Fat Fritz: what videocard to buy

by Albert Silver
11/14/2019 – With the release of Fritz 17 with Fat Fritz, you may be strongly motivated to upgrade your computer to enjoy the best experience, or actually upgrade your entire machine if you feel it is time. In this first guide you will be shown the simplest upgrade possible: just adding a fast GPU to your computer. All the major cards as as well current prices and what performance to expect.

ChessBase 15 - Mega package ChessBase 15 - Mega package

Find the right combination! ChessBase 15 program + new Mega Database 2020 with 8 million games and more than 80,000 master analyses. Plus ChessBase Magazine (DVD + magazine) and CB Premium membership for 1 year!


Important note: Before starting, I’d like to just point out that while I own a machine with the highest specs one could want, I run Fat Fritz regularly on my 3-year-old laptop which can run it only at 2000 NPS (bottom of the performance list below). It does not bother me one bit, and I get very useful and enjoyable analysis. While the guide is designed to help those seeking an upgrade, you should not think that all upgrades are required just to derive any benefit or enjoyment from Fat Fritz.

Even on a modest laptop at 2000+ nodes per second, Fat Fritz is quite enjoyable. True to its style, it doesn't need big guns to know it wants a Marshall Gambit here.

Technology never stops evolving and keeping up with all the changes in the market can be more than a little daunting. It is for this reason many people find the ready-made machines by companies such as Dell or others so attractive, but the truth is the budget-conscious can almost always do a lot better with smartly selected parts.

Upgrading for Fat Fritz (or Leela)

The good news if all you really want now is to enable your system for a good or great experience with Fat Fritz (or Leela) is that in most cases all you need to do is upgrade or install a good Nvidia graphics card. This explicit choice of Nvidia is not really optional unfortunately since it is the special codebase known as CUDA, exclusive to it, that gets such extraordinary performance for the neural networks. AMD may no doubt catch up in this area in the future, but as of now Nvidia is the hands down choice.

You won’t actually need a whole new system, and even a fairly modest computer won’t need more. Remember the powerful machine set up for Fat Fritz on the Cloud? Well the truth is that the core of the machine is a very modest quad-core processor, four years old. But it also has two powerful graphics cards, which are the heavy lifters here.

In other words, even if you think your computer is nothing special, if you just add a good video card you will be running Fat Fritz and Leela as fast as the best. No need to point out that a fancy graphics card also offers many other attractive options beyond just running them.

An example of additional perks from a good GPU

Here is a rough breakdown of the various graphics cards, the NPS performance (nodes per second) you can expect from them, as well as the current prices cited in the US as of this writing by the online electronics store NewEgg (for reference):

Graphics card Nodes per second Price US$
RTX2080 Super
RTX2070 Super
RTX2060 Super
GTX1660 Super

As you can see, there is a massive leap in performance when you buy one of the new RTX generation cards, and in terms of price performance it just makes the most sense. At the top, the 2080ti really is the king of the single card performance, but if you are considering buying just one, then consider also an alternative plan to instead buy two 2070 Supers, or two 2080s. You'll get superior performance for the same price, though there are other considerations, which will be considered in the next guide on complete new systems.

Topics: Fat Fritz

Born in the US, he grew up in Paris, France, where he completed his Baccalaureat, and after college moved to Rio de Janeiro, Brazil. He had a peak rating of 2240 FIDE, and was a key designer of Chess Assistant 6. In 2010 he joined the ChessBase family as an editor and writer at ChessBase News. He is also a passionate photographer with work appearing in numerous publications.
Discussion and Feedback Join the public discussion or submit your feedback to the editors


Rules for reader comments


Not registered yet? Register

retspan retspan 39 minutes ago
From what I've been reading on the internet, not all Nvidia cards have hardware support for fp-16. The old 10xx series only use software for fp-16, or if they do have hardware support, the number of dedicated cores to run fp-16 is so small that it's not worth it to use fp-16. So you won't get any speedup by using fp-16 on those cards. In fact it may even be slower. See https://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/5 for more details.

The RTX cards use tensor cores to run fp-16, and so you will get the largest speedup from those cards. See https://www.phoronix.com/scan.php?page=news_item&px=LCZero-NVIDIA-Benchmarks.

