Evaluating the visual quality and colour differences of fullscreen H.265 Video compression with YUV 4:2:0 and YUV 4:4:4 in Citrix HDX
Table of Contents
Storage performance remains one of the most decisive factors in delivering a high-quality Virtual Desktop Infrastructure (VDI), Deskop as a Service (DaaS) and Cloud PC and experiences. While key elements like compute and graphics often receive the most attention, storage latency, throughput, and consistency significantly define how responsive a virtual desktop feels to end users. In public cloud environments such as Microsoft Azure, selecting the right storage tier is therefore not just a sizing exercise but a critical architectural decision.
This research focuses on the performance differences between these configurations in Azure. The goal is to move beyond theoretical specifications and marketing claims, providing data-driven guidance to help organizations optimize both performance and cost in Azure-based Virtual Desktop deployments.
Azure Storage options
Azure offers multiple disk types, with the Standard SSD and Premium SSD options positioned as the most viable for Virtual Desktop workloads. On paper, the differences appear straightforward: higher IOPS, lower latency, and better reliability come at a premium cost. However, real-world Virtual Desktop workloads are rarely that simple. Performance characteristics can vary significantly depending on disk sizing, workload patterns, and how the platform enforces limits at both the disk and VM level.
This raises an important question for architects and EUC engineers: do Premium SSDs consistently deliver measurable benefits for Virtual Desktops?, and if so, are those benefits worth the additional cost?
The available options are Ultra Disk, Premium SSD v2, Premium SSD, Standard SSD, and Standard HDD. However, these apply to managed disks. For the default OS disk, the choices are limited to Premium SSD, Standard SSD, and Standard HDD.
Please note that Standard HDD option is scheduled for retirement on September 1, 2026. Given both its impending deprecation and its well-documented performance limitations, it is not considered a viable option for modern Virtual Desktop deployments. As such, Standard HDD is excluded from the scope of this research.
Ultra Disk and Premium SSD v2 are only available as managed disks and can be attached to a VM as a secondary disk. This can be useful for hosting I/O-intensive applications that allow data to be stored on a specific volume , such as a Microsoft SQL database. In the context of general Virtual Desktop use cases, however, this is typically not applicable since Virtual Desktop workloads rely heavily on overall operating system performance, making these options less relevant in most scenarios, aside from specific edge cases.
The following table shows the specifications, as well as the use cases recommended by Microsoft for the different disk types:
| Ultra Disk | Premium SSD v2 | Premium SSD | Standard SSD | Standard HDD |
|---|---|---|---|---|
| Disk type | SSD | SSD | SSD | SSD |
| Scenario | IO-intensive workloads such as SAP HANA, top tier databases (for example, SQL, Oracle), and other transaction-heavy workloads. | Production and performance-sensitive workloads that consistently require low latency and high IOPS and throughput | Production and performance sensitive workloads | Web servers, lightly used enterprise applications and dev/test |
| Max disk size | 65,536 GiB | 65,536 GiB | 32,767 GiB | 32,767 GiB |
| Max throughput | 10,000 MB/s | 1,200 MB/s | 900 MB/s | 750 MB/s |
| Max IOPS | 400,000 | 80,000 | 20,000 | 6,000 |
| Usable as OS Disk? | No | No | Yes | Yes (retiring Sept 8, 2028) |
Source: https://learn.microsoft.com/en-us/azure/virtual-machines/disks-types
The theoretical maximum throughput and IOPS are not only determined by the disk type, but also by the selected disk size.
