AI is driving hardware development hard. Do you need a 800GbE 51.2Tbps switch? – Yes you do!

TL:DR High end switches from Marvell now deliver 51.2 Tbps at 512 radix with a current draw of 15KW for a cluster of six switches or 90A that is equivalent to 131,500 kWh per year costing $39,500.00 per year in California or $64,500.00 in Germany.

My Data Sys (MDS) is increasingly working with Power Delivery Systems in Data Centers and the requirements are off the scale! This article is an extract / summary of a really good article published by Patrick Kennedy from ServeTheHome (STH) Youtube channel, one of your favorite hardware and server channels. Here is the link to the article. The Youtube Video from Patrick is at the end of our article.

Our interest, as a Vertiv partner, is in the power consumption of the high performance switches used in modern AI Data Centers. The Marvell Teralynx (previously Innovium) 10 51.2T 64-port 800GbE Switch has 64 OSFP 800GbE ports that can also be used as eight x 100GbE per OSFP port allowing for 512 (512 radix) ports!

Each port us an OSFP module that is larger than the standard QSFP ports to allow for larger heatsinks for cooling. The typical consumption is up to 15W for the 800GpE OSFP modules while the 8 x 100GpE can draw up to 28W each. If you fill each slot with the 8 x 100GpE modules you will draw 1.8kW at idle! This does not include the 500W required by the Teralynx 5 nano meter 10 MV-TX9180-A0 chip and and other components.

These high end switches draw on average 2.5kW of power and in a typical redundant configuration of six, instead of the usual ten switches, the combined power consumption is 15kW (131,500kWh and 90A) or around $39,500.00 per year. In a typical AI data center you could be saving 1mW per year by deploying these high end switches and that easily equates to 2.6 Million Dollars in California (30c/kWh = 8,766,000kWh) and much more in the EU where the cost in Germany is .4 Euros per kWh.

Of course a Switch cluster is only one component in a Data Center and this article is simply to show how much power the switch clusters alone will draw. That is why you need Vertix Power bus and specifically the Vertiv™ Powerbar iMPB system with tap off boxes in your Data Center providing a 1000 Amps per circuit.

We will be publish more articles soon and stand by for a look at Elons Musk’s 100,000x liquid-cooled Nvidia H100s in Memphis and that is a 100,000 of Nvidia’s best chips (for now), the H100 which retails for around US$30,000 each!

Thank you Patrick Kennedy from STH – The YouTube video – Inside the FASTEST New 800GbE 64-port Switch – and please remember to subscribe!