Microsoft adds more AMD-powered Azure VMs, while Intel is caught in a benchmarking scandal

Microsoft is expanding the variety of AMD-powered virtual machine types on Azure. Intel is feeling the heat after being caught misrepresenting benchmarks.

Intel and AMD announce new CPUs at Computex 2019 Intel and AMD presented new processors that will make their way into upcoming computers. Karen Roby and James Sanders discuss the next generation of CPUs.

Microsoft announced the general availability of new virtual machine instances for Azure at Ignite 2019 in Orlando on Monday, bringing additional configurations for AMD-powered instances and Generation 2 instances as part of an overall update to Azure's underlying infrastructure.

The AMD EPYC 7452-powered Azure Da_v4 and Das_v4 instances are targeted toward "enterprise-grade applications, relational databases, and application servers" according to Microsoft, and support up to 96 vCPUs, 384 GB RAM, and 2,400 GB SSD-delivered temporary storage, with support for Azure Premium SSDs and Ultra Disks.

Likewise, the Azure Ea_v4 and Eas_v4 virtual machines use AMD's EPYC 7452, with support for up to 96 vCPUs per instance, 672 GB RAM, and 2,400 GB SSD-delivered temporary storage, supporting Azure Premium SSDs and Ultra Disks.

SEE: Microsoft Azure: An insider's guide (free PDF) (TechRepublic)

While the Da/Das_v4 series were announced in preview in August, those and the Ea/Eas_v4 hit general availability on Monday. Microsoft's first (Zen 2) EPYC-powered instances were announced in February, though Microsoft's continued enthusiasm for EPYC CPUs underlies AMD's newfound competitiveness in enterprise compute, which--alongside AMD offerings from AWS, OVHCloud, and Oracle, and future offerings from IBM and Google Cloud Platform--underlie AMD's potential for loosening Intel's tightfisted grip on enterprise compute.

That said, Microsoft is also iterating Intel-powered virtual machine offerings, with Generation 2 virtual machines bringing support for higher memory capacities and Intel's Software Guard Extensions (SGX), additionally enabling UEFI boot architecture in Azure, which permits OS disk sizes in excess of 2 TBs, and RAM capacities up to 12 TB.

Intel publishes misleading benchmark

Given the heat that Intel is facing with a resurgent AMD, to say nothing of their own difficulties across the product line as the transition to 10nm has been a protracted, years-long problem form the firm, the "Performance at Intel" blog has been caught red-handed by Patrick Kennedy at STH for publishing a GROMACS benchmark claiming that a two-socket Xeon 9282 system is 20% faster than a two-socket EPYC 7742 system. 

The core contention is Intel's use of GROMACS 2019.3, which was published in June 2019, rather than GROMACS 2019.4, published on October 2, over a month before Intel's benchmarks were published. The minor version bump to the popular molecular dynamics package added support for detecting the Zen 2 microarchitecture, marking the first release that has any enablement of features in AMD's newest processors.

With the changed version, GROMACS enables 256-bit wide AVX2 SIMD instructions by default, as well as adding performance tuning for non-bonded kernel parameters, which the release notes indicate "[have] a significant impact on performance."

Kennedy's benchmarking eliminates any Intel advantage, though notes that further optimizations are still needed for AMD, and highlights other configuration oddities, concluding that Intel's post "is not a reputable attempt to present factual information."

Update (Nov. 7, 2019): An Intel spokesperson provided the following statement:

Intel is committed to always provide fair, transparent, and accurate performance results and would not intentionally mislead. We received feedback on our original blog and appreciate the community's passion about performance and the accuracy of benchmarks. Taking the community's feedback, we have updated this blog with data for the most recent GROMACS 2019.4 version and found no material difference to earlier data posted on 2019.3 version.

STH's Patrick Kennedy posted an update addressing Intel's updated test run, highlighting some disparities between the two compared products, including questioning the equability of a comparison between a 400W TDP niche part from Intel, compared to the (default) 225W TDP of the EPYC 7742, which is configurable to a 240W TDP, concluding that "We need to remember this is a marketing exercise. As such, we also need to expect that Intel is going to try presenting its best case."

Also see

Black abstract diamond pattern background

Image: James Sanders/TechRepublic