| SKU | Tier / Size | Size (GiB) | Max IOPS | Max Throughput (MB/s) |
|---|---|---|---|---|
| Ultra Disk | 4 GiB | 4 | 4,000 | 1,000 |
| 8 GiB | 8 | 8,000 | 2,000 | |
| 16 GiB | 16 | 16,000 | 4,000 | |
| 32 GiB | 32 | 32,000 | 8,000 | |
| 64 GiB | 64 | 64,000 | 10,000 | |
| 128+ GiB | 128 – 65,536 | Up to 400,000 | Up to 10,000 | |
| Premium SSD v2 | 1 GiB – 64 TiB | 1 – 65,536 | 3,000 – 80,000* | 125 – 1,200* |
| Premium SSD (v1) | P1 | 4 | 120 | 25 |
| P2 | 8 | 120 | 25 | |
| P3 | 16 | 120 | 25 | |
| P4 | 32 | 120 | 25 | |
| P6 | 64 | 240 | 50 | |
| P10 | 128 | 500 | 100 | |
| P15 | 256 | 1,100 | 125 | |
| P20 | 512 | 2,300 | 150 | |
| P30 | 1,024 | 5,000 | 200 | |
| P40 | 2,048 | 7,500 | 250 | |
| P50 | 4,096 | 7,500 | 250 | |
| P60 | 8,192 | 16,000 | 500 | |
| P70 | 16,384 | 18,000 | 750 | |
| P80 | 32,767 | 20,000 | 900 | |
| Standard SSD | E1–E30 | 4 – 1,024 | Up to 500 | Up to 100 |
| E40 | 2,048 | 500 | 100 | |
| E50 | 4,096 | 2,000 | 400 | |
| E60 | 8,192 | 4,000 | 600 | |
| E70 | 16,384 | 6,000 | 750 | |
| E80 | 32,767 | 6,000 | 750 | |
| Standard HDD | S4–S50 | 32 – 4,095 | Up to 500 | Up to 60 |
| S60 | 8,192 | 1,300 | 300 | |
| S70 | 16,384 | 2,000 | 500 | |
| S80 | 32,767 | 2,000 | 500 |
When it comes to pricing the biggest challenge with storage in Azure every tier is using a different pricing model, where Premium SSD v2 and Ultra Disk cannot be directly compared to Standard SSD and Premium SSD. Therefore they are not included in the price table. Both options are priced based on a combination of capacity, IOPS, and throughput, making a direct comparison less straightforward.
| Disk Size | SKU | Standard SSD ($/mo) | Premium SSD ($/mo) |
|---|---|---|---|
| 128 GB | E10 / P10 | ~$9.60 | ~$19.71 |
| 256 GB | E15 / P15 | ~$19.20 | ~$38.42 |
| 512 GB | E20 / P20 | ~$38.40 | ~$76.83 |
| 1024 GB | E30 / P30 | ~$76.80 | ~$135.17 |
| 2048 GB | E40 / P40 | ~$153.60 | ~$259.05 |
| 4096 GB | E50 / P50 | ~$307.20 | ~$522.24 |
It is important to note that these prices do not include any transaction fees, which are applicable for the Standard SSD’s. With each disk size for the Standard SSD there is a maximum paid transactions per hour, which is only used for billing. It is important to emphasize that there is no throttling applied when those transactions are reached. More details about the transaction fees are shared in a Microsoft blog post.
Source: https://azure.microsoft.com/en-us/pricing/details/managed-disks/
Scope and setup
For this research, the goal is to validate the performance differences per SKU in the context of a virtual desktop. Since overall performance is critical in a Virtual Desktop environment, the focus is on the OS disk. This directly excludes Ultra Disk and Premium SSD v2 from the research, as they cannot be used as OS disks.
In addition, Standard HDD is being deprecated and is not recommended for Virtual Desktop workloads, so it is also excluded from the scope. This results in the research focusing on the following scenarios:
- Standard SSD
- Premium SSD
For each SSD type, the following disk sizes are included:
- 128GB
- 256GB
- 512GB
- 1024GB
- 2048GB
- 4096GB
The default machine used for this research is a Standard_D4as_v6, which is a 4 vCPU, 16 GB memory VM running Windows 11 24H2. For each scenario, a new machine is deployed using Terraform and configured with Ansible. The configuration ensures the machine is fully up to date, the required tools are installed, and the system is optimized using the default Windows 11 template of the Citrix Optimizer. Once the machine is ready, a reboot is performed before the test starts to ensure a clean state.
When benchmarking, there are generally two types of tests: synthetic tests and load simulations. Usually, within GO-EUC, there is a preference for load simulations, as these reflect user activity and make the results more relatable to real-world scenarios. In this case, however, a synthetic test is used, as it allows the maximum capabilities of the component under test to be fully utilized. This approach is more applicable for this research, where the goal is to isolate and validate disk performance. Therefore, the tool DiskSpd is used, which is a Microsoft utility for disk performance testing. The following command line is used:
diskspd.exe -c4G -d60 -W3 -b4K -r -o8 -t2 -w50 -L -Sh testfile.dat
It is important to note that this test does not represent a fully realistic user scenario. The workload generated by DiskSpd is synthetic in nature and introduces a relatively heavy burst of I/O over a short period of 60 seconds. In a real VDI, DaaS and Cloud PC environments, user activity is more dynamic and distributed over time, rather than continuously stressing the disk at this level.
