In case you haven't heard, Intel hosted an event for the press today that it called "Data-Centric Innovation Day." The primary purpose of the event was to promote a bunch of new products: the Agilex FPGAs, new SSDs, new Ethernet adapters, and most critically, the Cascade Lake family of Xeon CPUs. We weren't privy to any press materials before the presentation, so this is just a quick overview of all the new hardware on display today.
Obviously, the biggest news of the day is that Intel has launched its second-generation Xeon Scalable processors. These chips are the successors to the Skylake-based platform launched in 2017, and they were supposed to arrive at the end of last year. That didn't happen, but the chips are out now, and according to Intel you should be able to order machine with one as I write this.
The new Xeons haven't showed up on ARK yet, but Intel says there are "over 50" chips in the family. The company was eager to point out during its presentation that that number does not include some portion of bespoke processors. The chips range from single-socket eight-core models all the way up to a 400-W, 56-core behemoth with two CPU dice and a full twelve memory channels. All of the new chips are based on a revised CPU core that includes "DLBoost" instructions to accelerate AI inferencing as well as hardware mitigations for most Spectre variants.
Navin Shenoy holds up a stick of Optane DC persistent memory.
Most of the new chips—but not all—also support Optane persistent memory. That sounds exactly like what it is: regular old DIMMs, but with Optane memory onboard instead of DDR4. As long as the machine has at least one module of DDR4 DRAM, the rest of its slots can be packed with extremely high-density Optane DIMMs. Intel says that this enables up to 4.5 TB of memory per socket, which is certainly extremely impressive, though we're curious to see what the performance implications of such a configuration would be once software is written to take advantage of it.
Besides the big Xeons, Intel briefly showed the next series of Xeon D processors. Like those that came before, the new Xeon D-1600 family uses the same full-fat CPU cores as its heavyweight cousins, but stuffs them into an SoC-style package that Intel says can go as low as 27W. These chips are primarily intended for network applications, and as such integrate Intel's QuickAssist technology for hardware-accelerated encryption. Unfortunately, that's about all Intel had to say on these chips.
Other than Xeons, Intel very quickly announced a fair few non-CPU components. On the storage side of things, we saw a version of the Optane SSD DC D4800X with dual U.2 ports. That feature isn't designed to improve performance, but instead to increase availability. The company also showed the SSD D5-P4326 in EDSFF "ruler" form factor. These drives use QLC NAND flash to cram nearly 31 TB into a stick roughly the size of a typical classroom straight-edge. Intel says you could use these drives to stuff around a petabyte of storage in a 1U enclosure.
On the network side of things, Intel announced the 800-Series Ethernet adapters, codenamed Columbiaville. The add-in cards Intel showed hook up to 100-gigabit Ethernet, which is the sort of transfer rate that us mere mortals can only imagine. Intel talked about several technologies that these controllers offer: Application Design Queues are a form of hardware quality-of-service assurance, while Dynamic Device Personalization—carried over from the 700-series adapters—allows extensive reconfiguration of the adapter without resetting the system.
Finally, Intel showed off its first-ever FPGAs developed and built right at home: the Agilex series. It's an all-new brand for the company, and it describes a series of 10-nm FPGAs that use Embedded Multi-die Interconnect Bridges (EMIBs) to hook up to chiplets offering various I/O and other functionality. These are the only products that Intel announced today which won't be coming to market anytime soon; the very first samples of Agilex will be available in Q3 of this year. Intel claims that these chips will deliver up to 40 TFLOPs of DSP performance.