Introduction
I just noticed three new TPC-E benchmark results for SQL Server 2022. These are the first new TPC-E results in nearly two years and the first results for SQL Server 2022.
Just in case you are wondering, TPC-E is an OLTP benchmark that only has official submissions for SQL Server (so far, since 2007). The database vendor has veto power over TPC-E submissions from the server vendors, and only Microsoft has approved any official TPC-E submissions.
This set of recent results includes two, one-socket AMD EPYC Genoa systems and one, two-socket Intel Xeon Sapphire Rapids system. Some of the basic details about the systems and results are below:
- Lenovo ThinkSystem SR655 V3
- (1) AMD EPYC 9654
- 96C/192T
- TPC-E throughput: 13,000.00
- TPC-E Score/Core: 135.41
- HPE Proliant DL345 Gen 11
- (1) AMD EPYC 9654P
- 96C/192T
- TPC-E throughput: 12,054.79
- TPC-E Score/Core: 125.57
- Lenovo ThinkSystem SR650 V3
- (2) Intel Xeon Platinum 8490H120C/240TTPC-E throughput: 12,436.66
- TPC-E Score/Core: 103.64
Having to use a two-socket, 120 core Intel Sapphire Rapids system to get less total OLTP capacity compared to a one-socket 96 core AMD Genoa system is not a good look for Intel. Having nearly 32% lower single-threaded CPU performance (in this benchmark) is also not helpful for OLTP workloads.
A two-socket server with two Intel Xeon Platinum 8490H processors will be significantly more expensive than a one-socket server with only one AMD EPYC 9654 processor.
Finally, having to buy 24 additional SQL Server 2022 Enterprise core licenses for the Sapphire Rapids system is going to cost you roughly $165K at full retail.
Here are some more details about these three systems:
Lenovo ThinkSystem SR655 V3
This one-socket Lenovo system had an AMD EPYC 9654 processor and 1.5TB of DDR5-4800 RAM. The AMD EPYC 9654 has 96C/192T and the TPC-E throughput score for the system was 13,000.00. If we divide 13,000 by 96, we get a score/core of 135.41, which is very good for a high core-count processor.
The AMD EPYC 9654 has a base clock speed of 2.4 GHz, an all-core boost speed of 3.55 GHz and a max boost clock of up to 3.7 GHz. This processor has 384MB of L3 cache. It supports 12-channels of DDR5-4800 RAM and has 128 PCIe 5.0 lanes. The default TDP is 360W and it has a configurable TDP of 320-400W. The retail cost for this CPU is $11,508.00, and it will work in one or two-socket systems.
HPE Proliant DL345 Gen 11
This one-socket HPE system had an AMD EPYC 9654P processor and 1.5TB of DDR5-4800 RAM. The AMD EPYC 9654P has 96C/192T and the TPC-E throughput score for the system was 12,054.79. If we divide 12,054.79 by 96, we get a score/core of 125.57, which is still very good for a high core-count processor.
The AMD EPYC 9654P has a base clock speed of 2.4 GHz, an all-core boost speed of 3.55 GHz and a max boost clock of up to 3.7 GHz. This processor has 384MB of L3 cache. It supports 12-channels of DDR5-4800 RAM and has 128 PCIe 5.0 lanes. The default TDP is 360W and it has a configurable TDP of 320-400W.
The retail cost for this CPU is $10,625.00, and it will only work in one-socket systems. The “P” at the end of the model number tells you that. Other than only one-socket support, the 9654P has identical specifications to the two-socket capable 9654. The TPC-E score is lower probably because Lenovo did a better job picking components and configuring their system than HPE did.
Lenovo ThinkSystem SR650 V3
This two-socket Lenovo system had two Intel Xeon Platinum 8490H processors and 1.5TB of DDR5-4800 RAM. The Intel Xeon Platinum 8490H has 60C/120T in each CPU (so 120C/240T total), and the TPC-E throughput score was 12,436.66. If we divide 12,436.66 by 120, we get a score/core of 103.64, which is not very good for a modern high core-count processor. Being in a two-socket system adds extra NUMA overhead, which does not help overall OLTP performance.
The Intel Xeon Platinum 8490H has a base clock speed of 1.9 GHz and a max turbo clock of 3.5 GHz. This processor has 112.5MB of L3 cache. It supports 8-channels of DDR5-4800 RAM and has 80 PCIe 5.0 lanes. The default TDP is 350W. The retail cost for this CPU is $17,000.00, and it will work in up to eight-socket systems.
How Much Generational Improvement Is There?
Let’s take a look at the two most relevant previous TPC-E results, both of which are on SQL Server 2019. These are both for the most recent previous generation processors from AMD and Intel.
