AMD has released three new frequency-optimized EPYC 7002 series SKUs that are ideal candidates for SQL Server usage. This is true for both OLTP and DW workloads. Historically, AMD EPYC 7002 Series have done extremely well in DW workloads such as TPC-H. At the same time, their single-threaded CPU performance has not quite matched the fastest Intel Cascade Lake-SP processors. This has hurt them on OLTP workloads such as TPC-E.
This situation has now changed, with the release of these three new AMD EPYC processor SKUs. These three models have much higher clock speeds and very large L3 caches, which result in them having better single-threaded performance than the latest Intel Cascade Lake-SP Refresh processors. This has never happened before, and I believe it is a very significant development for SQL Server.
Three New AMD EPYC 7002 Series SKUs
The top-end new SKU is the AMD EPYC 7F72. This model has 24C/48T, with a base clock speed of 3.2 GHz, and a max boost clock speed of 3.7 GHz. It has 192MB of L3 cache, and a TDP rating of 240 watts. This processor will cost $2,450.00.
Next down in the stack is the AMD EPYC 7F52. This model has 16C/32T, with a base clock speed of 3.5 GHz, and a max boost clock speed of 3.9 GHz. It has 256MB of L3 cache, and a TDP rating of 240 watts. This processor will cost $3,100.00.
Finally, we have the AMD EPYC 7F32. This model has 8C/16T, with a base clock speed of 3.7 GHz, and a max boost clock speed of 3.9 GHz. It has 128MB of L3 cache, and a TDP rating of 180 watts. This processor will cost $2,100.00.
All of these processor SKUs will work in both one-socket and two-socket servers. They have significantly higher base clock speeds than the previous highest clock speed SKU at each core count. This will also cause them to use more power and require better cooling (although water cooling won’t be required).
Why Is This Important for SQL Server?
This is a very big deal for SQL Server because of core-based licensing.
SQL Server license costs are based on the number of physical cores in your machine (on bare metal servers). The actual performance of each core has no effect on the license price. If you choose frequency-optimized processors, you can often get away with having a lower core count processor. This can save a huge amount in licensing costs, and it also give you better single-threaded performance.
The 24C/48T EPYC 7F72 has a 192MB L3 cache, compared to only 128MB in the older 24C/28T EPYC 7402. Even better, the 16C/32T EPYC 7F52 has a huge 256MB cache, compared to only 128MB in the older 16C/32T EPYC 7302. Having a larger L3 cache, and more L3 cache/core is very important for relational database performance.
In any processor, finding the instructions and data that you need to complete a request as quickly as possible is very important for performance. Finding what you need in the L1 cache is better than having to go to the L2 cache or the L3 cache. Going from the L3 cache to DRAM means a large jump in latency! Having a very large L3 cache, especially when it is shared among a lower number of physical cores, is very beneficial.
Dr. Ian Cutress of Anandtech measured the relative cache latency in an AMD EPYC 7F52 processor. His results are shown below:
- 1.0 nanoseconds for L1 (4 clks) up to 32 KB
- 3.3 nanoseconds for L2 (13 clks) up to 256 KB
- 4.8-5.6 nanoseconds (19-21 clks) at 256-512 KB
- 12-14 nanoseconds (48-51 clks) from 1 MB to 8 MB inside the first half the CCX L3
- Up to 37 nanoseconds (60-143 clks) at 8-16 MB for the rest of the L3
- ~150 nanoseconds (580-600+ clks) from 16 MB+ moving into DRAM
For this processor, finding what you need in the L3 cache is between 4X and 12X faster than having to go out to main DRAM memory. With 256MB of total L3 cache, it is more likely that more of your workload will be found in the L3 cache.
How Does This Compare to Intel Cascade Lake-SP?
Based on the comparative specifications and initial benchmark results, AMD finally has a server processor that has faster single-threaded CPU performance than Intel (at the same core counts). This has never happened before. This makes AMD a superior choice for many workloads, both for OLTP and DW workloads.
AMD released a slide, shown in Figure 2, that compared OLTP database performance between AMD and Intel processors at the same core count. As always, you should take vendor-supplied benchmarks with some skepticism.
My main problem with this slide is that AMD compared the 16C AMD EPYC 7F52 to the 16C Intel Xeon 6242 rather than the faster 16C Intel Xeon 6246R. They made a similar mistake by comparing the 8C AMD EPYC 7F32 to the 16C Intel Xeon 6244 rather than the faster 16C Intel Xeon 6250. AMD still would win against those newer, faster SKUs from Intel, but the margin would be lower.
Best AMD EPYC Processors for SQL Server
Here are the “best” AMD EPYC processor choices for a one-socket system for SQL Server. These scores and score/core numbers are from my estimated TPC-E score calculations for each processor. The score/core represents single-threaded CPU performance. The license cost is for SQL Server 2019 Enterprise Edition.
Best Intel Xeon Processors for SQL Server
Here are the “best” Intel Xeon processor choices for a two-socket system for SQL Server. These scores and score/core numbers are from my estimated TPC-E score calculations for each processor. The score/core represents single-threaded CPU performance. The license cost is for SQL Server 2019 Enterprise Edition.
A number of reviews of these processors have already gone live, if you want more detail.
- AMD EPYC 7F52 Benchmarks Review and Market Perspective
- AMD’s New EPYC 7F52 Reviewed: The F is for ᴴᴵᴳᴴ Frequency
- AMD EPYC 7F52 Linux Performance – AMD 7FX2 CPUs Further Increasing The Fight Against Intel Xeon
- AMD Claims World’s Fastest Per-Core Performance with New EPYC Rome 7Fx2 CPUs
These new frequency-optimized SKUs from AMD are a very welcome new development. This will increase the competitive pressure on Intel. Intel’s most likely response is further price reductions. Since the cancellation of Intel Cooper Lake, the existing Cascade Lake-SP Refresh processors will have to hold the line against AMD until probably early 2021.
As database professionals, we have two CPU vendors in heated competition. Intel still has a huge advantage in market share, and there is also a large amount of institutional inertia in Intel’s favor. Intel had much better processors for SQL Server usage from about 2008 until 2019, so it is harder for AMD to change people’s perceptions.
AMD has a more modern, modular architecture, higher memory density and bandwidth, and PCIe 4.0 support. AMD also seems to have fewer security vulnerabilities. Now, with these frequency-optimized SKUs, AMD has closed one of the final remaining gaps with Intel.