Cisco challenges Broadcom, Nvidia with a 102.4T switch of its own
As AI training and inference clusters grow larger, they require bigger, higher-bandwidth networks to feed them. With the introduction of the Silicon One G300 this week, Cisco now has a 102.4 Tbps monster to challenge Broadcom's Tomahawk 6 and Nvidia Spectrum-X Ethernet Photonics.
Much like those chips, the G300 packs 512 ultra-fast 200 Gbps serializers/deserializers (SerDes). The massive radix — that means loads of ports — means Cisco can now support deployments of up to 128,000 GPUs using just 750 switches, where 2,500 were needed previously. Alternatively, those SerDes can be aggregated to support port speeds of up to 1.6 Tbps.
None of this is unique to Cisco, however. That's just how bandwidth scales. Those same figures apply to Broadcom and Nvidia's 102.4 Tbps silicon just as they do to anyone else's.
Managing AI congestion
According to Cisco fellow and SVP Rakesh Chopra, what really sets the G300 apart from the competition is its collective networking engine, which features a fully shared packet buffer and a path-based load balancer to mitigate congestion, improve link utilization and latency, and reduce time to completion.
"There's no sort of segmentation of packet buffers, allowing packets to come in [and] be absorbed irrespective of the port. That means that you can ride through bursts better in AI workflows or front-end workloads," he said.
The load-balancing agent "monitors the flows coming through the G300. It monitors congestion points and it communicates with all the other G300s in the network and builds sort of a global collective map of what is happening across the entire AI cluster," he added.
This kind of congestion management isn't new by any means. Both Broadcom and Nvidia have implemented similar technologies in their own switches and NICs for this reason.
However, Cisco claims its implementation achieves a 33 percent better link utilization, which can cut training times by up to 28 percent compared to packet-spraying-based approaches.
And while Chopra doesn't say which vendor it's comparing the G300 to, we'll note both Broadcom and Nvidia's implementations rely on packet spraying. As usual, we recommend taking any vendor-supplied performance claim with a grain of salt. While there aren't that many 102.4 Tbps switches out there, there are countless ways to build networks using them, and some topologies may benefit from one vendor's tech more than another.
P4 programmability
Cisco's collective networking engine is only one piece of the puzzle. The other, claims Chopra, is the chip's P4 programmability.
"It means that we can take our device, we can reprogram it to add new functionality, new capabilities, and deploy the same equipment in multiple different roles," Chopra said, adding that this is particularly valuable for extending the useful life of the switches. As new features are introduced, they can often be added via a software update rather than requiring new hardware.
If any of this sounds familiar, that's because Cisco isn't the only AI networking vendor leaning into P4. AMD's Pensando NICs, like the Polara 400, also make use of the programming language. This actually came in handy for AMD as it allowed the chip designer to start shipping Ultra Ethernet compatible NICs before the spec had actually been finalized, since any changes to the spec could be implemented later via a software update.
As with past Silicon One switch chips, the G300 will compete directly with Broadcom in the merchant silicon arena in addition to powering Cisco's own networking appliances.
In particular, Cisco says the part will be available as part of its N9000 and Cisco 8000 product lines, both of which will come equipped with 64 1.6 Tbps OSFP cages.
1.6 Tbps optics have arrived, but still no CPO from Cisco
To support these appliances, Cisco is also rolling out new 1.6 Tbps pluggable optics, which can be broken out on the other end into up to eight 200 Gbps connections.
Alongside the ultrafast links, Cisco is introducing 800 Gbps linear pluggable optics (LPO), which ditch the onboard digital signal processor (DSP) and retimer in order to cut power consumption. This is possible because signal processing is handled entirely in the G300.
While pluggables don't consume that much power – usually 10-20 watts – a 50 percent reduction is still significant when you consider just how many pluggables may be deployed across a datacenter.
Cisco wouldn't tell us how much power its new LPO transceiver consumes, but did say that when combined with the N9000 or Cisco 8000 systems, customers can expect to see a roughly 30 percent reduction in switch power.
While Cisco is making gains on pluggables, it's not quite ready to follow Nvidia and Broadcom and take the plunge on copackaged optics (CPO) just yet.
The technology involves moving the optics from the pluggable module into the switch package, reducing power consumption and, in theory, improving the reliability of the network.
"We're not making any public announcements about CPO at this time," Chopra said, noting that Cisco has previously demoed the tech on its G100 silicon. "We certainly have the technology, we're looking for business alignment to sort of productize it."
Cisco expands availability of multi-datacenter routing silicon
Alongside the new optics, Cisco also announced broader availability for its Silicon One P200 routing silicon.
We took a closer look at the chip last fall, but in a nutshell, the 51.2 Tbps chip is designed to connect AI training clusters over distances of up to 1,000 kilometers.
Alongside the previously announced Cisco 8223 router, Switchzilla says the part is now available in its N9000 line and as a 28.8 Tbps line card.
Cisco's G300 chips, systems, and optics are slated to begin shipping later this year, just in time for Nvidia and AMD's next-gen rack systems. ®