Lenovo ThinkSystem SR655
This one-socket Lenovo system had an AMD EPYC 7763 processor and 1.0TB of DDR4-3200 RAM. The AMD EPYC 7763 has 64C/128T and the TPC-E throughput score for the system was 7890.66. If we divide 7890.66 by 64, we get a score/core of 123.29, which is very good for a high core-count processor.
The AMD EPYC 7763 has a base clock speed of 2.45 GHz and a max boost clock of up to 3.5 GHz. This processor has 256MB of L3 cache. It supports 8-channels of DDR4-3200 RAM and has 128 PCIe 4.0 lanes. The default TDP is 280W and it has a configurable TDP of 225-280W. The retail cost for this CPU is $7,890.00, and it will work in one or two-socket systems.
Lenovo ThinkSystem SR860 V2
This four-socket Lenovo system had four Intel Xeon Platinum 8380H Cooper Lake processors and 3.0TB of DDR4-3200 RAM. The Intel Xeon Platinum 8380H has 28C/56T in each CPU (so 112C/224T total), and the TPC-E throughput score was 12,162.59. If we divide 12,162.59 by 112, we get a score/core of 108.59, which was not very good for a modern high core-count processor. Being in a four-socket system adds even more NUMA overhead, which does not help overall OLTP performance.
The Intel Xeon Platinum 8490H has a base clock speed of 2.9 GHz and a max turbo clock of 4.3 GHz. This processor has 38.5MB of L3 cache. It supports 6-channels of DDR4-3200 RAM and has 48 PCIe 3.0 lanes. The default TDP is 250W. The retail cost for this CPU is $10,009.00, and it will work in up to eight-socket systems.
AMD Genoa managed to show a decent improvement over their previous generation flagship AMD Milan server processors in the TPC-E benchmark, while Intel Sapphire Rapids actually regressed in this benchmark compared to the older Intel Cooper Lake. That is not good news for Intel. Here are the results in tabular form:
Conclusions
Intel’s probable response to this is that general compute performance is an out-moded way of comparing server processor performance that ignores specialized workloads that Sapphire Rapids can handle much better than AMD. According to Intel, the specialized accelerators in Sapphire Rapids give them a significant advantage for the specific workloads that can use them effectively.
This shows the built-in accelerators in Intel Sapphire Rapids:
For SQL Server 2022 (and older versions of SQL Server) much of this Intel argument about accelerators is moot. SQL Server 2022 can use Intel QAT Backup compression in hardware assist mode or in software mode. In order to use hardware mode, you have to be using SQL Server 2022 Enterprise Edition and you must have Intel QAT-capable hardware. Unfortunately, most standard Intel Sapphire Rapids SKUs do not have any QAT accelerators. Intel has really hurt the prospective adoption of QAT with this product segmentation decision. The Intel graphic below shows default accelerator support by SKU.
The large cloud providers will be using custom Intel Sapphire Rapids SKUs, which may or may not have QAT support. Even if the host processor does have QAT support, the cloud providers may or may not expose QAT to guest VMs.
Unfortunately, SQL Server 2022 does not leverage any other specialized Intel accelerators. SQL Server 2022 can take advantage of AVX-512 support for some limited Columnstore batch mode query operations if your CPU has AVX-512 support (and you enable TF 15097). Both AMD Genoa and Intel Sapphire Rapids have “good” AVX-512 support and performance that does not throttle clock speeds when AVX-512 operations occur. Just to be clear, AVX-512 is a set of special vector instructions that is completely different from a specialized accelerator.
Final Words
You might be thinking that this is interesting, but it simply doesn’t apply to you because you are running some form of SQL Server in a PaaS or IaaS cloud scenario. Because of that, things like benchmark scores and gory hardware details don’t matter (or you have no control over them).
I think that is an incorrect point of view. Everything in the cloud ultimately runs on some host hardware, somewhere. New generations of processors with better ST performance and higher core counts, along with more memory and I/O bandwidth make a huge difference for the cloud providers and for you. Offloading specialized workloads (including infrastructure tasks) to specialized hardware makes more general compute capacity available to the host (and to guest VMs).
Since SQL Server 2022 (and older versions) does not use any “specialized” accelerators besides Intel QAT acceleration, general compute performance is still extremely important for SQL Server. These recent TPC-E results show an overwhelming general compute performance advantage on OLTP workloads for AMD Genoa compared to Intel Sapphire Rapids.
For SQL Server, this means that you should purposely choose a cloud provider offering that uses AMD Genoa if you can.
AMD Genoa cloud instances will be available relatively soon from Microsoft, AWS and Oracle Cloud Infrastructure (OCI).
If you have any questions about this post, please ask me here in the comments or on Twitter. I am pretty active on Twitter as GlennAlanBerry. Thanks for reading!