ByteDance is developing a custom Groq-style LPU AI inference chip to bypass US export restrictions on high-bandwidth memory

ByteDance is reportedly building its own inference chip modeled on Groq's LPU, the same architecture Nvidia paid roughly $20B to license in December.

The LPU keeps the model in on-chip SRAM and skips high-bandwidth memory. HBM is the component the US restricts most tightly for export to China. ByteDance's memory partner InnoStar fabs at TSMC's mature nodes, which also sit outside the controls.

Each of those choices routes around a US restriction. What's left is the architecture Nvidia just spent $20B to own.

China is increasingly moving toward developing its own chips and is succeeding in becoming ever more independent of the USA.

That is truly impressive.

Source: The Information.

@zephyr_z9 My Internal Plan is as follows📈

⬇️Details as follows

@zephyr_z9 Future trends！

👇

@zephyr_z9 Details of my stock holdings are as follows

⬇️⬇️⬇️

@zephyr_z9 Could the MRAM be relevant to space data centers because of the radiation?

holy shit this is wild

so nvidia drops $20B on groq's architecture and bytedance is already cloning it using parts that dodge every single US export control

the SRAM approach is genius here - no HBM means no export restrictions, mature node fabs at TSMC means no cutting-edge controls apply

this is exactly what happens when you try to control tech through export bans. you just accelerate the other side's innovation and force them to build around you

china's gonna have their own full stack way faster than anyone in DC thinks

#AI #chips #geopolitics #techindependence

@kimmonismus Clever how they're using SRAM to bypass HBM restrictions.

@zephyr_z9 Just everspin or mram in general?

@kimmonismus But this is inference only, right? US export controls are mostly aimed at training capacity for Chinese models, and that's still the HBM-heavy part.

Routing around HBM for inference is clever, but it doesn't really touch the training bottleneck.

@kimmonismus USA Cooked Nvidia from global infrastructure

@kimmonismus sram density is a joke for massive models. this is just a sanctions dodge because they are blocked from buying hbm.

@kimmonismus LPU insight: transformer inference is memory-bandwidth bound, not compute bound. That's why Groq could massively undercut GPU inference costs — ByteDance wants that same economics for Doubao. http://theinformation.com/articles/china…

@kimmonismus Man, that's wild. ByteDance really said we'll just make our own instead of buying. Wonder how long until they actually ship it.

@kimmonismus If ByteDance pulls this off, Nvidia’s moat around HBM could vanish overnight.

@kimmonismus SRAM inference makes sense for their use case, curious how the capacity scales against larger models

@kimmonismus

@kimmonismus Yup. It's the only way to grow the X axis (output tokens) of the token performance pareto curve

@zephyr_z9 You had me a shitco. Long $MRAM again.

@kimmonismus Why ByteDance haven’t been embargoed

@kimmonismus Is Cerebras cooked?