Compared to last year, 2022 is shaping up to be an exciting year for data center silicon.

The new year heralds, among other things, fresh new CPUs, artificial intelligence (AI) accelerators, ever faster networking, and of course a bevy of long-anticipated compute standards.

Here are the top five data center technologies we’ve got our eyes on in 2022.

1. Tantalizing CPUs

Any round up of data center silicon would be incomplete without a couple of shiny new CPUs. With new chips from Intel, AMD, and Ampere on the way, there’s plenty to be excited about.

Intel is expected to be first on the scene, due in part to a last-minute delay to its upcoming Sapphire Rapids Xeon Scalable Processors. The new Xeons are slated for release early this year and boast a number of tantalizing new features, including an all-new, multi-die architecture, Intel’s most advanced 10-nanometer manufacturing process, and on-die high-bandwidth memory.

The launch comes alongside a refresh to AMD’s Zen 3-based EPYC Milan processors, which promise up to 50% higher performance in “targeted technical computing workloads,” the company claims. However, AMD’s main event won’t come until the latter half of 2022, with the launch of EPYC 4, codenamed Genoa.

While little is known about AMD’s next-gen data center chips, we do know it’ll be based on Taiwan Semiconductor Manufacturing Co. (TSMC)’s 5-nanometer manufacturing process and feature up to 96 cores.

“When it’s introduced, we expect Genoa will be the world’s highest performing processor for general-purpose computing,” AMD CEO Lisa Su said during a keynote last November.

Not to be left out, up-and-coming Arm-based data center chipmaker Ampere, will launch its first chips based on a wholly custom core architecture later this year. The chipmaker's previous generation Altra processors were based on an Arm Neoverse core design, which enabled up to 128 cores in a 300-watt power envelope.

“This next-gen part will be significantly more performant than what we have with Altra Max. It’s more cores, it’s a new platform,” Jeff Wittich, chief product officer at Ampere, told SDxCentral in an earlier interview.

Like AMD, Ampere’s upcoming chip will utilize a 5-nanometer manufacturing process. However, the chips will reportedly offer significantly higher core counts. While Ampere hasn’t said how many cores the chip will have, we can expect it to be more than the 128 in Ampere’s Altra Max chips.

2. Gobs of Lightning-Fast Memory

All of the upcoming CPUs add support for DDR5 memory. While this might not sound all that exciting, it’s a boon for data centers running massive, memory-intensive workloads.

As with previous generational upgrades, DDR5 will feature higher per-module capacities over DDR4 while pushing memory transfer rates substantially higher. DDR5 also marks a departure from previous generation memory with on-die error-correcting code as a standard feature and integrated power management for greater efficiency.

To give you an idea of how big an upgrade DDR5 is, Samsung, one of the largest memory manufacturers in the world, teased 512-gigabyte modules with transfer speeds north of 7 Gb/s at Hot Chips 2021.

This means we could see servers with up to eight or even 16 terabytes of onboard memory before long. By comparison, most server-bound DDR4 topped out at just 128 gigabytes and 3.2 Gb/s per module.

3. PCIe 5.0, CXL Open Composable Infrastructure Door

The next generation of CPUs also introduces PCIe Gen. 5.0, doubling the bandwidth available to peripherals to twice that offered by the previous generation.

Last month, Samsung announced its PM1743 PCIe 5.0 SSDs had achieved read speeds of 13 gigabytes per second, write speeds of 6.6 gigabytes per second, and input/output operations of 250 thousand. That’s a massive increase in performance which will allow for significantly faster offload of AI workloads to AI accelerators and NVMe storage.

Beyond higher-performance storage and bandwidth-hungry AI workloads, PCIe Gen 5.0 also introduces Compute Express Link (CXL), which enables a direct memory interface between the CPU and peripheral devices.

One of the early applications for CXL will be to enable large pools of persistent memory that don’t require proprietary interfaces like Intel’s Optane. Instead, system integrators and OEMs can utilize standard PCIe interfaces found on every modern motherboard.

Beyond persistent memory, the CXL Consortium is eyeing opportunities for other peripherals including GPUs, FPGAs, and Ethernet NICs. In theory, CXL could enable a direct memory interface over high-speed Ethernet, enabling data center operators to compose pools of resources without significant performance bottlenecks.

4. 400G Networking Arrives

Speaking of networking, PCIe 5.0 is also essential for the next generation of NICs and data center processing units (DPUs), of which several are slated to launch later this year.

As networking bandwidth increases, PCIe is now the limiting factor. Today’s best NICs are limited to a pair of 100 Gb/s interfaces, which are more than enough to completely saturate a PCIe Gen. 4.0 16x interface.

And while the 400 Gb/s of raw throughput enabled by PCIe Gen 5.0 might sound like overkill for a single server, that hasn’t stopped networking vendors from pushing the envelope ever higher.

Nvidia’s ConnectX-7 NIC, sampling this month, is among the first to take advantage of these bandwidth advances to enable a pair of 200 Gb/s interfaces. Meanwhile, Nvidia’s BlueField-3 DPUs, which build on the ConnectX-7 but target a broader set of input/output (I/O) workloads like networking, security, storage, and virtualization, will launch in May.

Nvidia isn’t the only vendor launching high-throughput DPUs this year. Marvell’s Octeon 10 DPUs are slated to launch in the second half of 2022, boast PCIe Gen. 5.0, and, you guessed it, up to 400 Gb/s of aggregate bandwidth.

5. AI Accelerators Galore

Nvidia’s A100 GPUs dominated the high-performance computing (HPC) space in 2021, with the chip powering many of the most powerful supercomputers in the Top500 and Green500 today.

However, that lead is already under threat, following the launch of AMD’s second-generation Instinct GPUs last November.

AMD’s MI200-series features two GPU dies with a total of 58 billion transistors based on a 6-nanometer manufacturing process. This translates to 220 compute units and 880 matrix cores, a 1.8 times increase over the previous generation MI100.

The GPUs are fed by up to eight on-package, high-bandwidth memory modules for a total of 128 gigabytes of total memory.

On paper, the GPU should have AMD leapfrogging Nvidia in the AI and HPC spaces. AMD’s Forrest Norrod, SVP and GM of data center and embedded solutions, claimed the MI200 is 490% faster than Nvidia’s A100 in highly-precise floating-point 64 calculations. “This is critical for HPC workloads requiring the highest level of precision, like weather forecasts,” he said.

The U.S. Department of Energy’s Oak Ridge National Laboratory plans to use the new GPUs in its upcoming Frontier supercomputer. And assuming Hewlett Packard Enterprise (HPE), which is co-developing the system, doesn’t accidentally wipe the research data again, the supercomputer is expected to deliver 1.5 exaFLOPs of computing power when it comes online early this year.

Of course, don’t count Nvidia out just yet. You can expect Nvidia CEO Jensen Huang, clad in his signature leather jacket, to clap back with the launch of its Ampere Next architecture later this year.

Unfortunately, we’ll have to wait until at least March for an update from team green.