Hacker News new | past | comments | ask | show | jobs | submit login

One thing I don't see anyone mention is that the M1Max is the probably the cheapest GPU memory you can buy. The only other way to get >=64Gb GPU memory is with the A100 (?) which is like 20k by itself.

So this would be great specifically for finetuning large transformer models like GPT-J, which requires a lot of memory but not a lot of compute. Just hoping for pytorch support soon..




Do you have any evidence one can use >48GB RAM on transformer tuning on the new M1 Max? That would be for me the only reason to buy it as I dislike notch very much, have a beefy 3080-based notebook already and can use 2x3090 for 48GB transformer tuning.


Would it make sense to mine BTC with it?


BTC mining isn't memory-bound, it's compute-bound. The more sha256 hashes you can compute per second, the better. And I'm highly doubtful that any general-purpose hardware at all could even begin to compete with mining ASICs.


Would you or someone happen to know/guess the hash rate of the M1 Max?


I don't think there's a definite answer to this. There's no existing bitcoin miner software optimized for Apple hardware. So, do you use the CPU? The GPU? Both? If you do use the GPU, which API do you use to program it?


At this point it doesn't make sense to mine BTC on any GPU.


Hmm, technically no, an R7 5700G system and 2x 32GB sticks would be around a fifth of the price. That's if you don't need a lot of compute. It's also right around 1/8th the performance of a 3080 but it doesn't have Tensor cores which is big downside in ML.

In theory you could even push the 5700G to 128GB if you figured out a way to get ECC to work on it.


But that’s system memory. Not GPU memory. M1 shares that memory, so it’s addressable by both directly, but with ryzen (and almost every other consumer platform) the cpu and gpu memory are separate.


No, you misunderstand. The 5700G is an APU. Memory is shared between it's GPU and CPU. Hence the "G".


It's unlike 5700G would push 200-400GB/s for GPU tasks, assuming one gets pytorch/tensorflow to use the shared memory and BIOS allows setting such a large shared window, all of them unlikely unfortunately.


Now that's a completely different argument, and still mostly incorrect. The 5700G will access as much memory throughput from its GPU as you can feed it. The limitation is not the GPU, it's how fast you can clock your RAM.

The BIOS doesn't set the shared memory. The BIOS sets the dedicated memory. The shared memory is set by the OS and driver as you need, and the only limit is how much memory you have and how much is used by other processes.

You can force any program to use shared memory by making dedicated memory low. As I said, these programs don't really choose to use it, it's a driver/OS responsibility.

The 5700Gs memory controller indeed can't go above 100GB/s. However 200-400GB/s is not what the M1 Max GPU can do, it's combined performance. You'd have to substract CPU performance. The M1 Max GPU would still be faster of course. But the premise is that GPU performance doesn't really matter.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: