
Pexels/Angel Bena
Microsoft and Intel announced an expanded Azure collaboration on Wednesday, Nov. 18, at 11:47 a.m. ET, and its two headline speedups, "up to 2.3x" for Compute Express Link (CXL) memory and "up to 2.5x" for SQL Server backups, come from footnotes that compare unlike systems. Azure customers sizing memory-heavy or SQL Server workloads should treat both multipliers as vendor marketing figures until matched-configuration tests exist.
The release also names Microsoft as the first company to announce use of Intel Foundry's EMIB advanced packaging for Azure first-party silicon, and confirms multi-year Xeon supply including the next-generation Diamond Rapids chips. It does not name the Azure chips or give dates, volumes or prices, a gap TechTimes covered earlier on Wednesday.
The CXL footnote compares two different Azure virtual machines. The test VM, Standard_M76ds_2_v4, has 76 vCPUs and 1,764 GB of memory using Intel Flat Memory Mode. The baseline, Standard_M64lds_v4, has 64 vCPUs and 1,007 GB of DRAM only, across two NUMA nodes. The footnote's own breakdown for the test VM, 1 TB of DRAM plus 1 TB of CXL memory, does not add up to 1,764 GB, so the real DRAM-to-CXL split is unclear.
That is 1.19 times the vCPUs and 1.75 times the memory before CXL enters the picture (author calculation from the footnote figures). The workload ran cube-hypervisor microVM sandboxes of 2 vCPUs and 24 GB each, with a persistent Python interpreter and a 12 GB dataset, and testing was completed on Nov. 10. At 2 vCPUs each, 64 vCPUs hold 32 sandboxes and 76 hold 38; the 24 GB allocations would occupy 768 GB of the baseline's 1,007 GB and 912 GB of the test VM's 1,764 GB (author calculation).
The footnote as published does not say what the 2.3x measures, whether sandbox density, throughput or latency. Without that definition, a reader cannot separate the effect of extra vCPUs and extra memory, or of the baseline's two-node layout, from the effect of CXL.
Flat Memory Mode is hardware-managed memory tiering. According to research slides on Intel's design, the CPU's memory controller handles placement in 64-byte cache lines, treats local DRAM as a direct-mapped cache of CXL memory, and pairs equal capacities of each in one address space. Software does not decide which pages live where.
The tradeoff is that performance depends on how often a workload touches data currently sitting in the slower CXL tier. The Memstrata paper presented at OSDI 2024 reports that hardware tiering stays within 5% of regular DRAM for more than 82% of workloads, but outlier workloads slow by up to 34%. One sandbox workload therefore cannot establish what Flat Memory Mode does for latency-sensitive databases or caches.
The QAT footnote, with testing completed on Nov. 2, compares a Windows Server 2025 system on a two-socket Intel Xeon Gold 8362 (32 cores per socket, 1,024 GB of memory) against an Azure VM on a Xeon 6973P-C with 384 processors and 2.06 TB of memory, running SQL Server 2025. The footnote does not mention QuickAssist, does not list a SQL Server version for the baseline system, and does not isolate QuickAssist from the newer CPU generation and doubled memory.
QuickAssist (QAT) is an Intel accelerator that offloads DEFLATE-class compression from CPU cores. SQL Server 2022 and later can use it for backup compression through a QAT_DEFLATE algorithm option alongside the default MS_XPRESS. The release describes the first QAT-enabled virtual machines in the cloud.
SQL Server 2025 adds ZSTD as a backup compression algorithm that Microsoft's documentation calls faster and more effective than MS_XPRESS. Independent community tests report large gains over the default: dbi services measured backup time falling about 41%, and Pythian's low-compression setting cut time by about 52%, a 2.1x speedup. Those tests used different hardware and databases from Intel's footnote (dbi used an 8-core, 12 GB virtual machine), so the numbers are not comparable to 2.5x.
Neither the release nor the independent tests cited here compare QuickAssist head to head against ZSTD on the same system. For a buyer, the relevant comparison is QuickAssist against the best software algorithm, not against the default; that comparison is absent from the release.
Microsoft's documentation says integrated acceleration and offloading is supported only on Windows and is not available on Linux or containers. QAT backups work when the Intel QAT drivers are installed and configured, and the documentation describes a software fallback through Intel's ISA-L library when the hardware is unavailable. Linux SQL Server deployments cannot use the feature as documented, so the 2.5x scenario does not apply to them. Intel has also published a QAT Windows driver security advisory, dated May 2026, which adds a patching responsibility to the setup.
The release says Intel Trust Domain Extensions (TDX) provides memory and CPU register state encryption as confidential VMs move between hosts. It offers no performance numbers.
A Google security assessment of Intel TDX with live migration support, conducted with Intel and posted to arXiv on Feb. 11, 2026, found five vulnerabilities. The most severe, CVE-2025-30513, rated CVSS 3.1 7.9, is a time-of-check/time-of-use flaw in which a compromised destination hypervisor can interrupt a migration import and leave a migratable trust domain debuggable, exposing its migration key and decrypted state. The other four, CVE-2025-32007, CVE-2025-27572, CVE-2025-32467 and CVE-2025-27940, are memory-read flaws rated 4.1 to 4.4 on the same scale.
Per the assessment, Intel delivered remediated TDX Module versions to customers between September and December 2025, starting at 1.5.24/1.5.25 and 2.0.14 depending on platform. The release does not say which TDX Module version Azure runs or whether its confidential VMs use live migration. Customers can check attestation reports and migration policies for the build they depend on.
Diamond Rapids is the Xeon 7 generation, which Intel has said will launch in 2027 on Intel 18A-P. The supply commitment is therefore forward-looking, and the release gives no instance counts or delivery dates. According to eWeek's report, Intel's design uses UCIe-S links rather than EMIB for longer connections between compute blocks and fabric hubs, so EMIB is not used for all of that part's internal links.
Microsoft's Saurabh Dighe, corporate vice president for Azure System & Architecture, said customers need infrastructure built around the workloads that matter to them, and Intel's Naga Chandrasekaran, executive vice president and chief technology and operations officer, said EMIB packaging will help scale future Azure infrastructure.
Ask Microsoft or Intel for matched-configuration results: the same vCPU count and memory, with and without CXL for the memory claim, and QuickAssist against SQL Server 2025 ZSTD on one system for the backup claim. Test ZSTD first, since it needs no QAT driver or accelerator setup, and verify TDX attestation and migration settings before relying on live migration for confidential workloads. Until those results exist, the figures describe the footnote systems, not Azure workloads in general.
It is a hardware-managed cacheline-granular memory tiering feature on Intel Xeon processors that manages local DRAM and CXL-attached memory in 64-byte cache lines, pairing equal capacities of each in one address space.
It can offload backup compression in SQL Server 2022 and later on Windows, but Intel's 2.5x figure compares different CPU generations and memory sizes, and SQL Server 2025 offers ZSTD as an alternative.
A Google assessment, done with Intel, found a high-severity migration flaw, CVE-2025-30513, that Intel says it remediated in TDX Module updates delivered in 2025, so the fixed version and attestation state matter.
EMIB (Embedded Multi-die Interconnect Bridge) is Intel Foundry's 2.5D packaging technology that links chiplets with a silicon bridge embedded in package substrate, as described in Intel's EMIB technology brief, rather than a large silicon interposer.
