Introduction
On June 18, 2020, Intel announced the release of their 3rd Generation Intel Xeon Scalable Processors, otherwise known as “Cooper Lake”. This is another new family of 14nm Xeon server processors that is meant for four-socket and eight-socket servers. These new processors use a new Cooper Lake/Cedar Island platform with a new Socket P+. This means that they will require new model servers.
Intel Xeon Scalable Roadmap
Back in February, Intel cancelled the 1-2 socket version of Cooper Lake. They also released the Cascade Lake Refresh processors. Cascade Lake Refresh was actually a good release for SQL Server, with clock speed increases and large price reductions.
What SKUs Are Available?
There are eleven public SKUs in this release. As usual, Intel also makes custom SKUs for their largest customers that have slightly different specifications, that are not listed in the ARK database.
These SKUs range from 16C/32T models up to 28C/56T models. The RCP prices range from $1,273.00 to $13,012.00. As you can see from the table below, most of these SKUs have lower base clock speeds than the equivalent Cascade Lake Refresh SKUs (at the same core counts). OTOH, many of them do have higher Turbo speeds than Cascade Lake Refresh.
Normally, you won’t see full Turbo clock speed from very many cores very often on a database server, unless the server is pretty idle. Cooper Lake-SP processors do have higher TDP ratings than Cascade Lake Refresh, so that could mean more time spent at Turbo clock speeds.
The Intel slide below (I apologize for the quality) has a few more details.
Other Coverage
There have already been several other posts and reviews of this release. Here are a few of them:
- Taking A Deep Dive Into “Cooper Lake” Xeon SP Processors
- 3rd Generation Intel Xeon Scalable Cooper Lake Launched for 4P and 8P
- Intel Launches Cooper Lake: 3rd Generation Xeon Scalable for 4P/8P Servers
- Intel Powers On Sapphire Rapids Processors, Scheduled for 2021
Final Words
Unfortunately, this seems to be another example of Intel hurriedly pushing a release out the door as a stopgap measure. This is mainly aimed at some extremely large customers with very specialized needs for AI and ML. The general consensus is that this is an underwhelming release outside of a few niche scenarios.
Due to the very low base clock speeds on most of the SKUs, these are not going to be a good choice for SQL Server usage. The flagship Xeon Platinum 8380H does have a 2.9 GHz base clock, which is higher than the previous Xeon Platinum 8280 (at 2.7 GHz). This makes it somewhat interesting if you want slightly better ST CPU performance from an eight-socket server. But most organizations are using two-socket Intel systems for their database servers. Most SQL Server workloads will run much better on Cascade Lake Refresh processors.
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!
I have done extensive tests recently, but my assessment made several years ago was that (true random) transaction processing work loads was most core and (hyper-)thread count dependent, not frequency. I do recall DW, especially column store, being frequency dependent. It is a good that Intel is separating 4S+ from 1S-2S, with the 4S needing more interconnect (UPI) and less PCI-E per socket. Of course, on account of the memory latency sensitivity of transactions, 1S has a meaningful per core advantage over even 2S
Hi Joe,
I guess it depends on what you mean by “dependent”. Having more total cores and threads in a CPU gives you more total CPU capacity (unless the base clock speed is so low that a lower core count, but higher clock speed processor from the same family is actually better for total capacity).
When you are comparing different SKUs from the same family (with the same core count) the SKU with the highest base clock speed is going to have better single-threaded performance. Assuming you don’t have storage or memory bottlenecks, then single-threaded CPU performance is going to important for query performance.
For example, if you compare a 24C/48T Xeon Gold 6248R with a base clock of 3.0 GHz to a 24C/48T Xeon Gold 6240R with a base clock of 2.4 GHz (with all other specifications except TDP being equal), the higher base speed CPU (at the same core count) is going to have better ST performance.
I completely agree that one-socket servers are the future for many workloads.