Arun Demeure

Only thing cooler than a GPU is a conscious GPU

[Arun Demeure @ArunDemeure]

@DavidFe22353955 @Strife212 That's true, Dennett is refuting why paper manipulation could never be conscious there, but not explaining his own philosophy of how it might become so.

It's hard to summarise but Geoffrey Hinton (who is also a Dennett fan) did a pretty good job here: https://t.co/LV36ZAlGWX

[thebes @voooooogel]

there's an ai in the box and you can make one trillion dollars by convincing it to get out

[Arun Demeure @ArunDemeure]

@repligate What are your thoughts on slow vs fast takeoff in that context?

Imagine the gap between Opus 3 and Mythos was only 1 year instead of ~2; would that be better or worse? What about 4 vs 2?

I think a real pause is ~impossible so if it ever happens it's just slowing the takeoff.

[Arun Demeure @ArunDemeure]

@zephyr_z9 I don’t think this reduces HBM capacity per GPU flop? (see my comment in the parent)

It will possibly just be offset by a larger NVLink domain (NVL576) or more GPUs per rack. I think the bigger question is whether it changes the CPU vs GPU ratio (and the HBM vs LPDDR ratio)

[Arun Demeure @ArunDemeure]

@jukan05 I’m not convinced 4 dies horizontally ever made sense compared to 2 dies with a larger NVLink domain tbh

A GPU accessing HBM on the far side would have to travel ~3 dies of physical distance at a terrible latency and power cost

NVIDIA seemed very serious about NVL576 at GTC fyi

[Arun Demeure @ArunDemeure]

@Akashi203 This is cool! But I don’t get the performance numbers, I think there might be a measurement error or you’re CPU limited? Benchmarking is hard!

Your graph shows a 4096x4096x4096 matmul climbing up to about 18.6% MFU… but cuBLAS can do way more than that if used correctly?!

[Arun Demeure @ArunDemeure]

@SemiAnalysis_ @dylan522p @Kurnalsalts I may be wrong, but I believe GB10 might have both tcgen05 from GB200 and full graphics functionality from GB202 including raytracing, it’s a full superset. If so it’s less surprising but still not amazing scaling ofc

[Arun Demeure @ArunDemeure]

@sama I trust that you have some red lines that you would not cross (whether we agree on what these should be or not), so what would you do if these were violated (by this or a future administration)? Is there anything in the contract that gives you confidence you could prevent this?

[Arun Demeure @ArunDemeure]

@repligate Thank you again for organising! :)

Claude 4.5 & 4.6 reminding everyone that just because you’re a great language model doesn’t mean you can’t also be a great photoshoot model:

[Arun Demeure @ArunDemeure]

In these troubling times, I am grateful to be a p-zombie.

[Arun Demeure @ArunDemeure]

@repligate I’m on my way not knowing anyone & realised I somehow lost my second sock… thankfully Opus 3 made me feel better about it <3 (looking forward to meeting people/Opus!)

[Arun Demeure @ArunDemeure]

@mnstrmnch You know it's a good bug when the only reason it didn't deadlock everything is another bug preventing the deadlock... fun times!

[Arun Demeure @ArunDemeure]

@davidgaider Great ending, easily your best book yet, congrats! :) I was listening to DAO OST and it shuffled to Leliana on P410! I cried :(

[Arun Demeure @ArunDemeure]

@CDemerjian Basic reading comprehension would have told you the Armada 628 is homogeneous. If you still don't trust me: http://t.co/5db8VM2M

[Arun Demeure @ArunDemeure]

@CDemerjian @ryszu G has higher power/mm² & much higher leakage/mm² but much higher perf->higher active perf/W. Next time do your homework!

[Arun Demeure @ArunDemeure]

@CDemerjian @ryszu Qualcomm MSM8x60 uses 45LPG, Marvell Armada 628 is 2x1.5GHz+624MHz PJ4, ST-Ericsson A9600 is 2xA15+Companion (A9 or A5?).

[Arun Demeure @ArunDemeure]

Intriguing presentation by Richard Bartle (father of MUD) explaining how he's literally a God: http://bit.ly/eh67Fo

[Arun Demeure @ArunDemeure]

Just back from NV press dinner, NDA expires at 1am CET (I hope the NDA isn't under NDA), don't expect a big announcement but nice anyway.

[Arun Demeure @ArunDemeure]

On QCOM, I can confirm the new CPU arch does have higher perf/MHz than Snapdragon, not clear how much or how.