IBM Power11 Raises the Bar for Enterprise IT

[TheLostSwede]

Today, IBM revealed IBM Power11, the next generation of IBM Power servers. Redesigned with innovations across its processor, hardware architecture, and virtualization software stack, Power11 is designed to deliver the availability, resiliency, performance, and scalability enterprises demand, for seamless hybrid deployment on-premises or in IBM Cloud.

Organizations across industries have long run their most mission-critical, data-intensive workloads on IBM Power, most notably those within the banking, healthcare, retail, and government spaces. Now, enterprises face an onslaught of new technologies and solutions as they transition into the age of AI. IDC found that one billion new logical applications are expected by 2028, and the proliferation of these systems poses new complexities for companies. IBM built Power11 to deliver simplified, always-on operations with hybrid cloud flexibility for enterprises to maintain competitiveness in the AI era.




Power11 is designed to be the most resilient server in the history of the IBM Power platform, with 99.9999% of uptime. Together with zero planned downtime for system maintenance and less than one-minute guaranteed ransomware threat detection with IBM Power Cyber Vault, Power11 sets a new bar for business continuity, addressing both planned and cyber-incident-related downtime.

For the first time ever, the Power11 general availability will simultaneously include high-end, mid-range, and entry servers as well as IBM Power Virtual Server in IBM Cloud. IBM Power Virtual Server offers a fast path to the cloud for Power workloads, and is certified as a hyperscaler platform for RISE with SAP. Power11 will also be the first IBM Power server to support the IBM Spyre Accelerator, IBM's system-on-a-chip available Q4 2025 that is purpose-built for today's AI-intensive inference workloads. Available across IBM's enterprise systems portfolio, including Power11, IBM z17, and LinuxONE 5, Spyre will deliver advanced AI acceleration to help organizations scale AI across hybrid cloud environments.

With support for autonomous operations, Power11 delivers intelligent performance gains that reduce complexity and improve workload efficiency. Power11 offers up to 55% better core performance compared to Power9 and has up to 45% more capacity with higher core counts in entry and mid-range systems compared to Power10. This upgraded performance means that with Power11, enterprises are positioned to achieve enhanced flexibility and security, to transform their enterprise processes through automation.

With Power11, customers can expect:

• Zero planned downtime for system maintenance. Power11 provides enterprises with a solution that can avoid costly planned downtime and help reduce operational risk. Through advanced technologies like autonomous patching and automated workload movement, planned system maintenance events can occur without ever taking critical applications offline. This feature can free IT professionals from spending time planning, testing, and executing upgrades to their systems, to focusing on higher-value work and innovation. Furthermore, IBM Power will interoperate with IBM Concert using generative AI to help identify operational risks, provide actionable insights, and automate remediation, starting with security patch management. IBM Technology Lifecycle Services (TLS), IBM's global infrastructure support provider, complements this feature with premium, AI-powered services designed to help proactively monitor system health, reduce downtime, and optimize system reliability and performance.
• Less than one minute ransomware threat detection with IBM Power Cyber Vault. The Power Cyber Vault solution is an integrated cyber resiliency solution following NIST cybersecurity framework to help identify, protect, detect, and automatically respond to cyber threats. Cyber Vault provides protection against cyberattacks such as data corruption and encryption with proactive immutable snapshots that are automatically captured, stored, and tested on a custom-defined schedule. Power11 also uses NIST-approved built-in quantum-safe cryptography designed to help protect systems from harvest-now, decrypt-later attacks as well as firmware integrity attacks.
• Transformed business processes with AI integration. Power11 delivers AI-ready infrastructure with built-in, on-chip acceleration for inferencing and will be able to scale to support mission-critical AI workloads through the IBM Spyre Accelerator. Combined with Red Hat OpenShift AI and a broad ecosystem of open-source software and toolkits, Power11 is built to provide the flexibility and performance needed to operationalize AI across hybrid environments. To further modernize application development, IBM watsonx Code Assistant for i will help developers extend critical RPG applications for greater ease and productivity. IBM will also make watsonx.data, its hybrid, open data lakehouse, available on Power11 by the end of 2025.
• Efficient IT that saves time and money. Beyond the autonomous processes that support zero planned downtime and IBM Power Cyber Vault, Power11 delivers meaningful efficiency gains across the IT stack. Separately, in terms of energy efficiency, Power11 offers twice the performance per watt versus comparable x86 servers and up to 28 percent better server efficiency with the new Energy Efficient Mode compared to Maximum Performance Mode on Power11.

