Need for small port count 200Gb/400Gb Network Switches High Performance Computing small 3 or 4 person AI Research labs or any small business with next generation Servers and NVMe PCIe Gen4 SSD.
In today’s market where you have very small businesses working in 4 or 5 person AI Deep Learning Research Labs, they have very high Networking Performance Requirements.
The Neural Network modeling they are using will be writing out huge amount of temporary (intermediate compute sets) to very high-speed Server Internal Storage which is very costly and expensive so these set must be written out to Flash Arrays to keep for post modification. Once the modeling/learning solutions have been found then typically you’d purchase all the data up to 1.5 TB from very expensive internal U.2 SSDs and get ready for the next run.
It turns these small AI Research Labs are using “Data Center Level” Servers like the Dell PowerEdge R750 because when those are configured with 14 Intel Optane U.2 P5500 SSDs (NVMe PCIe Gen 4) SSDs configured in RAID 10 the write out speed to external Flash Array will be at approx. 24,500MB/s which means they would be required to use a Mellanox or equivalent 200Gb NIC card to meet that write out rate.
However, the Server in our case the two R750’s needs to connect to a Network Switch through their 200Gb NIC cards and of course the Flash Array needs to connect to the Network Switch through its 200Gb NIC card.
So, it’s really nice that Cisco can produce huge VM farm corporate conglomerates network switches with 32, 48 and 192 200Gb port network switches and beyond. But they are entirely missing a huge new area of business coming online primarily because of AI Deep Learning intense compute performance and many businesses have had to downsize after COVID-19 to very small research labs. With a need for 4 & 8 port 200Gb switches and even 4 port 200Gb and 4 port 400Gb Network Switches.
Cisco obviously hasn’t realized that the AMD (EPYC 7763 processor) or Intel’s 3rd (Platinum 8380 processor) with 40 cores are here today. These servers are shipping to small labs with PCIe Gen 4 X16 lane slots loaded with very high speed NIC cards that are required for today’s Hyper Computing application.
Cisco “should” be producing specifically for small business high performance sector specifically a line of very high-performance small port count Network Switches examples below:
- Minimal multi-gig handling switch configuration: 4 (four) 25Gb ports + 4 (four) 100Gb Ports + 2 (two) 200Gb ports. That’s when using slow Intel Optane P5500 SSDs and only in RAID 10 RAID 0 would easily saturate 200Gb NIC card.
- More useful 4 (four) 200Gb ports and 4 (four) 400Gb ports. When using high performance Intel Optane P5800X SSDs RAID 10 with on 14 U.2 NVMe SSDs.
This of course already would fall way behind where they need to be since if you put 16 of these Intel P5800 SSD in RAID 0 and easily saturate 400Gb NIC cards by over 60%.
Detail The server middle performance R750 mentioned above with crippled (slow 3500MB/sec write) Intel P5500 SSDs would be ok for 200Gb network switch. However, the Server equipped with Intel Optane Performance U.2 SSDs (P5800X) which have a writing out speed of 6,200MB/sec when using seven (7) SSDs in RAID 0 will write out at: 43,400MB/sec which is equal to 347,200Gbs which would require a 400Gb NIC and Network Switch. Remember we are talking a big emergence of post COVID-19 very small high performance AI Research and development labs.
Cisco should be keeping up with NVIDIA – because NVIDIA has already learned how to bring giant performance to individual users. Cisco isn’t even capturing Small Business users. Search for Intel Optane P5800X Product Brief. Unable to provide link in the text.
To understand where technology is today. BTW this same capability will soon be coming to the individual Workstation Owner. So, do you think the individual PC owner needs a 48 port 400Gb Network Switch for? Start thinking small. COVID-19 taught everyone else I guess Cisco didn’t here.
