An Intel Xeon Granite Rapids-SP processor has leaked out, courtesy of YuuKi_AnS. The Engineering Sample (ES) consists of 80 cores and 160 threads, each paired with 2MB of L2 cache. The shared L3 cache comes in at 336MB, up from 320MB on Emerald and a paltry 112.5MB on Sapphire Rapids-SP. It features 48KB of L1 Data Cache (12-way) and 64KB (16-way) of L1 Instruction Cache (16-way) per core. These are “Redwood Cove” P-cores, a minor upgrade from the Raptor Cove architecture powering Emerald Rapids.
Unlike the client version in Meteor Lake, the Granite Rapids-SP P-cores will feature AVX512 support with a dedicated 512-bit wide FMA execution unit (Port 05) and 2MB of Machine Learning Cache. These are just some of the data center-centric optimizations exclusive to the Granite Rapids cores.
The accompanying CPU-Z score of the chip is inconclusive as the cores were clocked at a dismal 600MHz, much lower than the targeted launch clocks (3GHz+). The multi-threaded score of 30,299 points places it on par with AMD’s 64-core Epyc Rome offerings, but as already mentioned, these are preliminary. The final launch scores should be much higher, similar to the 80-core Epyc Genoa parts. Of course, AMD has its army of Turin “Zen 5” chips planned for later this year and the next.