Supermicro has become the latest of the big OEMs to add Arm-based systems to its portfolio, with the launch of its Mt. Hamilton platform based on the Altra line of Arm CPUs from Ampere Computing.
Since Amazon’s Graviton chips aren’t available outside of AWS, and Nvidia has yet to ship its “Grace” CPU superchips, that means these systems have to use Ampere’s Altra and Altra Max CPU families if Supermicro wants to have an Arm server play.
Introduced in 2020, the Altra CPU is available with up to 80 Neoverse N1 cores running at 3.3G Hz. Ampere’s Altra Max, released a year later, boosted the core count to 128, albeit at the expense of a lower 3 GHz operating frequency. Both chips are produced using 7 nanometer manufacturing process from Taiwan Semiconductor Manufacturing Co – that’s the same process used in AMD’s “Rome” Epyc 7002 and “Milan” Epyc 7003 chips.
Unlike its X86 competitors, neither the Altra nor Altra Max chips support simultaneous multithreading (SMT). Ampere contends this as an advantage in cloud-native workloads where it says the 1:1 core-to-thread ratio allows for more predictable performance within a core and across the cores on a die. And it is in the cloud where Ampere has seen some of its largest successes, with Google, Microsoft, Oracle, and Cloudflare having deployed its chips in their respective datacenters. So, it’s hardly surprising that Supermicro would aim its Ampere-based Mt. Hamilton platform at the cloud-native arena.
The systems are available in 1U and 2U single-socket configurations, with support for up to 4 TB of 3.2 GHz DDR4 memory across 16 DIMMs – the max supported by Ampere’s Altra processors.
Like many Supermicro systems today, the Mt. Hamilton platform is modular and supports a variety of storage and PCI-Express configurations, including up to four double-width GPUs or two dozen 2.5-inch U.2 NVM-Express SSDs. For networking, the motherboards are equipped with redundant Nvidia ConnectX4 NICs good for 25 Gb/sec using SFP28 cages.
The air-cooled systems are also designed to operate in extended temperatures ranging from 35 C to 55 C. This means these systems are more than capable of handling the higher datacenter temps prescribed by the American Society of Heat, Refrigerating, and Air Conditioning Engineers (ASHRAE), which top out at 32 C or about 90 F.
Supermicro is far from the only OEM that’s embraced Ampere’s Arm CPUs. Foxconn, Gigabyte, Wiwynn, and most recently Hewlett Packard Enterprise have launched Altra-based systems in either single- or dual-socket configurations.
But with 2023 on the horizon, Ampere’s Altra chips are starting to show their age. They lack support for DDR5 memory and PCI-Express 5.0 peripherals (and thus CXL on the latter). The chipmaker is working on a new processor called AmpereOne that will have more cores and ditch Arm’s Neoverse cores in favor of an in-house design, but details on the upcoming chip remain elusive.
However, Ampere won’t be Supermicro’s only supplier of Arm CPUs for much longer. At Computex this spring, Supermicro showed off an HGX reference server that will use Nvidia’s Arm-based Grace or Grace-Hopper Superchips.
The design, which will be used by six vendors, including Supermicro, Asus, Foxconn, Gigabyte, QCT, and Wiwynn, will begin shipping early in 2023.
Warm-up for Grace. Customers, Mfg lines, integrators, gremlins, get it aaaaalll figured out in advance.
“Unlike its X86 competitors, neither the Altra nor Altra Max chips support simultaneous multithreading (SMT).”
The only ARM or RISC-V CPU core that I know of that supports SMT is the Neoverse E1, and that is a small core not suited for servers and other HPC. Neoverse Nx and Neoverse Vx are all derived from the Cortex-A and Cortex-X cores that Qualcomm, MediaTek, Samsung etc. use in mobile devices and those are all single-threaded. The same is true of the Apple Icestorm and Firestorm cores used in their Ax and Mx SOCs for iOS and macOS devices: single-threaded.
Supermicro is actually using old Ampere Altra CPUs for these servers. Ampere’s latest and greatest is AmpereOne. It isn’t based on Neoverse, but similar to Apple they have a true architectural license – obtained by way of AppliedMicro somehow – and used it to design their own AmpereOne CPUs that are manufactured on TSMC’s 5nm process.
So, Supermicro is probably using Altra because it is cheaper. Or it may be that AmpereOne wasn’t available in time. In either case, AmpereOne doesn’t support SMT either.
In any event, Intel in 2024 is releasing Sierra Forest, a Xeon server CPU product where all the cores are smaller and single-threaded for power efficiency, on Intel 3, a 6nm process. They refer to it as a cloud-native direct competitor to ARM servers, and as such will support up to 128 cores, matching all but Nvidia’s 144 core Grace CPU.
AmpereOne is not available yet, hence it is not yet supported. But Supermicro will no doubt do it.
SMT does mess with deterministic performance. It does present an extra set of security concerns. And it does sometimes boost performance for applications when all of the conditions are right. But if you are going to treat all threads as cores — as many of the clouds do for their VMs — then it does make sense to just have a lot of actual cores — as the Ampere designs do. If you have a workload, like Java plus relational databases, that really likes threads, then SMT makes sense. IBM has eight-way threading on Power8, Power9, and Power10 chips and it really makes these things hum. But IBM knows that 99.2 percent of its customers are doing this kind of stuff, which is a luxury that the cloud providers do not have. They never know what kind of junk code their customers are going to try to run in the cloud.
I completely forgot about the IBM Power series. Which is a great example of why I should just leave this to the experts!
It is curious that you state that Java applications are better with SMT as the cloud providers heavily promote Java-based web applications for the very same serverless platforms that they are going to run in VMs on “cloud native” hardware like AWS Graviton. Azure is currently promoting VMs running on Ampere Altra for application servers, which is clearly going to be primarily Java servers like Apache Tomcat and Oracle Weblogic. So it isn’t merely not knowing what code the customers are going to run, but rather the code that in many cases the cloud providers are advertising and pushing the customers to migrate to and even do original development for their clouds.