The Tech Report recently wrapped up the first part of their review of Intel's new Core i9-7900X, focusing on its effectiveness in production machine. Their benchmarks cover a variety of scientific tasks such as PhotoWorxx, FPU Julia and Mandel as well as creativity benchmarks like picCOLOR, DAWBench DSP 2017 and STARS Euler3D. During their testing they saw the same peaks in power consumption as Ryan did in his review, 253W under a full Blender load. Their follow up review will focus on the new chips gaming prowess, for now you should take a look at how your i9-7900X will perform for you when you are not playing around.
"Intel's Core i9-7900X and its Skylake-X brethren bring AVX-512 support, a new cache hierarchy, and a new on-die interconnect to high-end desktops. We examine how this boatload of high-performance computing power advances the state of the art in productivity applications."
Here are some more Processor articles from around the web:
- Intel Core i9 7900X Linux Benchmarks @ Phoronix
- Intel Core i7 7740X Benchmarks On Linux @ Phoronix
- Ryzen 5 1400 @ Hardware Secrets
That AVX 512 support is why
That AVX 512 support is why Intel’s Xeon server SKUs are outperforming Epyc’s Zen CPU cores is some Benchmarks, but AMD can just offer up its Radeon Pro WX/Radeon instinct GPU accelorators for any HPC workoads that need FP performance or for other AI/Inference ability.
AMD’s Epyc SKUs use less power than Xeon, likely due in part to Intel’s extra AVX 512 hardware on a per core basis. But with GPUs and some FPGAs and ASICs(Tensor Processors) taking over more of the heavy FP crunching and other AI/Inferencing workloads, Epyc’s Zen core AVX deficiency is not as much of a disadvantage as it would be in the past.
With Epyc and the Infinity Fabric any Epyc systems can communicate with Vega accelorators via the Infinity Fabric. So there is a possible direct and coherent connection between Epyc’s Zen cores and any Vega GPU that will support the Infinity Fabric, and that includes Vega GPU to Vega GPU coherency, in addition to Zen core/s to Vega GPU Infinity Fabric coherent protocol communication.
28C Skylake Xeons are
28C Skylake Xeons are outperforming 32C AMD Epyc on benchmarks that don’t use AVX512, including CB or SPECint.
Yes, AMD can just offer up its Radeon Pro WX/Radeon instinct GPU accelorators for any HPC workoads that need FP performance or for other AI/Inference ability. But that is a slower and less efficient solution, not to mention programming issues. AMD GPU accelerators aren’t used in HPC, period:
“A total of 91 systems on the list are using accelerator/co-processor technology, up from 86 on November 2016. 71 of these use NVIDIA chips, 14 systems with Intel Xeon Phi technology (as Co-Processors), one uses ATI Radeon, and two are using PEZY technology.”
AMD’s Epyc SKUs use more power than Xeon. Another reason why AMD CPUs are being masssively rejected.
More power on the platform
More power on the platform maybe but not the Zen/Epyc CPU cores themselves if you include the Epyc platform’s extra memory channels(8 DDR4 memory channels which can support more of the smaller less expensive DDR4 modules for some workloads/Extra PCIe lanes). And all those extra PCIe lanes that AMD provides on its 1P(128 PCIe lanes) and 2P(128 PCIe lanes per Epyc/socket available with each CPU giving up 64 PCIe lanes over to infinity fabric inter-socket communication, leaving 128 total PCIe lanes on the 2P systems as well). So on Epyc systems, AMD does no platform product segmentation like Intel does in reducing/gimping with the PCIe lanes and charging $$$$ to get the PCIe lanes back.
AMD’s Epyc HPC systems will be using the Infinity Fabric to directly communicate with the Vega GPU accelerator and Vega AI/Inferencing SKUs, so that’s direct attached Zen/Epyc to Vega GPU AI/Accelerator fully cache coherent communication via the Infinity Fabric protocol(Very similar to Nvidia’s NVLink) and Infinity Fabric hardware IP. Vega GPUs will be able to communicate with other Vega GPUs also via the Infinity fabric, in addition to any Zen/Epyc processors.
Both AMD’s and Nvidia’s GPUs offer much more FP Tflops/Gflops per watt than Intel’s Xeon/Knights SKUs and AMD is just getting back into the workstation/server/HPC market with Epyc(And Radeon Pro WX/Radeon instinct GPU accelerators will ride along wired up and communicating with full coherency via the Infinity Fabric). AMD is also along with IBM/others a founding member of OpenCAPI so there will be the possibility of Power9 systems using AMD’s Radeon Pro WX and AI products.
AMD probably made Zen with the smaller AVX units in mind with the knowledge that AMD has its GPU IP to bring online to work any heavy HPC FP workloads, and and AMD’s next product stack includes a Zen2 Epyc/Rome platform that may see 6 Zen2 cores per CCX, at 12 cores per Die and 48 Zen2 cores. AMD’s Zen/Epyc cores take up less area than Intel’s and use less power. AMD’s Epyc platform Motherboards are a bit less crowded owing to the fact that Zen/Epyc is a true SOC with northbridge/southbridge included on the Zeppelin die/MXM. So Epyc CPUs are very power efficient and make for good lower power consuming options for customers that do not need any extra FP workload power on the CPU, and there are plenty of server workloads to do not need AVX.
AMD is so busy with Zen/Epyc and getting Vega to market that AMD still has not begun to discuss its in development workstation/server/HPC Interposer based APUs targeting the professional markets, so that’s something to watch for in 2018. AMD also has its Zen/Vega APUs fro the mobile market to introduce with mobile versions of the first Zen/Vega consumer/mobile APUs arriving 2H 2017, with any PC/Pro desktop Zen/Vega arriving in 2018.
So let’s revisit those Workstation/Server/HPC Epyc/Vega WX figures over the next year or two following AMD’s Epyc/Vega Professional WX/AI Full release to market and see where AMD will stand on those market metrics. For the Server/Workstation/HPC market AMD can only go up at this early point in AMD’s return to this market.
Let’s wait for some Price/Performance and some Price/Features metrics for Epyc systems and Epyc paired with Vega Pro WX/INS. Intel(On their own) and Nvidia(On their own) can offer the same level of x86 CPU/Vega GPU Professional package/Pricing deals like AMD has the ability to offer. The TCO figures for AMD will be very attractive as well as the Price/Performance that AMD/Epyc/Vega Professional will offer.