The latest GTX cards (1650, 1650 Super, 1660, 1660 Super, 1660 Ti) have special hardware (not tensor cores, though) to run fp-16, so you will get about 2x speedup (theoretically - I haven't seen any benchmarks yet) from those cards.
rav555 rav555 11/29/2019 07:49
Does Fat Fritz support dual GPU AIB? It would be interesting to see Fat Fritz on a 2080 Ti take on Stockfish running on a Threadripper 64 core 128 thread machine. The TR3 machine would need at least 128 gb ram though. At least 1 gig per thread I would think.
Ricbennet Ricbennet 11/25/2019 02:17
I update my Nvidia driver to version 441.20 et now Fat Fritz cuda seems to work fine (not very fast but fine).
Ricbennet Ricbennet 11/25/2019 10:51
In addition to my cuda problem, can you provide the minimum Nvidia driver version to be used ? Thank you
Actually I have the 398.36
Ricbennet Ricbennet 11/24/2019 09:46
Same problem as Graand
""The Fritz (lc0) cuda caused an exception. This is problem in the engine." -> this is a message I receive after trying Fat Fritz cuda to run. "
I have a Surface Book 2 - i7 processor- GeForce GTX 1060 (6 Go) de Nvidia - 16Go Ram -

The Nvidia 1060 support CUDA so where is the problem ?
The Fat Fritz engine is not optimized ?
OlivierEvan OlivierEvan 11/20/2019 06:34
Chess computer, config file FatFritz guide for blitz and long game by @AlbertSilver.
Stay connected <:o))
misterpunch misterpunch 11/19/2019 04:53
I built your $500 machine from your 2017 article, I'll have to see what kind performance I get.
Can Fat Fritz use more then 1 or 2 cards? Im mining crypto on a 11 x gtx 1080 card rig and was wondering if Fat Fritz or Leela will use all the cards.
johnfmichael johnfmichael 11/16/2019 07:40
I run Leela Chess Zero in CB15 with an RTX2060, but from what I'm seeing I can't tell that Leela is even using the GPU at all. How can I tell? And if I purchase Fritz 17, how will I be able to tell if Fritz is using the GPU?
mamak mamak 11/16/2019 12:55
Can fat fritz run in GTX 1070? its not listed above
graand graand 11/16/2019 08:33
"The Fritz (lc0) cuda caused an exception. This is problem in the engine." -> this is a message I receive after trying Fat Fritz cuda to run.

Is it a matter of engine settings or NVIDIA issue on my three year old laptop (Win10 Prof, 840M, i7,16MB)? Frits 17 runs at 2500kN/s. Fat Fritz opeccl -> 200N/s :(

Where could I read about settings in general or ask for help? I'd really like to solve these problems by myself and don't bother you any more, but I can't.

Chessbase support doesn't answear :(
Samtd17 Samtd17 11/16/2019 06:18
Very nice 👍 article . Thank you!
glanmaster glanmaster 11/16/2019 04:17
I have a laptop with an NVidia Geforce 930/ Intel Core i7 and 8 megs of ram. Will Fat fritz or Leela be usable on my system?
Albert Silver Albert Silver 11/15/2019 11:17
@Olivier - La semaine prochaine, c'est promis.

@mvychessspecial - The model I have is the MSI GeForce RTX 2080 TI VENTUS 11G OC and it has been running without any issues for a year now.

@Promatix - Those speed numbers you quote are inaccurate. I know someone with a dual 2070 setup and his speed on the start position is around 40knps. With the 2080ti alone I get around 36-37knps. Remember also that two GPUs also means double the power consumption and a much larger and more expensive power supply.

@Samtd17 - There is nothing wrong if you own one, but only the Nvidia cards can use CUDA and the CUDNN and CUDNN-FP16 backend. In a nutshell, a card of equal value by Nvidia can get over 5x the speed of an Nvidia card. This is specific to Fat Fritz and Leela, not games or other uses.
basler88 basler88 11/15/2019 08:03
Thank you very much Albert Silver!! I really appreciate your help.
Samtd17 Samtd17 11/15/2019 06:36
Why Is AMD bad? I did not understand. External AMD gpu will work w apple. Nvidia does not.
EatMyShorts EatMyShorts 11/15/2019 05:53
@Elektrosmog Thank you for that info, and I hope that settings will be addressed in a future guide. Yes, it is somewhat frustrating that information is being given to use piecemeal, but it's better than nothing I guess. In the meantime, perhaps you can expound upon this fp=16 issue, for the benefit of us config-ignorants?
Elektrosmog Elektrosmog 11/15/2019 05:27
You fail to explain that the option fp=16 is the major issue, which boosts the speed more than twice. Config-ignorants waste more than half the power and you should help them.
mvychessspecial mvychessspecial 11/15/2019 03:51
Can you name a specific rtx 2080 ti card to use.Money is not an issue.
promatix promatix 11/15/2019 01:01
Best value per1$ RTX 2070 61 ns/1$ 2080 Ti 33 ns/1$ so its better buy two 2070 than one 2080Ti
Abhyudit98 Abhyudit98 11/15/2019 12:20
is gtx 960M with intel i7 6700HQ decent?
graand graand 11/15/2019 08:31
On which NVIDIA card Fat Fritz CUDA will run? Is GeForce 840M too old for it?
And also - can I play games like Friendly or Rated games against other engines in Fritz 17 enviroment/program or is it only possible against Fritz 17 engine there?

