Last month Intel expanded its Xeon 6 lineup with new models. Whereas the Intel Granite Rapids family members launched last year strutted 64 or more CPUs and 8 or 12 DRAM channels, the newer models have as few as eight CPUs and only 8 DRAM channels. Last year’s high-core-count (ultra-core-count or UCC in Intel’s parlance) targeted hyperscalers, seeking to regain ground lost to AMD. The newer devices target enterprise customers, a group Intel has done better holding onto as its manufacturing operations stumbled.
Special Models
Beyond scaling down, the new models include processors with distinctive features.
- For workloads such as in-memory databases and scientific visualization requiring both many cores and more memory bandwidth than is achievable with the XCC chips, the expanded lineup includes processors supporting four- and eight-socket motherboards. It’s a niche AMD doesn’t address because Epyc scales to only two sockets.
- For designs requiring prodigious I/O, Intel added models supporting 136 PCIe lanes. Customers can choose SKUs with 16–80 cores; the low end is for designs such as storage servers with a lot of peripheral devices but requiring little computation. These models more than close a gap AMD opened by offering Epyc processors mating only a few computing dice with its standard I/O die.
- Disclosed last year, Xeon SoC (formerly known as Xeon D) integrates a different I/O tile than other Xeon 6 models and targets mobile infrastructure, wired communications, and embedded designs. This tile provides 200 Gbps of Ethernet connectivity, video transcoders, QAT (not the narcotic leaf but the encryption and compression engines found in other Xeon 6 models), and vRAN Boost. The lattermost is an engine performing forward error correction (FEC), a computationally intense signal-processing function included for Intel’s mobile-infrastructure customers, such as Ericsson. Intel’s main Xeon SoC competitor is Marvell, which offers Octeon and serves OEMs such as Ericsson rivals Nokia and Samsung.
Two Asides
- Intel also launched Xeon models based on its Raptor Lake PC processor but with the less-powerful E-cores disabled and ECC enabled. This branding exercise may benefit OEMs offering pedestal servers.
- Intel points out Xeon 6’s superior performance on AI tasks compared with Epyc. Intel’s advantage stems from Xeon’s AMX matrix-math unit. While it boosts CPU-based inferencing, AMX adds manufacturing costs to every Granite Rapids chip and requires software to use AMX instructions.
Bottom Line
Intel is refreshing its enterprise Xeons as AMD is training its sights on the market. Corporate IT managers and the OEMs that supply them are conservative and less prone to switching suppliers. Moreover, they’re less inclined to buy the expensive many-core processors that have differentiated AMD Eypc from the incumbent Intel Xeon. Thus, this market is Intel’s to lose, and the new mainstream Xeons will help the company defend its position. Meanwhile, the company remains unchallenged for four- and eight-socket designs. As for Xeon SoC, Intel faces not so much competition but the consequences of poor execution and limited Open RAN uptake.
The company has been manufacturing challenged, and the new Xeon models are a positive sign. Fabricated in the Intel 3 process, they indicate it has matured to the point where the company can economically produce high-volume mainstream chips. At the same time, the Granite Rapids family is another Xeon generation reflecting a questionable chiplet strategy. Whereas AMD Epyc uses a regular organic packaging substrate, Intel uses a costlier albeit more capable silicon substrate. Intel bears further costs by having three different computing tiles and two different I/O tiles. Manufacturing costs should be a concern as the company struggles with profitability.
On balance, the Granite Rapids rollout has shored up Intel’s data-center business. By the company’s own admission, it must do at least as well delivering the next generation (Diamond Rapids) to restore customer confidence. Although the company has faced criticism for failing to capitalize on the AI boom, Intel’s priority must be defending its core PC- and server-processor businesses by successfully introducing new products.