Timothy Prickett Morgan, HPCWire, Intel ‘Haswell’ Xeon E5s Aimed Squarely at HPC, here.
The Xeon E5-2600 v2 processors topped out at 12 cores per die, but the new Haswell chips, which are implemented in the same 22 nanometer processes as their predecessors, can have as many as 18 cores per die. So that is up to a 50 percent increase in the number of cores, which drives per-system performance and which allows HPC customers to cram a lot more performance into the same space. The two top-bin parts with high core counts, the E5-2699 v3 with 18 cores and the E5-2698 v3 with 16 cores, run at 2.3 GHz and have 45 MB and 40 MB of L3 cache across those cores, respectively. They also run a little hot – in part due to the integration of the voltage regulator on the die on all Haswell Xeon E5s. And what also makes these two top-end Haswell Xeon E5 processors unique in Intel chip history is that they do not come with an official list price. These are not customized chips and will be made available to all customers, but they are, according to an Intel spokesperson, “unique offerings that fall outside of our traditional, publically available 2S product offering” and were created for HPC, virtualization, and cloud customers looking for maximum performance. The 18-core chip is obviously the most interesting one for a lot of HPC workloads that can use threads and are not as sensitive to clock speeds, and it is also the one that Intel is using to show off relative performance compared to prior Xeon E5s.
The L1 and L2 cache memory bandwidth on the Haswell cores has been doubled up, and one of the reasons why is that the second generation of Advanced Vector Extensions (AVX) integer and floating point math units have a lot more performance than the AVX1 units etched into the prior Sandy Bridge and Ivy Bridge cores. The AVX1 unit had eight 256-bit registers for floating point (four for AVX add and four for AVX multiply) and could double peak floating point operations per second (flops) in these two chips compared to their many Xeon predecessors that have had 128-bit SSE math units to 8 flops per clock. The Haswell core has 256-bit registers with its AVX2 unit, based on two Fused Multiply Add units, which doubles the peak performance to 16 flops per clock at double precision and 32 flops per clock at single precision.
Kmiecik explained to HPCwire that the new FMA instructions in the AVX2 unit would potentially increase performance for structural analysis, computational fluid dynamics, and electromagnetic field and cosmology simulations. The AVX2 feature also supports full 256-bit wide integer calculations, rather than the 128-bit width for the prior AVX1, which will be useful to accelerate image and signal processing, genomics, and cryptographic workloads.
The trick is adding up the effects of the increased core counts, better single-threaded performance, memory bandwidth, and AVX2 math units as a whole running real workloads. And here is what the initial test results look like on a variety of HPC applications:
The tests above compare a two-socket server with the 18-core Haswell Xeon E5-2699 v3, which runs at 2.3 GHz, against a machine using the 12-core Xeon E5-2697 v2 processor, which runs at 2.7 GHz.