"IBM Power11 changes the game for enterprise computing," said Tom McPherson, GM, Power Systems at IBM. "With Power11, clients can accelerate into the AI era with innovations tailored to their most pressing business needs. We are taking advantage of the full IBM stack to deliver hybrid cloud, AI, and automation capabilities while building on our decades-long reputation as a trustworthy hybrid infrastructure for essential workloads."

"With just 20 minutes and the help of watsonx Code Assistant for i on Power, I was able to investigate a report, trace the field logic, understand the calculation, and document the issue," said Jasmine Kaczmarek, VP of Technology at MR Williams. "What had taken a senior developer six hours the day before, I was able to accomplish 18 times faster. That's truly remarkable when you think about it. This kind of productivity gain is exactly what Power11 is designed for. By empowering developers to modernize core business applications faster and more confidently, AI-augmented tools built into the platform can help increase long-lasting business value."

"We're committed to helping the people of Florida receive the quality health care they can afford," said William Allarey, Sr. IT Manager at GuideWell. "With the new IBM Power11 automation capabilities, we are very interested in faster and more frequent maintenance updates with no planned downtime to keep the servers secure, stable, and current, so our team can focus on delivering benefits and services for better health. In addition, the native AI capabilities and significant system performance gains of the Power11 will allow our company to leverage AI for a clearer and faster claims experience."

"Temenos' collaboration with IBM has already proven the value of running mission-critical workloads on IBM Power," said William Moroney, Chief Revenue Officer at Temenos. "With the launch of Power11, that potential grows even further by bringing an AI-ready infrastructure, zero-downtime resilience, and even greater performance to the financial services industry. We look forward to our continued partnership to help clients modernize their core banking on this built for hybrid cloud platform."

IBM Power11 will be generally available July 25, 2025. The IBM Spyre Accelerator is expected to be available in Q4 2025.

View at TechPowerUp Main Site | Source

 

[Daven]

Man I haven’t seen news about POWER in forever. No POWER10 on the top 500 supercomputers. In fact, no new POWER entries on that list at all in the last five years. I’m guessing each version of POWER is for upgrading a fixed number of instances locked into that technology from long ago. I can’t believe anyone is developing new applications based on POWER.

 

[AleksandarK]

Man I haven’t seen news about POWER in forever. No POWER10 on the top 500 supercomputers. In fact, no new POWER entries on that list at all in the last five years. I’m guessing each version of POWER is for upgrading a fixed number of instances locked into that technology from long ago. I can’t believe anyone is developing new applications based on POWER.

Same! I miss seeing Power news, but it got pushed out. It offers nothing an x86/Arm server can't offer, and if you want maximum uptime you go for a mainframe. Soo a weird place in the market. Definitely there are buyers, but not like there used to be.

IIRC physics labs used to run a lot of Power servers (back in Power8 era), but I am sure they moved on these days.

 

[ncrs]

Man I haven’t seen news about POWER in forever. No POWER10 on the top 500 supercomputers. In fact, no new POWER entries on that list at all in the last five years. I’m guessing each version of POWER is for upgrading a fixed number of instances locked into that technology from long ago. I can’t believe anyone is developing new applications based on POWER.

