compiled information on Intel’s future Xeon server processors. The upcoming Sapphire Rapids will feature a chiplet design that enables higher efficiency than Ice Lake-SP. It will be equipped with a maximum of 56 cores, using the Golden Cove architecture, manufactured with a 10nm Enhanced SuperFin process, and a TDP of 350W. The new platform also supports PCIe Gen5, CXL 1.1 (Compute Express Link), and eight-channel DDR5 memory, and will continue Intel’s built-in AI acceleration strategy and support Intel’s advanced matrix expansion. Sapphire Rapids is said to have a base clock of 2.3 GHz.
In addition, Sapphire Rapids will be available in two versions. The HBM version will be equipped with four sets of HBM2e, each with a capacity of 16GB, a total of 64GB of memory, a peak bandwidth between 1.432 TB/s and 1.640 TB/s, and will share the socket with the regular version of the Sapphire Rapids processor.
Sapphire Rapids is followed by Emerald Rapids, which still uses the 10nm Enhanced SuperFin process, which is the Intel 7 process, and the overall efficiency will be higher. Emerald Rapids is expected to use the Raptor Cove architecture core, which is an optimized version of the Golden Cove architecture, with an expected 5% to 10% IPC improvement. Emerald Rapids can provide 64 cores and 128 threads, the base clock is increased to 2.6 GHz, it is equipped with 120MB of L3 cache, supports DDR5-5600 memory, and the TDP is increased to 375W. Emerald Rapids will remain compatible with Sapphire Rapids, still using the Eagle Stream platform.
Replacing Emerald Rapids will be Granite Rapids, and Intel will make some big changes. Granite Rapids will be manufactured based on the Intel 4 process (the original 7nm EUV process), the launch time is not clear, it may appear between 2023 and 2024, and it should be a transitional solution. Granite Rapids will use Redwood Cove architecture cores and will increase the number of cores and threads, which is expected to reach 120 cores and 240 threads, with an L3 cache of 240MB. Its base clock is 2.5 GHz, and it offers 128 PCIe 6.0 lanes and 12-channel DDR5 memory. Intel has previously revealed that Granite Rapids will include multiple chiplets in a single SoC, through EMIB packaging, and there will also be HBM and Rambo cache chips.
Starting with Granite Rapids, Intel’s Xeon server processors will feature the latest AVX-1024/FMA3 to improve performance across a variety of workloads, albeit at a power increase. Granite Rapids will use the new Mountain Stream platform and maintain compatibility with next-generation products.
Since AMD launched the first-generation EPYC processor in 2017, it has gradually suppressed Intel in performance in the server market. Intel is pinning its hopes of a turnaround on Diamond Rapids, which is expected to compete with AMD’s EPYC processor codenamed Turin and based on the Zen 5 architecture in 2025. Intel hasn’t revealed much about Diamond Rapids so far. It is expected to use the Intel 3 process (the original 5nm process) and use the enhanced Lion Cove architecture core, which can provide 144 cores and 288 threads, the L3 cache is 288MB, the base clock is increased to 2.7 GHz, and DDR6-7200 memory will be introduced.
There may also be a variant of Diamond Rapids codenamed Sierra Forest, possibly also with HBM memory, for select customers. It will be later than Diamond Rapids, which should not appear until 2026.
