Intel has announced the latest performance and internal forecasts for its upcoming 5th generation Emerald Rapids and next-generation Granite Rapids Xeon CPUs. During the SC23 conference, Intel unveiled new performance metrics for its Xeon CPUs, including actual results for the upcoming 5th generation Emerald Rapids and predicted results for the next-generation Granite Ridge Xeon CPUs. The chipmaker compared these results to its previous-generation chips, and also showed HPC metrics comparing Xeon Max CPUs to AMD's PYC Genoa 96-core chips.
Intel's 5th generation Emerald Rapids CPU is pin-compatible with the Sapphire Rapids CPU on the EagleStream platform. They use the same LGA-4677 socket and primarily offer an optimized design with a smaller number of cores/threads, larger cache pools, and a focus on higher performance per watt.
The company emphasized that Emerald Rapids Xeon CPUs will provide workload-optimized performance and energy-efficient computing. At the same time, the addition of faster DDR5-5600 memory will bring higher bandwidth and transmission speeds. The fifth-generation Xeon will also provide CXL1.0 and 2.0 support, as well as 80 PCIeGen5.0 lanes. Intel claims that performance will increase by 40%. Details are as follows:
Comparison between Intel Emerald Rapids (Xeon 8592+64 cores) and Sapphire Rapids (Xeon 8480+56 cores):
Artificial intelligence speech recognition: 1.4 times improvement
HPCLAMMPS (Copper): 1.4 times improvement
Media transcoding (FFMPEG): 1.2x improvement
Some of the features of the fifth-generation Xeon "EmeraldRapids" CPU include:
3x faster LLC cache
Faster memory speed
More CPU cores (up to 64 cores)
CXL Type 3 High Bandwidth Interface
Workload Optimization (Intel AE)
Optimized power mode
General performance/power consumption improved by up to 17%
Utilize Intel Accelerator Engine to offload CPU cores and improve energy efficiency
Dramatically improve inference and training performance
Intel AMX for built-in AI acceleration
Optimized software stack deployment out of the box
Intel also announced internal performance predictions for its upcoming Granite Rapids Xeon CPUs. This series of processors will only have P cores, while Sierra Forest will add E cores and provide up to 288 cores. The chips will be compatible with the next generation BirchStream platform, which has two different sockets: LGA4710 and LGA7529. The company confirmed that Granite Rapids Xeon CPUs will offer increased core counts, frequencies and the latest Intel AMX (Advanced Matrix Extensions). GraniteRapids will also add FP16 to expand accuracy support for AI-based developers. The platform will provide 12-channel MCRDIMM support to cope with large memory-constrained LLM models.
In terms of performance, according to internal estimates compared to the existing fourth-generation Sapphire Rapids CPU, the Granite Rapids Xeon chip will improve artificial intelligence inference (DeepMD+LAAMPS) by 2.9 times, memory bandwidth by 2.8 times, and artificial intelligence workload by 3 times. These are rough estimates only and final performance may vary.
Finally, Intel also showed some HPC performance benchmarks of its Xeon Max 9480 CPU, which comes with 56 cores and 64GB HBM memory, which is comparable to the AMD PYC9654 with 96 cores. XeonMax CPU performance can be improved by up to 30%. The EPYC9645 is priced at $11,805, while the XeonMax9480 is priced at $12,980.
Both CPUs are rated at 360/350W, but the SapphireRapids draw more power, so the total cost of ownership is likely to be lower than the AMD solution. In addition, AMD also provides Genoa-X series of 3DV-Cache enhanced products that can provide competitive performance in bandwidth-limited workload scenarios.
Intel will launch its fifth-generation Xeon CPU series on December 14, while Granite Ridge CPUs will be available in the second half of 2024, shortly after the launch of Sierra Forest.