To approximate typical operating system behavior, the test is configured with a 4 KB block size, random I/O, and a 50/50 mix of read and write operations. This reflects common Virtual Desktop activities such as opening applications, reading and writing small files, and general user interaction. The test runs with a limited level of concurrency, using multiple threads and a queue depth to simulate parallel disk access from different processes within the OS. A 4 GB test file is used to ensure the workload exceeds cache boundaries, while both software and hardware caching are disabled, by the -Sh parameter, to guarantee that the measured performance reflects the actual disk capabilities. Each run includes a short warm-up period, followed by a 60-second measurement window in which metrics such as IOPS, latency, and percentile values are collected.
While this setup generates a more intensive workload than typically observed in production environments, the objective is not to mimic exact user behavior, but to maintain a consistent and controlled testing baseline. By applying a uniform mixed I/O pattern, this methodology isolates storage performance from other external variables like workload variability. This approach makes it possible to make a direct and objective comparison between different disk types and sizes. As a result, inherent differences and limitations between the storage tiers become more visible than they would under typical, user-driven activity patterns.
Hypothesis and results
Based on Microsoft documentation, Premium SSD performance has a structured, stepwise scaling model where performance increases predictably with disk size, approximating linear progression across tiers. In contrast, Standard SSD does not scale proportionally with size but instead it maintains a largely fixed IOPS ceiling across lower and mid-tier disks, with performance increases occurring only at higher capacity tiers. This results in a fundamentally non-linear scaling behavior.
DiskSpd collects various metrics during disk operations, but in this case, the primary metrics are IOPS, latency, standard deviation, and the 99th percentile. The combination of these metrics provides a clear overview of the disk’s capabilities.
The results show a clear difference in scaling behavior between Premium SSD and Standard SSD. For example, Premium SSD shows a very consistent increase in IOPS as disk size grows, which follows the predicted structured and stepwise progression. However, this scaling is not strictly linear, as performance gains diminish at higher tiers, with only marginal improvements observed between the largest disk sizes.
In contrast, the Standard SSD metrics show no significant performance scaling with increased disk size. IOPS remain effectively constant across all tested configurations, which indicates a fixed performance ceiling that is independent of capacity.
The data show a fundamental difference in how performance is delivered. While Premium SSD allows performance to be tuned through disk sizing, Standard SSD does not provide a mechanism for scaling IOPS in line with capacity. For Virtual Desktop workloads, this distinction is critical, as predictable and scalable performance is often required to maintain a consistent user experience.
Overall, the latency shows a similar pattern compared to the IOPS results. For Standard SSD, latency remains relatively consistent across all disk sizes, without any clear improvement as capacity increases. This aligns with the earlier observation that performance does not scale with disk size for this disk type.
For Premium SSD, latency decreases significantly as the disk size increases, showing clear performance scaling. This results in a noticeable difference between Standard SSD and Premium SSD, where Premium consistently delivers lower and more predictable latency.
The standard deviation indicates how stable the overall latency is, where a lower value results in more consistent performance.
For Standard SSD, the standard deviation remains relatively consistent across most disk sizes, ranging between ~55 ms and ~57 ms, with a clear spike at 1024 GB reaching ~86 ms. This confirms that, while the average values appear stable, there is still significant variability in latency. When analyzing the individual runs, this is also reflected in consistently higher deviation values, indicating less predictable performance.
As expected, the data shows that Premium SSD is more stable compared to Standard SSD. The standard deviation decreases as the disk size increases, from ~72 ms at 128 GB down to ~20 ms at 2048 GB, showing a clear improvement in consistency. It is interesting to note that at 128 GB, Premium SSD has a higher standard deviation compared to Standard SSD, but this quickly improves with larger disk sizes, where Premium SSD becomes significantly more stable.
The 99th percentile latency represents the upper bound of latency, showing the higher latency peaks during the test. This is important, as when these values are high and occur frequently, they will directly impact the overall performance and user experience.
When evaluating disk performance, it is important to consider these outliers, as they reflect the moments where the system becomes noticeably slower. For Standard SSD, the 99th percentile values are consistently high across all disk sizes, exceeding 300 ms and even peaking above 400 ms at 1024 GB. This indicates frequent latency spikes, which can negatively impact the user experience.