The main reason POWER systems were on the Top500 list is NVIDIA's close relationship with IBM that resulted in integration of NVLink into POWER8+ (1.0) and 9 (2.0). It was of course a stopgap measure until NVLink 3.0 providing more lanes per GPU allowing complex internal topologies, and 4.0 with NVSwitch chips were ready. After that AMD64 was enough as a base platform which removed the cost of supporting a separate POWER software stack, with the prospect of NVIDIA ARM CPUs taking over in the future.

Unfortunately Power10 lost a bit of good will when they didn't support "open-sorce everything" like in 8 and 9, so no new Talos workstation for example.

 

[Daven]

The main reason POWER systems were on the Top500 list is NVIDIA's close relationship with IBM that resulted in integration of NVLink into POWER8+ (1.0) and 9 (2.0). It was of course a stopgap measure until NVLink 3.0 providing more lanes per GPU allowing complex internal topologies, and 4.0 with NVSwitch chips were ready. After that AMD64 was enough as a base platform which removed the cost of supporting a separate POWER software stack, with the prospect of NVIDIA ARM CPUs taking over in the future.

Unfortunately Power10 lost a bit of good will when they didn't support "open-sorce everything" like in 8 and 9, so no new Talos workstation for example.

The first Power processors were launched in 1990 long before Nvidia was anything. Power has been in the top500 supercomputer list since the first list published in 1994. Their presence in mainframes and servers was because of the 'Big Iron' era that was dominated by DEC, IBM, Intel (Itanium) and others. Any recent relation to Nvidia was due to shifting compute paradigms that forced a shift in IBM product topography.

 

[BoggledBeagle]

Power11 offers up to 55% better core performance compared to Power9 and has up to 45% more capacity with higher core counts in entry and mid-range systems compared to Power10.

What does this even mean?????? Why do they compare the performance to Power9, and not Power 10?! And what is the "capacity" they talk about?

Why dont they compare the performance to competing server CPUs from AMD, Intel and ARM?

 

[Daven]

What does this even mean?????? Why do they compare the performance to Power9, and not Power 10?! And what is the "capacity" they talk about?

Why dont they compare the performance to competing server CPUs from AMD, Intel and ARM?

To be fair, they do mention Power10:

With support for autonomous operations, Power11 delivers intelligent performance gains that reduce complexity and improve workload efficiency. Power11 offers up to 55% better core performance compared to Power9 and has up to 45% more capacity with higher core counts in entry and mid-range systems compared to Power10.

IBM knows that customers are buying Power, X86 and ARM for different reasons. Comparing across ISAs is probably not their focus. Back in the day, Itanium looked like a calcuator compared to X86 in some tasks. In others, it looked like alien technology.

 

[BoggledBeagle]

To be fair, they do mention Power10:

Only in relation to the mysterious "capacity".

 

[Daven]

Only in relation to the mysterious "capacity".

I assume there is a spec sheet somewhere with core counts. I also assume that the IPC is the same between 10 and 11. So with 45% higher core counts (in a single chip), you would expect 45% higher performance if the application usage is linear.

Edit: BTW, @TheLostSwede , you might want to link to the actual press release. Your source link just goes to the Power main page.

IBM Power11 Raises the Bar for Enterprise IT

IBM revealed IBM Power11 - redesigned w/innovations across its processor, HW architecture, & virtualization SW stack, & is designed to deliver the avail., resiliency, perf., & scalability enterprises demand, for seamless hybrid deployment in IBM Cloud.

newsroom.ibm.com

 

[BoggledBeagle]

So the Power11 CPUs are not new at all beside higher core count compared to Power10?

55% improvement from Power9 to Power 11 is in
IBM internal measurements of a commercial core banking solution running on IBM Power E950 compared to an E1150.

Improvement in "capacity" is in
current IBM Power rPerf and CPW estimates for E1150, S1124 and S1122 versus E1050, S1024 and S1022 respectively.

From footnotes in original press release.

 

[Daven]

So the Power11 CPUs are not new at all beside higher core count compared to Power10?

