Sabitlenmiş Tweet
Why is @MarvellTech partnering with @nvidia and supporting NVLink?
My hypothesis: Because @AWS wants them to.
Marvell is the XPU design partner for Trainium, and AWS announced back in December that Trainium4 would support NVLink. So this feels like Nvidia saying, “Yes, please Marvell, go ahead and use NVLink. Happy to help you”.
Vik the interconnect guy can't help but point out the press release mentions “the companies will also collaborate on silicon photonics technology.” Could that involve Celestial AI’s photonic fabric? Could there be a future where NVLink scales up over optical links instead of copper? And if so, could Marvell/Celestial’s photonic fabric show up in future Trainium chips?
After all, in Marvell’s press release about the Celestial acquisition, AWS was on the record saying they “believe optical interconnects will play an important role in the future of AI infrastructure” and that “Celestial AI has made impressive progress, and we expect their combination with a large-scale semiconductor company like Marvell will help further accelerate optical scale-up innovation for next-generation AI deployments.” Maybe a Trainium4 (or 5+?) will scale up using a photonic fabric? Good take @vikramskr!
So there is definitely a believable angle about customer demand pulling this Nvidia/Marvell partnership into existence.
This is a great move for Nvidia and helps them participate in the expanding XPU TAM. But it’s more than just getting a cut of the XPU TAM; it’s another step in the multi-silicon datacenter story.
If I were Nvidia, I’d frame conversations with hyperscalers like this: hey AWS, if you’re buying Vera Rubin GPUs (and maybe some @GroqInc LPUs?), why not deploy some racks of Trn4 in the same datacenter hall? Same MGX rack design, same Spectrum-X scale-out fabric, same management plane. Route cheap inference to the XPUs, send the hard reasoning workloads to our GPUs. And while you’re at it, maybe you look into our storage racks for hardware-accelerated KV-cache, Vera CPU racks for orchestration, BlueField DPUs… it's about Nvidia owning the broader "walled garden" and letting silicon in. That's not a knock on Nvidia --- the value prop to the customer is that vertically integrated datacenter has the best performance compared to a mix and match, less coherent approach.
So this NVLink fusion + XPU play is making the whole datacenter floor Nvidia-native so everything speaks the same language, even when the compute is heterogeneous, and even if the physical medium shifts to optics in the future. $MRVL $NVDA
youtube.com/watch?v=oWQG20…
YouTube
English