For Premium SSD, the 99th percentile values are significantly lower and decrease as the disk size increases, dropping from ~157 ms at 128 GB to around ~36 ms at 2048 GB. This follows the same pattern as the average latency and standard deviation results, confirming that Premium SSD delivers more consistent and reliable performance compared to Standard SSD.
It is expected that Premium SSD is more expensive than Standard SSD when comparing base disk pricing alone. However, Standard SSD introduces an additional cost component in the form of transaction costs. These costs are incurred based on the number of I/O operations (reads and writes) performed on the disk and are billed per transaction. In scenarios with high I/O activity, such as Virtual Desktop workloads, these transaction costs can accumulate rapidly.
For this analysis, the maximum possible transaction cost for Standard SSD are included to represent a worst-case scenario. Under these conditions, the total cost of Standard SSD becomes much closer to Premium SSD, and in some cases nearly equal. It is important to note that this represents an upper-bound estimate, and actual transaction costs in a typical Virtual Desktop environment may be lower depending on workload characteristics.
When reflecting to the performance results, Premium SSD consistently delivers higher IOPS and lower latency than Standard SSD. Based on the results collected in this research, certain disk sizes and configurations showed up to ~8× higher IOPS and latency reductions of up to ~80–90%.
When evaluating the balance between cost and observed performance gains, the 2048 GB Premium SSD emerges as a strong candidate. At this size, the increase in performance remains substantial compared to smaller disks, while the cost increase is still proportionate. Moving to larger disk tiers results in significantly higher costs, while the relative performance improvement becomes less pronounced.
Conclusion
When deploying a Virtual Desktop solution in Azure, there are several storage and performance considerations to take into account. Based on Microsoft recommendations for general Virtual Desktop workloads, Premium SSD is advised to achieve the best performance.
Based on this research, it can be concluded that Premium SSD indeed provides the best performance within the scope of this research, with IOPS increasing from ~1800 up to ~8100, while latency decreases from ~9 ms down to ~2 ms. In addition, Premium SSD shows significantly lower variability, with the 99th percentile dropping from ~157 ms to ~36 ms. Compared to Standard SSD, which remains around ~850–980 IOPS as it does not scale based on the size like the Premium SSD. With ~16–19 ms latency and 99th percentile values exceeding 300 ms, this results in a much more consistent and responsive user experience.
The size of Premium SSD disks also affects overall performance, with a clear tipping point around 2048 GB. Beyond this point, the increase in performance is minimal, with IOPS only increasing from ~7844 to ~8124 and latency improvements becoming negligible.
From a cost perspective, the maximum transaction cost for the Standard SSD can result in a similar total price compared to the Premium SSD. This is very dependent on the total amount of transactions per hour, which is related to the workload. This is an important factor to take into account as the Premium SSD is a more consistent cost. When diving into the Premium SSD, the sweet spot appears to be the 2048 GB Premium SSD, offering the best performance-to-cost ratio. At this size, the monthly cost is $284.94, while moving to 4095 GB increases the cost significantly to $545.10 with only marginal performance gains.
However, it is important to understand the financial impact of these choices at scale. Let’s have a theoretical example of a 2,500-seat Virtual Desktop environment where each user is backed by a dedicated virtual machine with an associated OS disk. Disk pricing is calculated based on Azure list prices for Premium SSD, using the selected disk size per VM. For example, increasing the disk size from Premium SSD 2048 GB to a Premium SSD 4096 GB doubles the storage capacity per VM, but does not result in a proportional increase in performance based on the test results observed in this research.
When applied across the full environment, this change increases the total monthly storage cost from approximately $712,350 to $1,362,750. This represents an increase of over $650,000 per month, or more than $7.8 million annually. This example highlights how relatively small architectural decisions at the individual VM level can have a substantial financial impact when multiplied across thousands of users. While larger disks may offer incremental performance benefits, the cost increase at higher tiers is significantly steeper, making it critical to carefully evaluate whether those gains justify the additional expense in a Virtual Desktop scenario.
In conclusion, there are many options for storage. In VDI, DaaS and Cloud PC environments, it is essential to strike the right balance among usability, performance, and cost. It is important to note that this conclusion is based in the context of this research, which is a synthetic test, which might be different in your environment. This also raised the question in GO-EUC, does this conclusion hold up when testing with a load simulator, which is planned on our backlog.
What are you using in your environment, and have you considered changing it? Let us know in the comments below.
Photo by Aron Yigin on Unsplash