I'm new to this stuff so please don't laugh guys :)
OlivierEvan OlivierEvan 11/15/2019 07:40

Un article ou des informations sur le ''setting'' de FF que vous avez utilisé pour les matchs serait intéressant.
Nordlandia Nordlandia 11/15/2019 04:09
Werewolf: I recommend that you start with 3 threads first then compare it with four.

On my system Lc0 use about 1.5 core under load (i7-5960X 4.5GHz) so 2X 2080 Ti may benefit of 3 cores instead of 2.

NVLink is the new term that is replacing Sli.
ramon au cardero ramon au cardero 11/14/2019 09:29
A question.
The fritz 17 brings several Fat Fritz engines how to know which one I should use.
Albert Silver Albert Silver 11/14/2019 07:30
@clamic - It is a complicated question since it depends a lot on your budget. Still, if I had to find a balance, I would say 2x 2070 Super might be ideal. The 2070 Super is still a good 15% faster than the 2070, and you lose a bit of value I suppose, but not enough to disregard it.

@Basler88 - 51% is the total winrate, meaning if you play it 1000 times, it expects you to score 51% of the points. +23% means it thinks it will win 23% of the time, =56% means it expects 56% of the games to draw, and of course =21% means it think it will lose 21% of the time.

@werewolf - It varies for Round Robin and Multiplexing, so I recommend testing on a couple of middlegame positions. Regardless, if you use RoundRobin set the threads to 3, and if Multiplexing, set it to 4.
Werewolf Werewolf 11/14/2019 06:23
@ Albert for 2 x 2080 Ti would you change threads to 3 (as recommended by the Lc0 team for Lc0) or 4?
And do you favour Round Robin or Multiplexing?
basler88 basler88 11/14/2019 05:34
I just have a question when we already taking about Fat Fritz - can somebody explaing it to me how to read and what it's mean (51% + 23% = 56% - 21%)? Thanks!
goeland goeland 11/14/2019 05:04
The new digital world. I'm able to play World of Warcraft on my laptop and maybe one day, will be forced to buy a new computer for chess.
clamic clamic 11/14/2019 04:59
Hi Albert,
If you went for 2 Nvidia cards ,what would your choice be?I mean relating to cost and performance.
Albert Silver Albert Silver 11/14/2019 04:08
@mamak - You can get performance benefits from even 3 GPUs, though it becomes a case of diminishing returns for more than 2, however there is no need for SLI. They do not need to be connected in any way to work.

@EatMyshorts - Correct, and I explain it has to do with the CUDA code. Note that this is also true of AI development in general where Nvidia cards reign supreme as they alone enjoy acceleration with Tensorflow. I kept the table to consumer cards, but if you want to know what the single best GPU possible for it is, it is the RTX TItan. The previous Titan V was just a fraction faster than a 2080ti, but of course far more expensive. Quadro cards sadly are very poor price/performance for Fat Fritz I have heard, but don't have hard numbers at hand. I will discuss multi-GPU setups in the next article on complete computer configurations.
EatMyShorts EatMyShorts 11/14/2019 02:52
OK...so should one assume that AMD cards are inferior to NVIDIA cards insofar as Fat Fritz' chess performance is concerned? (Interesting if so, as the opposite was the case for crypto mining...not that the two applications are equivalent in any way...but interesting nonetheless...)

Also, if 1 card is good, and 2 cards are better, what benefits would a 3 or 4 card system offer? Are any numbers available showing the amount of gain? Is it linear?

Lastly, is it possible to use NVIDIA Tesla or Quadro RTX cards instead (super expensive, I know), and if so, would there be any benefit to using them over using, say, 2080Ti cards?
mamak mamak 11/14/2019 01:57
Have you tried testing for Nvidia SLI configuration? whats the performance like?