have you used?

[thirdfogie]

On a related question, not every notebook has all navigation keys as dedicated keys. When using go AI software, which navigation keys are needed frequently? Arrows keys? Pos1 and End? PageUp and PageDown? Is a mouse needed or will any touchpad do?

There are two different phases in my use of Lizzie and KataGo. First analyse the whole game, which takes about 90 minutes at 7000 visits per move. Then save the resulting SGF file which contains the analysed results so that it can be reviewed, also using Lizzie and KataGo. (Bad things happen if you analyse with LeelaZero and then try to review with KataGo, or vice versa.)

Lizzie has several single-key commands, but none that should be a problem on a laptop. For example: Up, Down, Left and Right arrow to move through variations; '.' to switch territory shading off and on, 'a' to start an analysis run, Space to switch analysis on and off at the current point in the game and Ctrl-1 or Ctrl-2 to switch between the two engines, which are LeelaZero and KataGo. PageUp and PageDown go forward or back 10 moves, but you can do without them if they are inconvenient on a laptop.

A mouse is used to select Lizzie's menu items and to move through the graph of results to focus on blunders etc. It is also useful to move the cursor off the board display during analysis, otherwise Lizzie thinks you only want to see the results of a move at the point under the cursor. As others have said, a plug-in USB mouse would be cheap and easy to carry around.

I have briefly used q5Go instead of Lizzie. From memory, it uses more mouse movements and fewer single-key commands than Lizzie, and also updates the screen less often. Indeed, Lizzie updates the screen rather manically.

It feels likely that a German-language installation would be little or no different to an English one, but you should be the expert there.

Whether you buy a laptop or a desktop really depends on your expected usage. Do you want to perform analysis at tournaments or in a summer house (Kleingartenhaeuschen)? If so, buy a laptop. If not, buy a desktop. In my experience, a desktop is likely to be better value, longer-lasting and more upgradeable. Your mileage may vary.

Here is an English proverb for you: There's nothing certain, save death and taxes.

[RobertJasiek]

My choice between desktop and notebook depends on availability of preferred hardware, price, noise, speed, ergonomics of keyboard / mouse / display, longivity, and flexibility at home (where the garden may be unrealistic without power cable). The decision is somewhat difficult for me because there are pros and cons.

[kvasir]

Lenovo laptop with a 3080 [in] high performance mode [...] 3250 playouts per seconds

goame reported 2x RTX2080 TI KataGo 40-block 384 channel ~1667 playouts per second. To compare this to your number, which network (blocks, channels) have you used?

40 block 256 channel (are there b40c384 networks, thought there were b40c256 and b60c320) and I use a katago built for the tensorRT backend. I also made a modest attempt to tune the batch size (96) and the number of threads (32) for higher performance, there is a command in katago to help with finding such a setting. The tensorRT backend has a big drawback that it takes forever to start katrain (but I have the habit of leaving it open).

I measured it again with the same sgf and this time it was 16 minutes and 12 seconds, which is about 2600 playouts per second. It is a different number every time for me. Maybe it is due to temperature throttling, I don't really know because it is a bit too large of a difference between the measurements for me to explain it.

The number of playouts depends a lot on the position, probably it comes down to the branching factor of the search and there is a also a cache that maybe speeds up things (I don't know how this cache works). This means it might not be directly comparable to benchmarks for a single position, for example an empty board, and to benchmark the analysis of every position in a game to 10k playouts.

Btw I might use less playouts, the thing is that the number of playouts in analysis and when playing katago in Katrain appears to be the same setting and I can't be bothered to change it back and forth. I have a project to practice playing certain positions against the computer and I have set the number of playouts for this activity It is too annoying when the computer responds instantly (I have hard enough time playing seriously against the machine as it is) and I rather prefer a stronger opponent for my little project anyway.

[RobertJasiek]

Interesting! So it is hard to distinguish numbers of playouts per second within a factor of roughly 2.