55% improvement from Power9 to Power 11 is in
IBM internal measurements of a commercial core banking solution running on IBM Power E950 compared to an E1150.

Improvement in "capacity" is in
current IBM Power rPerf and CPW estimates for E1150, S1124 and S1122 versus E1050, S1024 and S1022 respectively.

From footnotes in original press release.

I must admit that I have limited knowledge in this area so my definition of capacity is a guess. Processors like Power can really only be compared to past generations of the same processor. The technology is highly specialized and applications are highly optimized for the hardware. Increase in core counts is just one way of many to increase 'capacity'. Just to give you and indication of how specialized IBM hardware can get, look at this photo from 2010:

If you can believe it, those are quad core IBM processors.

IBM ships 5.2GHz chip, its fastest yet

A chip in IBM's new zEnterprise System clocks in at 5.2GHz. The fastest IBM microprocessor to date is targeted at mainframes. It packs in four cores, plus a respectable helping of DRAM.

www.cnet.com

 

[ncrs]

The first Power processors were launched in 1990 long before Nvidia was anything. Power has been in the top500 supercomputer list since the first list published in 1994. Their presence in mainframes and servers was because of the 'Big Iron' era that was dominated by DEC, IBM, Intel (Itanium) and others. Any recent relation to Nvidia was due to shifting compute paradigms that forced a shift in IBM product topography.

I should've written "were on the recent Top500 list", it's what I meant. I am aware of their history before AMD64 started dominating Top500.

Back in the day, Itanium looked like a calcuator compared to X86 in some tasks. In others, it looked like alien technology.

Too bad the aliens didn't leave behind a compiler capable of exploiting EPIC's design.

I must admit that I have limited knowledge in this area so my definition of capacity is a guess. Processors like Power can really only be compared to past generations of the same processor. The technology is highly specialized and applications are highly optimized for the hardware. Increase in core counts is just one way of many to increase 'capacity'. Just to give you and indication of how specialized IBM hardware can get, look at this photo from 2010:

View attachment 407099

If you can believe it, those are quad core IBM processors.

IBM ships 5.2GHz chip, its fastest yet

A chip in IBM's new zEnterprise System clocks in at 5.2GHz. The fastest IBM microprocessor to date is targeted at mainframes. It packs in four cores, plus a respectable helping of DRAM.

www.cnet.com

Most likely capacity refers to how a defined workload fits into a set server size (2U, 4U and so on). For example a 2U server of previous generation can handle streaming 1080p to 200 clients, the next generation can do 300 so that's a 50% capacity increase.

POWER isn't far removed from AMD64 server designs. Similar to how ARM servers relate to AMD64.
What you showed is a mainframe processor, and mainframes do have a radically different design from "standard" computers.

 

[BoggledBeagle]

... look at this photo from 2010:

View attachment 407099

Is that a CPU acupuncturist?

 

[Catzilla]

I worked with Power systems from mid 1990's to a couple years ago writing software for HPC systems. The first IBM HPC systems were SP1 and SP2. I remember them being Power-3 based but Wikipedia shows Power-1 and Power-2.

Power is a RISC based processor, where x86 is CISC. At the time some thought RISC was capable of better performance than CISC.

Originally, these ran IBM's version of UNIX, AIX. I forget whether it was System-V or BSD based (I think some of both). It had a set of sysadmin tools that were rather unique (SMIT) which had a command line and GUI. Someone was allowed to have a sense of humor because when the SMIT gui fired off a command, there was an animated running guy. If the command ran successfully, the running guy put his hands in the air like touchdown and if it failed, the guy fell over. Also, the initial value for registers was 0xDEADBEEF (AIX was an IBM Austin lab project)

The systems IBM built up to 2013 or so when I left were CPU only, and at that time Power-8. The HPC systems also had a proprietary high speed network switch for fast communication between nodes in the system using MPI or in some cases proprietary communication protocols (LAPI)

The HPC systems also started running Linux sometime around 2010 where you could run either AIX or Linux.

