Ask HN: Best (not crippled) laptop configuration for AI/ML/RL? | Hacker News Reader

Ask HN: Best (not crippled) laptop configuration for AI/ML/RL?

3 pointsDrNuke9 comments

We are seeing gaming laptops with RTX 3080 graphic cards these days being sold for just $2k, only for those being crippled and performing like RTX 3060s under stress. We are also seeing 11th gen, octa-core Intel CPUs trying to compete with AMD Ryzen processors, again being crippled from working into pretty thin laptops. It is very difficult to get unbiased reviews by top magazines and reputed influencers, though. Any reliable comparison chart out there? I suspect the best compromise for office & AI/ML/RL in mid 2021, in terms of efficiency and value for money, is a 10th gen Intel hexa core laptop with 32 GB RAM and an RTX 3060 6 GB graphic card? This would cost $1.5k and make the most of its inner limitations. Any help with actual benchmarks? Thanks!

View on HN

I did this type of research last year. I came to the conclusion that you're likely over paying for the portability premium of a laptop with gimped hardware that generally does quite poorly on ventilation. That means your hardware is getting abused by heat, meaning wear and increased likelihood of failure so overall shorter life.

If you want a laptop to actually run AI/ML/RL workloads that would benefit from a graphics card, you're better off optimizing your search for decent specs, no dedicated GPU, portability and good battery life, and spending the money on a powerful desktop setup instead. Your workflow can look like this:

1 - design experiments on laptop or use laptop to remote into desktop with GPU 2 - scale experiments on desktop with GPU till you reach a point where it's maxed out 3 - use Google co-lab with GPU or Paperspace or other similar services (usually offer a free tier for experimentation) 4 - scale on cloud if you really want to parallelize

If you have no choice but to buy only a laptop due to budget, I still think you'd be better off purchasing a laptop without a dedicated GPU and going for points 3/4 above. Seriously, the gains from using an RTX on a laptop are just not worth the cost.

The benefit of using colab or other types of cloud services, is that your investment is low to 0 and will allow you to determine if you actually should invest in your own dedicated hardware.

What's your budget? The usual answers involve workstations like the Dell Precision line and Lenovo's P-series. There will always be some level of throttling if you want portability and power-efficiency. But if you want a desktop in a laptop chassis, battery be damned, there is Sager, Clevo, and their many ODMs/resellers: Origin PC (now under Corsair), Maingear, Eurocom, Digital Storm, Falcon Northewest, among others.

The advantage of these resellers is that they're more responsive to non- standard requests if you really know what you want and are willing to pay for it. In Origin's case, I remember a time where you could have a liquid cooled Xbox 360 or PS3 built right into your desktop computer! I don't know how much things have changed since then, but, with regards to purchasing monstrously powerful laptops, you have as much say as your wallet does.

System76 has some nice Workstation options. You're looking for Thick and capable of Moving Dat Air correct?

I got a Serval I thoroughly enjoy. Even with a 2070 and a Ryzen 3950. I'm not sure as to whether they've got any more up to date AMD offerings slated for the 5xxx line of Ryzen's and 3xxx Nvidia cards; but if they deliver the same build quality, come upgrade time, I know who I'll be looking into for purchasing from. Had no issues with my blowtorch yet.

Bonus point: the cats are more than content to sleep behind or beside it instead of on the keyboard.

Get a desktop. You’d have to kidnap the president of intel and give him the patty hearst treatment if you want the industry to quit with the thin-n-lite cancer.

I get customized, water cooled MSI GT76 (from XoticPC store) with RTX 2080. I don't think it any crippled. Fully charged battery lasts maybe 30 minutes :)

Power supply for this thing is like 2 bricks with cords.

Honestly for AI/ML/RL I would recommend spending your budget on a nice desktop workstation and remoting into it with whatever cheaper laptop you prefer.

Another option is to buy a powerful laptop with thunderbolt, and use an eGPU.

How is straightforward is using an eGPU? Are there issues with drivers, throttling, etc? Also how is the eGPU support on Linux?

I know this is off topic, but I have an M1 16 GB, and an Alienware with 32 GB and a nice GPU. My M1 Mac is snappier, runs better, can handle the same workloads, runs high docker workloads fine. My alienware was 2200 dollars. I'll probably never buy another intel chip laptop. Only draw back is when I need a GPU, I use the alienware. Lately I just ssh in from my Mac and then run the workload.

Just to reiterate, its shocking how much better the M1 chip is.

Reply on news.ycombinator.com