I am going to buy a new computer. Price range is between 1500 and 3500 USD.

I want something that will let me get lots of playouts/rollouts (?) quickly when replaying a game using Lizzie+Elf. I don't like waiting if I'm quickly reviewing a game, and it takes a minute to get just a couple hundred rollouts/playouts on my current machine.

I slightly prefer a laptop.

I get that more memory, faster CPU, and I guess a good graphics card with a GPU will help. But the best balance to give me good perf for running LeelaZero, etc. is not clear to me.

Any recommendations?

_________________
be immersed

I'm fairly certain that just for machine learning purposes you want to focus on the GPU power as the main factor influencing performance. This is due to, in simple terms, how GPU cores specialize in mathematical operations and, in that respect, overpower the CPU by orders of magnitude. If you want to pour money into the unit, I heard you can't go wrong with Nvidia's 10x0 series, especially 1080Ti. I'm sure someone will post some direct comparisons soon. Other components are not that important, but I always suggest building machines that last, especially nowadays when there isn't going to be any major breakthroughs in performance of units in the home computer range any time soon (one could argue that Ryzens and Threadrippers were such a breakthrough, but those are best in multiprocessing workloads, which, oddly, people still can't program very well, aside from workstation uses like video editing or numerical computation). With this in mind, I would suggest looking at the aforementioned Ryzen processors (very affordable for the performance they offer), some 16GB of RAM and an SSD.

(1080 Ti desktop is about 80% faster than a 1070 mobile.)

If the computer I get has a graphics card with a GPU (eg. gaming laptop), will the GPU automatically be utilized if I run LeelaZero on Windows using Lizzie? Or do I need to configure something?

How about if I run Ubuntu?

_________________
be immersed

On windows Lizzie has separate releases for GPU and CPU, so I don't think you need to configure anything.

GPU - graphics processing unit, what graphics cards are colloquially referred to.
Yes, it will be used automatically, provided you use a GPU version (there are CPU only versions too, they are all labeled, so impossible to mix up).

It is important to be aware of two caveats of doing ML on a laptop. For one, the mobile versions of popular graphics cards are significantly limited in processing power compared to their desktop counterparts (e.g. http://gpu.userbenchmark.com/Compare/Nv ... 5564vs3603). The other thing is that laptops must have some protections in place to prevent overheating, so the moment you start utilizing the GPU at 100%, the power output to the card will be automatically adjusted to prevent battery and component damage, even if you're plugged to mains. If you actually want to get significant results from LZ analysis, it's important to skip laptops altogether and get a desktop PC.

For 3500 USD you can rent an preemptiv instance on Google Cloud with a V100 for 171 days!

One thought: external GPUs apparently work fairly well these days, though I haven't heard if they're a good fit for Go engines. If you're only slightly pro-laptop, that tells me you're pretty comfortable with studying in a fixed location. You could buy a less expensive laptop with a passable discrete GPU for a pinch, and use the eGPU when you're at home.

You mentioned Ubuntu, so I'll caution against the Razer line. I'd say positive things about them if you were only going to use Windows.

_________________
Occupy Babel!

One can buy both a decent desktop (2000) for home use and a laptop (1000) for mobile use.


And how about a smartphone too and then a tablet additionally because the screen of the smartphone is too small for your clunky fingers.

How about something like this?

https://www.amazon.com/MSI-GS65-Stealth ... B07BNZ6Z9M

_________________
be immersed

Good performance and nice looks, why not if you want to spend the money?

I like the CPU performance too on this one, very good for a laptop.


(I use a fanless touch screen notebook for 1000, very nice and silent, good for mobile analyze but not at home. Your choice is much faster and will show a very good performance with GO AI, near desktop performance. Enjoy!)

Of course the lure of a shiny new box is hard to resist.
That said, you may want to run LZ on an Amazon EC2 instance first and find out how much you really need it. I think you are not going to find a mobile rig that will outperform an NVIDIA K80 so Amazon's "p2.xlarge" instance should give you a good idea of what it would be like. For you in the U.S. this would seem to cost less than $1 per hour in their "on-demand" plan and less than $0.30 in their "spot" plan (interruptible). The latter is what the young pros in Japan seem to be using. I heard that they carry their laptops to their study groups and run Lizzie/LZ using the cloud for under 50 yen per hour, which is about p2.xlarge spot pricing for Japan. With your IT background it should be relatively easy to sort out within a day (a lot faster than the decision on what new box to buy ).

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

Good point. In some ways, I kind of just want to spend money somehow. I don't buy things that often, and the allure of buying something new makes me feel good. The LeelaZero thing is kind of an excuse in this regard.

That being said, I do want to be able to review games faster than what I do now - usually I'm busy, and just want to run through the game quickly and not spend a lot of time waiting for the computer to come up with moves.

Cloud seems like a good option.

_________________
be immersed

You have to install the open cl drivers separately which shouldn't be a problem, search google for the suitable driver and install it with apt-get.

Additionally you can install Nvidia xserver settings to view the GPU utilisation.

For cloud, earlier people have quoted more specific prices, but if you want a more general rule of thumb, my experience has been that for computing-heavy tasks (GPU or otherwise), the break-even point is often around 3 months to 1 year, depending on what the task is, whether you're okay with interruptible vs with on-demand, etc.

That is, for the price of buying the machine yourself, you can usually get at least 3 months of the equivalent machine in the cloud chugging continuously 24 hours/day at full power. If you only ever run your heavy computation in short bursts that are nowhere even slightly close to using it continuously 24 hours/day, then cost-wise of course the cloud can easily be much cheaper. If you're running something on a much more sustained basis such as regular overnight computations, or any of a variety of other considerations apply (slightly greater convenience, or you'll get a lot of other value from upgrading from an old machine to something new and shiny, or even something as mundane as wanting things to still work when your internet goes down), of course buying a machine directly can make a lot of sense too.

Also, if you are like me, you keep a computer for 5 years. For comparison with using the cloud, how quickly do more powerful GPUs become available (for around the same price to use, OC)?

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

It is very nice that today we have the choice.
Hope it stays like this for the next decades.

Just saw this in the german dgob forum:

https://shadow.tech/

Maybe this is easier to use, if you are not to tech savy. But I never tried it so...