Ｍａｒｋ

Ex-SWE, 2x Co-founder, always building something. Jpeg and compute collector, 4 dgx sparks + 2 rtx pro 6000

[Ｍａｒｋ @0xHyudryu]

Sneak peak of the cooling shroud WIP, noctua 40x20mm PWM fan

[Ｍａｒｋ @0xHyudryu]

New toy arrived. Using this to bridge my RTX PRO 6000 workstation into my DGX Spark cluster via ConnectX-7.

Going to start experimenting with a hybrid GPU + Spark setup to speed up large MoE models on the sparks

[Ｍａｒｋ @0xHyudryu]

@hanindh Ty for the heads up, i’ll experiment with a custom fan shroud and monitor temperatures. Worst case I fry a $1300 networking card lol

[Ｍａｒｋ @0xHyudryu]

@cguldogan Yup! I plan on printing a fan duct for this before installing, so it can get good airflow through the heatsink

[Ｍａｒｋ @0xHyudryu]

@puffybsd Haven't installed it yet, but I will probably print a duct + a noctua fan to get airflow through the heatsink

[Ｍａｒｋ @0xHyudryu]

The meta is using the GPU as the coordinator for attention/KV cache/scheduling/routing etc while the sparks act as distributed workers for the MoE expert compute

Here’s the repo I came across that shows how much faster it speeds things up by: https://t.co/6IwzEQOJfE

Tons of potential going this route

[Ｍａｒｋ @0xHyudryu]

@TechMDAI @wrldsuksgo2mars @JamesMeadlock Yup I got this card because I saw @wrldsuksgo2mars ‘s DS41RT and the numbers are insane 🔥

[Ｍａｒｋ @0xHyudryu]

@0xSero Why not max q

[Ｍａｒｋ @0xHyudryu]

@Morrowmake @Tech2Wild Their margins on sparks were low compared to GPUs, so consumers were getting a hell of a deal on them at the previous price

[Ｍａｒｋ @0xHyudryu]

@Chahatusharma @jun_song They aren't even out yet

[Ｍａｒｋ @0xHyudryu]

@BigOneWD @thsottiaux You have all night to use it, ez

[Ｍａｒｋ @0xHyudryu]

@liu5269 @ItsmeAjayKV Good choice on their end, A3B is too small, better to focus on the mid sized models like qwen 3.8 flash next or even a ~400B sized

[Ｍａｒｋ @0xHyudryu]

@mr_r0b0t He hates anything that's not Claude

[Ｍａｒｋ @0xHyudryu]

@cpaek72 @autonomous_labs It's a dangling carrot for sure, but frontier models getting cheaper with less energy costs? Then why are they cutting our quotas for subscriptions by more and more every time a new model is released?

[Ｍａｒｋ @0xHyudryu]

@SloptimistPrime Thank you!

[Ｍａｒｋ @0xHyudryu]

3 Asus GX10, 1 Nvidia DGX spark + Microtik CRS812 switch. Next step is clustering with RTX pro 6000

[Ｍａｒｋ @0xHyudryu]

@LLMJunky Hopefully Qwen 4 or Deepseek 5 will save us 🤪

[Ｍａｒｋ @0xHyudryu]

@superalesha Having 8gb and 12gb means they got their GPU for gaming, and not for AI. Lowering the model's size and quality just to cater to people who aren't actually using AI in a meaningful way is a waste of resources

[Ｍａｒｋ @0xHyudryu]

@sudoingX After you get 2, the next goal is to have 4, and then you’ll always want more compute 😭

[Ｍａｒｋ @0xHyudryu]

@anitdotguru @TechMDAI Microcenter is real for keeping the prices low 💪

[Ｍａｒｋ @0xHyudryu]

@giadotai 2 years is generous