I got invited back about 2015 and IBM's last HPC systems (Frontier/Summit) were being built using Power-9 and only ran Linux. These were the first that NVidia GPUs. The CPU was (I think) 44 physical core with 4-way hyperthread, so 176 processor threads. These systems also used Infiniband switches for their internal network.

Every Power processor up to Power 9 was big-endian. Power 9 was little-endian (and I think could be booted in big-endian mode if needed), I see that as not much more than switching the order of bytes when they moved between registers and memory.

Power architecture used to be proprietary, but has been opened up a bit by IBM forming the OpenPower foundation.

A couple other side notes,:

There were plans for OS/2 to run on the RS/6000 processor which was the risc-based predecessor to power but that got dropped.

Also, back in the 80's and early 90's, IBM had add-on hardware for running HPC applications on Mainframes that included a vector processor with up to 512 elements in the vector (AVX-512 has up to 8 double precision and I think Power only had 4-element vactors)

 

[Daven]

I worked with Power systems from mid 1990's to a couple years ago writing software for HPC systems. The first IBM HPC systems were SP1 and SP2. I remember them being Power-3 based but Wikipedia shows Power-1 and Power-2.

Power is a RISC based processor, where x86 is CISC. At the time some thought RISC was capable of better performance than CISC.

Originally, these ran IBM's version of UNIX, AIX. I forget whether it was System-V or BSD based (I think some of both). It had a set of sysadmin tools that were rather unique (SMIT) which had a command line and GUI. Someone was allowed to have a sense of humor because when the SMIT gui fired off a command, there was an animated running guy. If the command ran successfully, the running guy put his hands in the air like touchdown and if it failed, the guy fell over. Also, the initial value for registers was 0xDEADBEEF (AIX was an IBM Austin lab project)

The systems IBM built up to 2013 or so when I left were CPU only, and at that time Power-8. The HPC systems also had a proprietary high speed network switch for fast communication between nodes in the system using MPI or in some cases proprietary communication protocols (LAPI)

The HPC systems also started running Linux sometime around 2010 where you could run either AIX or Linux.

I got invited back about 2015 and IBM's last HPC systems (Frontier/Summit) were being built using Power-9 and only ran Linux. These were the first that NVidia GPUs. The CPU was (I think) 44 physical core with 4-way hyperthread, so 176 processor threads. These systems also used Infiniband switches for their internal network.

Every Power processor up to Power 9 was big-endian. Power 9 was little-endian (and I think could be booted in big-endian mode if needed), I see that as not much more than switching the order of bytes when they moved between registers and memory.

Power architecture used to be proprietary, but has been opened up a bit by IBM forming the OpenPower foundation.

A couple other side notes,:

There were plans for OS/2 to run on the RS/6000 processor which was the risc-based predecessor to power but that got dropped.

Also, back in the 80's and early 90's, IBM had add-on hardware for running HPC applications on Mainframes that included a vector processor with up to 512 elements in the vector (AVX-512 has up to 8 double precision and I think Power only had 4-element vactors)

Thank you, Catzilla, for that excellent overview! As a big enthusiast of 'Big Iron' and supercomputer history, I really loved reading your comment. Whenever I go home to see my parents, I get the opportunity to visit a very unique museum near my parent's house:

Mimms Museum of Technology and Art · 5000 Commerce Pkwy, Roswell, GA 30076

★★★★★ · Museum

maps.app.goo.gl

Mimms Museum of Technology

They have so many nodes and racks of the technology you describe.

 

[Catzilla]

Man I haven’t seen news about POWER in forever. No POWER10 on the top 500 supercomputers. In fact, no new POWER entries on that list at all in the last five years. I’m guessing each version of POWER is for upgrading a fixed number of instances locked into that technology from long ago. I can’t believe anyone is developing new applications based on POWER.

