Comments on: Nvidia Shows What Optically Linked GPU Systems Might Look Like https://www.nextplatform.com/2022/08/17/nvidia-shows-what-optically-linked-gpu-systems-might-look-like/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Thu, 25 Aug 2022 16:18:46 +0000 hourly 1 https://wordpress.org/?v=6.7.1 By: Just a Realist https://www.nextplatform.com/2022/08/17/nvidia-shows-what-optically-linked-gpu-systems-might-look-like/#comment-196241 Sun, 21 Aug 2022 16:25:07 +0000 https://www.nextplatform.com/?p=141050#comment-196241 HP Labs has publicly put forward the same vision, graphics, etc. regarding photonics for at least 10 years now. They’ve made significant strides in research, prototypes, etc., but have failed to commercialize it. If you replaced “Nvidia” with “HP Labs” you’d be hard pressed to see any difference (really no surprise as HPE has suffered significant brain drain to Nvidia’s benefit). Intel has been pushing its silicon photonics approach for nearly as long, but with, at best, tepid success as there are many shortcomings. In general, silicon photonics as put forward by HP Labs / Nvidia continues to suffer from repeated industry leadership failures and intestinal fortitude when it comes to committing the resources and funding to create a viable market.

The confluence of climate change, high energy costs, the urgent need for multi-trillion dollar global energy infrastructure upgrades / replacement, and growing national security demands in multiple dimensions make it abundantly clear that now is the time to invest and drive high-volume adoption of silicon photonics technology in board-to-board and ultimately, for chip-to-chip communication. The high-volume SFF TA-1002 connector was designed to support photonics from its inception which, in turn, enables multiple SFF memory, storage, and I/O form factors to support photonics which, in turn, enables multiple existing enclosures to support photonics (sometimes architects with enough foresight do get it right). The net is the technology and infrastructure to transparently insert photonics into nearly any solution exists today. All it requires is for industry leadership to invest and commit in earnest. Sure, some volume can be achieved in nice Ethernet technologies like HPE Slingshot or high-end GPGPUs from Nvidia, but that will never be enough to drive costs to achieve ubiquity. Industry and government leadership needs to look beyond simple BOM and OPEX-based accounting decision making, and instead focus on how photonics can drive new performance and efficiency capabilities as well as new solution architectures that fundamentally change how energy is managed and consumed. Doing so will yield so much more than simplistic bottom-line gains.

To quote Youtube’s Beau, “It’s just a thought.”

]]> By: Paul Berry https://www.nextplatform.com/2022/08/17/nvidia-shows-what-optically-linked-gpu-systems-might-look-like/#comment-196083 Fri, 19 Aug 2022 13:40:11 +0000 https://www.nextplatform.com/?p=141050#comment-196083 Having seen almost exactly this same presentation in 2004, I agree this is likely where we’re heading, but any promise of when we get there is pure speculation.

]]> By: vang https://www.nextplatform.com/2022/08/17/nvidia-shows-what-optically-linked-gpu-systems-might-look-like/#comment-196019 Thu, 18 Aug 2022 14:42:41 +0000 https://www.nextplatform.com/?p=141050#comment-196019 ..its a little bit unclear why logic-to-logic is 100Tbps and logic-to-memory is 16Tbps in this particular example. Is logic-to-logic on the same die/chiplet, while memory goes through an interposer?

also there is a small typo: “The goal for co-packaged optics using DWDM is to have lower power consumption than an electrical cable but with a similar cost, have a reach comparable to an active electrical cable, and offer signal density that is on par with a printed circuit board.” – the ‘active electrical cable’

Great read though!

]]>