Intel 5th Gen Xeon Performance Benchmarks: Impressive Efficiency Gains With "Optimized Power Mode"

[phoronix]

Phoronix: Intel 5th Gen Xeon Performance Benchmarks: Impressive Efficiency Gains With "Optimized Power Mode"

With the new Intel 5th Gen Xeon "Emerald Rapids" processors there is a new feature called the Optimized Power Mode (OPM). This Optimized Power Mode can be enabled via the system BIOS for Emerald Rapids for helping to reduce the CPU power consumption when not running at full utilization. With Intel claiming that up to 110 Watts of server power savings can be conserved when running at 30~40% utilization, I was curious and set out to run my own Optimized Power Mode benchmarks with the Xeon Platinum 8592+ processors. Here is a deep dive on Intel Optimized Power Mode benchmarks.

https://www.phoronix.com/review/xeon-optimized-power-mode

[Mark Rose]

They couldn't make it the default or the new chips would look worse from a performance perspective.

[bug77]

So first you pounded the server as hard as you could. And then turned on the "Optimized Power Mode" (that supposedly reduced power draw when the server is not at 100% usage) and pounded it as hard you could again. Did I get that right?

[acobar]

So first you pounded the server as hard as you could. And then turned on the "Optimized Power Mode" (that supposedly reduced power draw when the server is not at 100% usage) and pounded it as hard you could again. Did I get that right?

Did you miss the part where the graphics were presenting performance/watt? From a general POV, you can't be sure what will be the demand applied to your server, but you can set it to give the best performance/watt, this is what Michael did.

[blackshard]

This demonstrated only one thing: Intel is pushing their processor way too much outside their power/performance sweet spot to not trail behind in benchmarks.

This is very ironic: they started pushing and pushing their processor to insane power usages from 9th family onward, inventing "Power Levels" and several other gimmicks to keep the performance high enough to compete, and now they come out with an "optimized mode" as a feature to reduce power to "reasonable" levels because their processors draw too much power.

[bug77]

Did you miss the part where the graphics were presenting performance/watt? From a general POV, you can't be sure what will be the demand applied to your server, but you can set it to give the best performance/watt, this is what Michael did.

That was my question: is that perf/W representative of a server at 100% load or is it representative of a server in the ballpark where "Optimized Power Mode" is supposed to help?

[sophisticles]

From the article:

In particular, around the 30~40% range that many data centers and server environments operate at are supposed to realize significant power savings

This right here is why i keep saying that AMD's TRs and EPYs are a scam and a waste of money.

Most of the time, most computers, are operating as a fraction of their maximum performance.

If this is the case, then a 96C/192T TR can be replaced with a 32C/64T TR and end users will not notice the difference.

The person paying for the system will notice when he doesn't have to spend 5+ grand a CPU.

This is also why I keep saying that we need CPUs with dedicated hardware accelerators and not more cores and why i say that Intel's upcoming E-core only offerings will probably dominate most server deployments.

It's not about raw horsepower, it's about what is their operating range and how much juice do they need.

With using FFmpeg for some live video handling, there was around a 10% performance hit running in the Optimized Power Mode. But there was a 95 Watt power saving! The peak CPU power consumption was nearly 100 Watts lower. This led to a 31% boost in power efficiency for FFmpeg workloads on the Xeon Platinum Emerald Rapids server.

This right here speaking volumes, 90% of the performance at about half the power consumption.

This is the formula that allow Microsoft and Intel to beat Big Iron back in the day and it's the formula that will allow Intel to beat back AMD.

[bug77]

sophisticles I believe the idea is not to run things maxxed out, but keep some spare room for spikes. But yes, like most parts handling a wide range of workloads (electronic, electric, mechanic or whatever), there's sweet spot and there are area where efficiency won't shine.

[Keats]

From the article:



This right here is why i keep saying that AMD's TRs and EPYs are a scam and a waste of money.

Most of the time, most computers, are operating as a fraction of their maximum performance.

If this is the case, then a 96C/192T TR can be replaced with a 32C/64T TR and end users will not notice the difference.

The person paying for the system will notice when he doesn't have to spend 5+ grand a CPU.

This is also why I keep saying that we need CPUs with dedicated hardware accelerators and not more cores and why i say that Intel's upcoming E-core only offerings will probably dominate most server deployments.

It's not about raw horsepower, it's about what is their operating range and how much juice do they need.



This right here speaking volumes, 90% of the performance at about half the power consumption.

This is the formula that allow Microsoft and Intel to beat Big Iron back in the day and it's the formula that will allow Intel to beat back AMD.

Too bad AMD already beat Intel to the punch with Siena and Bergamo

[NobodyXu]

sophisticles AFAIK cloud computing is usually over-provisioned, there are way more, there are endless tasks and it scales well since there are enough customers to take full advantages of 192 cores provide AMD.

They usually don't care single-core perf as much as multi-core, so Intel is lagging behind a lot.