Cisco has piled on the AI networking bandwagon, becoming a member of Broadcom and Nvidia with a 51.2Tbit/sec change it claims is able to bringing collectively a minimum of 32,000 GPUs.
The change ASIC, codenamed the G200, was developed below Cisco’s Silicon One portfolio, and is focused at bandwidth-hungry web-scale networks in addition to bigger AI/ML compute clusters.
The chip itself affords twice the bandwidth of Cisco’s older G100 ASIC, doubling the variety of 112Gbit/sec serializer/deserializers (SerDes) from 256 to 512. This enables for as much as 64x 800Gbit/sec, 128x 400Gbit/sec, or 256x 200Gbit/sec ports, relying on the appliance and port density desired.
Realistically, we anticipate the majority of switches powered by Cisco’s G200 to cap out at 400Gbit/sec – that is the utmost bandwidth supported by PCIe 5.0 NICs as we speak, and there simply aren’t that many functions – aside from aggregation – that may reap the benefits of 800Gbit/sec Ethernet within the first place.
If you happen to’re feeling a little bit of déjà vu at this level, that could be as a result of we have seen related ASICs and switches from Broadcom and Nvidia, with their Tomahawk 5 and Spectrum-4 traces, respectively. Each switches boast 51.2Tbit/sec of bandwidth and are positioned as alternate options to InfiniBand networks for giant GPU compute clusters.
The truth is, the Cisco G200 guarantees lots of the identical AI/ML-centric options and capabilities promised by rival community distributors. All three promise issues like superior congestion administration, packet-spraying strategies, and hyperlink failover.
These options are necessary, as along with a GPU with the ability to absolutely saturate a 400Gbit/sec hyperlink, the best way workloads are distributed throughout these clusters makes them notably delicate to latency and congestion. If site visitors will get backed up, GPUs may be left sitting idle, leading to longer job completion occasions.
All three distributors will let you know that these options, mixed with one thing like RDMA over Converged Ethernet (RoCE), can obtain extraordinarily low-loss networking utilizing customary Ethernet and, by extension, diminished completion occasions.
“There are clearly a number of 51.2Tbit/sec switches in the marketplace. My view of the scenario is that not all 51.2Tbit/sec switches are created equal. Lots of people make claims; few individuals ship on these claims,” Rakesh Chopra, who heads up advertising and marketing for Cisco’s Silicon One line, instructed The Register.
Whereas Broadcom and Nvidia might have overwhelmed Cisco to market, Chopra argued one of many G200’s key differentiators is assist for 512x radix configurations. With out moving into the nitty gritty, a bigger radix means smaller, tighter change materials.
Cisco claims this permits the G200 to scale to assist clusters in extra of 32,000 GPUs – or round 4,000 nodes – utilizing 40 p.c fewer switches and half as many optics in comparison with an analogous 256x radix change. From what we are able to inform, this can be a shot at Broadcom’s Tomahawk 5, which highlights a 256x 200Gbit/sec radix, although it might additionally apply to Nvidia’s Spectrum-4.
Cisco contends its Silicon One G200 ASIC permits for denser, extra environment friendly networks, as depicted on this provided graphic – Click on for greater
Chopra claimed Cisco is in a position to do that as a result of Silicon One’s applied sciences are “massively” extra environment friendly than competing switches. How rather more environment friendly, we do not know – Chopra would not inform us how a lot energy the G200 makes use of.
He did say, for a community topology able to supporting 32,000 GPUs, utilizing G200-based switches would translate into practically a megawatt of energy financial savings in comparison with competing merchandise. Whereas that will sound spectacular, you in all probability would not discover given a cluster of that dimension is apt to drag greater than 40MW below load, and that is earlier than you take into accounts datacenter cooling.
Even nonetheless, Chopra contended that any financial savings are worthwhile, particularly for corporations whose sustainability targets mandate offsetting the facility they use.
Nevertheless, in an e-mail to The Register, Ram Velaga, SVP of Broadcom’s core switching group, argued the instance described by Cisco was unrealistic and would by no means be deployed in the actual world.
“For instance, they present a 100Gbit/sec hyperlink within the material transitioning to a 400Gbit/sec hyperlink to the GPUs, which suggests that you would not use cut-through operation. You’d as an alternative want to make use of the next latency store-and-forward operation,” Velaga wrote.
Velaga added that Tomahawk 5 can the truth is assist a 32,000 GPU cluster utilizing two-tier community utilizing the 768 switches – the identical quantity as Cisco.
We additionally reached out to Nvidia for remark, and had not heard again on the time of publication.
Silicon One ASICs are presently in clients’ palms for integration into finish merchandise. Nevertheless, Chopra would not decide to a timeframe once we can anticipate to see finish merchandise make their means into the datacenter. ®
