aelaron (e/acc)

presented for your amusement: twitter.com/aelaron_/statu…

@gregjhogan @__tinygrad__ that is awful

@aelaron_ @__tinygrad__ They have progressively restricted P2P on geforce over time to the point where ada lovelace cards have no support at all. It is purely a software limitation, read here: tomshardware.com/news/nvidia-co…

There was an error in the tinybox FP16 tflops on the website. It was stated as 738 TFLOPS. Using JAX, we are getting 774!

@__tinygrad__ cuMemcpyDtoD of the cuda driver api. I unfortunately don't have a second GPU to verify this, but I would be surprised if it goes through cpu. There is unfortunately not much documentation on this function.

@aelaron_ On NVIDIA you can memcpy direct from device to device? I haven't figured out how, are you sure it's not going through CPU RAM?

@__tinygrad__ Also, I'm curious, does tinygrad hit theoretical max. in matmul throughput on AMD yet?

@__tinygrad__ The p2p situation is bad, but you can still mem copy from device to device directly. Does AMD have something better?

@__tinygrad__ 5*180 tflops = 900 The white paper is bs, 162tflops isn't max. Closer to 4 than 5

@aelaron_ Yes, fp32. Similar FLOPS to 5x 4090

In the west, at a young age, we're taught to doubt our own innate abilities. I'm like that too, I'm not sure why. I'm trying to be less like that, even though it seems kinda icky. like, "drop the humble, just brag" was a prescient (now deleted) tweet

32x+ compute multipliers are hiding in plain sight waiting to be discovered just because they aren't easy to express in dl frameworks

@tszzl Least weird OpenAI employee be like

all the people that can make eye contact at openai joined in the last 6 months and they’re making me uncomfortable with their eye contact

@luciascarlet Average apple user be like

I wish macOS felt a bit more "squishy" like iOS does

@zebulgar Unfortunately

lolllll if not for ASML the whole continent would only produce vapidity

@GregSalazarYT Not saying no to a free GPU

The winner of this RX 6800 XT from YouTube never responded. I'll randomly select someone who replies to this post instead. You have 24 hours. Go!😁

@teknium I am certain if OpenAI wins, the future will be uniquely bad. Think of the goal all you want, there is something to be said in general against companies with delusional goals. Open source must win at all cost, even if it means keeping OpenAI in the dark about compute multipliers.

WorldCoin is doing great! Only down 50% since the launch!! Can't believe eyeball scanning UBI crypto is doing so well 🤑

roon wordcel comfirmed

@akbirthko I agree with that. I meant semantic as in "the choice to focus on complexity in implementation rather than the effective operation". A parameterized general operation (which often happen to be simple algorithmically) can be molded into more things.

@akbirthko Seems like a semantic difference in viewpoint then. But just because the programmer doesn't have to think about the complexity, that doesn't mean its not there.

@akbirthko Just because dense matrix multiplication is in principle a simple operation, doesn't mean a parameter configuration cannot perform advanced operations. Think of it as a small FPGA, but differentiable.