IBM's last HPC systems were the Summit and Sierra systems. Intel, AMD, and NVidia had hardware with good enough performance that Power really wasn't competitive any more.
I guess whatever Power systems are out there now are workstations and servers.

 

[Operandi]

I must admit that I have limited knowledge in this area so my definition of capacity is a guess. Processors like Power can really only be compared to past generations of the same processor. The technology is highly specialized and applications are highly optimized for the hardware. Increase in core counts is just one way of many to increase 'capacity'. Just to give you and indication of how specialized IBM hardware can get, look at this photo from 2010:

View attachment 407099

If you can believe it, those are quad core IBM processors.

IBM ships 5.2GHz chip, its fastest yet

A chip in IBM's new zEnterprise System clocks in at 5.2GHz. The fastest IBM microprocessor to date is targeted at mainframes. It packs in four cores, plus a respectable helping of DRAM.

www.cnet.com

I manage Power hardware as my job so I can speak to this a bit. Power CPUs are their own thing with different design approach compared to how AMD and Intel build a core and the IA is obviously unique but thats not a new thing in this space, you can compare any processor that runs on the same application. For our use case which is high performance database you need four a socket system to match what you can do with a two socket Power10 system and if you needed the type of performance a four socket Power10 system brought there wasn't an option in a single x86 system until recently with what AMD has brought to the space.

IBM's last HPC systems were the Summit and Sierra systems. Intel, AMD, and NVidia had hardware with good enough performance that Power really wasn't competitive any more.
I guess whatever Power systems are out there now are workstations and servers.

Its not so much that they are not competitive its more so that the GPU does most the work so the CPU has become somewhat irrelevant and x86 is simply more ubiquitous and easier to develop for.

 

[R-T-B]

There were plans for OS/2 to run on the RS/6000 processor which was the risc-based predecessor to power but that got dropped.

Slightly offtopic but there are actually a few copies of OS/2 for PPC that somehow shipped and thus, images are available. It's a very weird animal. It includes a complete seamless PPC Win16 / Windows 3.1 VM.

 

[TheinsanegamerN]

Power is a RISC based processor, where x86 is CISC. At the time some thought RISC was capable of better performance than CISC.

The last CISC x86 CPU was the i486. Starting with P5 Pentiums, x86 began running x86 code in micro operations like RISC CPUs.

 

[Catzilla]

The last CISC x86 CPU was the i486. Starting with P5 Pentiums, x86 began running x86 code in micro operations like RISC CPUs.

The instruction set on x86 is CISC. There's translation hardware that internally converts to a set of RISC operations and then the back end in the processor runs those. My understanding is that translation hardware uses a lot of power to run. At best, x86 is a hybrid of some sort.

 

[igormp]

The last CISC x86 CPU was the i486. Starting with P5 Pentiums, x86 began running x86 code in micro operations like RISC CPUs.

The instruction set on x86 is CISC. There's translation hardware that internally converts to a set of RISC operations and then the back end in the processor runs those. My understanding is that translation hardware uses a lot of power to run. At best, x86 is a hybrid of some sort.

Fun fact, even though the x86 ISA is CISC with a front-end that decodes that into µops that are more RISC-like, the mapping between those actual ops to µops on average is pretty close to 1 inst:µop .
Same applies to other "CISC" ISAs such as ARM, even though those rely a lot on instruction fusion to turn multiple ARM instructions into a single µop internally, those also average pretty close to 1 inst:µop.

 

[TheLostSwede]

Some specs

via

https://bsky.app/profile/aschilling.bsky.social%2Fpost%2F3ltkf2hpfzc2h

 

[BoggledBeagle]

Some specs
View attachment 407216

So they claimed 55% over Power 9, while 47% can be attributed just to core count and frequency increases?

That is 6% IPC performance increase AFTER SEVEN YEARS?! IBM is asleep at the wheel. Or it is simply a dead platform and they do not want to invest any effort in it anymore.

Yes, that is just 6. 1,33*1,10*1,06=1,55