Supermicro announced an Arm-based server – ARS-210M-NR – in December 2022 and Servethehome take it for a spin, load it with four Nvidia A16 cards. The test server came with a 128 core Ampere Altra Max Arm processor, running at 3GHz, with 16 DDR4 memory modules (512GB memory in all), two SPF28 ports with 25GbE support, 16 trays 2, 5 inches and a pair of 2Kw power supplies, all nicely assembled in a tool-free chassis.
The cherry on the cake is a quartet of Nvidia A16 graphics cards based on Ampere. Each card comes with 64GB of GDDR6 ECC memory and four GPUs housed in a maximum-height full-length, passively cooled PCB. The A16 isn’t the most powerful Data Center card Nvidia has; that would be the A100; however, it gains traction for service providers who want to get the maximum number of concurrent users on a table.
In terms of pure performance, each of these GPUs has half as many CUDA cores as the Geforce RTX 3050, so it should be able to run the legendary Crysis game without too much of a problem, which means in terms of performance. theoretically, you could have 16 multipliers. iterations of one of the most engaging games ever run concurrently. You only need to make 16 virtual machine and download 16 copies of cry.
No pricing details are provided but assume that The A16 retails for around $3,000 and a fully loaded server with the same login costing around $10,000, you would have to shell out $22,000 for such a system.
Powering the cloud
Now who would need such a depraved power? For starters, cloud gaming is mainstream right now (even if Google bailed out earlier this year when it was bogged down stadium), and it’s things like the ARS-210M-NR that make that possible.
Beyond this, however, it’s about VDI (Video Desktop Infrastructure) for business and enterprise. the rise of future work act as a catalyst for the adoption of Virtual client (and virtual PC) like homemade has become a reality for millions of users. And only one of these servers can hold many virtual machines (virtual machines); a recent experiment by a cloud computing company Nutanix shows that two Nvidia A16 graphics cards can handle 128 virtual machines, which means four cards should be able to run 256.
Of course, you may need to implement a more powerful CPU than the Altra Max and more memory (Nutanix went with 1.5 TB of RAM and two Intel Xeon Gold 6354, each with 18 physical cores). Increased SaaS support for virtual GPUs across multiple applications (e.g.: photoshop or Chrome browser) explains why dense servers with very high GPU counts are becoming popular.
Other apps like web hosting can benefit from a high CPU core count, that’s why VPS (virtual private server) now very affordable with the price gap with dedicated server and bare metal high residual solution.
Servethehome was running a 2U server on Ubuntu and while the experience wasn’t as smooth as it should be, it’s definitely worth the effort. Arm is slowly becoming a formidable competitor to both Intel’s Xeon and AMD’s Epyc, and with Amazon getting ready for its fourth generation. gravitation CPU family, there has never been a better time to try Arm.