Broadcom Ships 25.6 Tbps Network Switch on 7 nm Chip

[AleksandarK]

Broadcom has today started shipping its latest Tomahawk 4 chip for network switches, used by all hyperscale cloud providers like Google, Microsoft, and Amazon. The Tomahawk 4 is a processor built using a 7 nm manufacturing node from TSMC. Featuring over 31 billion transistors, the chip is one of the largest ever Broadcom built. It is an improvement to the Tomahawk 3, which achieved 12.8Tbps. This new chip is having two times the bandwidth available, resulting in 25.6 Tbps speeds while still achieving the form of a single die.

To achieve the massive throughput of 25.6 Tbps, Broadcom put 512 PAM4 SerDes blocks running at 50 Gbps on a single monolithic design. Capable of supporting 64 ports of 400GbE or 256 ports of 100GbE, the Tomahawk 4 is said to bring costs of operations down by 75% due to the less required power and hardware to operate 25.6 Tbps switch, where in the past it was required to run multiple switches to achieve such throughput. Additionally, to control all the telemetry processing and run the switch firmware, four Arm CPUs running at 1 GHz are embedded in die to help out.


View at TechPowerUp Main Site

 

[_larry]

This is a huge achievement. Out-pacing Moore's Law?
Whoa.

 

[Jack1n]

This thing better be super reliable as you can lose 25Tbps of your bandwidth on a single point of failure.

 

[Solaris17]

This thing better be super reliable as you can lose 25Tbps of your bandwidth on a single point of failure.

this is just the CPU the title is a little misleading. But the hardware used in the network core generally has massively long support and life cycle times compared to almost everything else. the switches this would be used in are probably stacked and not alone so a failure should not bring an entire network down, but that has less to do with the switch or in this case the CPU and more to do with network design.

 

[Valantar]

Cool. Now can they make a reasonably priced consumer switch chip providing something more sensible, like 8xnGbE? Or even 4xnGbE+4x2.5GbE?

Infrastructure improving is well and good, but this won't lead to tangible benefits for end users in a long, long time.

 

[SmokingCrop]

This thing better be super reliable as you can lose 25Tbps of your bandwidth on a single point of failure.

Why would you not have this in a redundant setup?

 

[TheUn4seen]

Why would you not have this in a redundant setup?

The short answer is "you must". The long and more correct answer is "you absolutely, positively, never, under any conceivable circumstances, would agree to have such a massively crippling single point of failure. Never." To reduce cost you can use cheaper and slower failover hardware.

 

[candle_86]

Cool. Now can they make a reasonably priced consumer switch chip providing something more sensible, like 8xnGbE? Or even 4xnGbE+4x2.5GbE?

Infrastructure improving is well and good, but this won't lead to tangible benefits for end users in a long, long time.

Why, I'm still happily using my 16 port gigabit switch for my home network, I'm not pushing enough data across it to need anything faster often enough. Even streaming 4k content from my nas to my htpc doesn't use close to the bandwidth, I only get close when I'm doing a network deploy to reimage a customer's computer and that isn't often.

 

[Valantar]

Why, I'm still happily using my 16 port gigabit switch for my home network, I'm not pushing enough data across it to need anything faster often enough. Even streaming 4k content from my nas to my htpc doesn't use close to the bandwidth, I only get close when I'm doing a network deploy to reimage a customer's computer and that isn't often.

You don't use network storage, then. Try editing photos or video off a NAS with GbE, then let me know your opinion.