Difference between revisions of "POWER9"
(→Nimbus) |
|||
Line 17: | Line 17: | ||
|data8 = January 2018 | |data8 = January 2018 | ||
|label9 = Production stepping(s) | |label9 = Production stepping(s) | ||
− | |data9 = DD2.2 | + | |data9 = DD2.2, DD2.3 |
|label10 = [[POWER8E|← POWER8E]] | |label10 = [[POWER8E|← POWER8E]] | ||
|data10 = POWER10 → | |data10 = POWER10 → |
Revision as of 17:00, 28 January 2020
POWER9 | |
---|---|
Power ISA | 3.0B |
Process node | 14nm |
Maximum slices | 24 |
Maximum cores | 12 SMT8 / 24 SMT4 |
L2 cache / slice | 512kB |
L3 cache / slice | 10MB |
Production availability | January 2018 |
Production stepping(s) | DD2.2, DD2.3 |
← POWER8E | POWER10 → |
POWER9 is IBM's most recent POWER compatible server and workstation CPU (POWER ISA v3.0B). Built on a 14nm process, each CPU package can contain up to 24 SMT4 cores or 12 SMT8 cores. Each pair of SMT4 cores, or singleton SMT8 core, comprises a slice; each slice in turn contains 512kB L2 cache and 10MB L3 cache. Raptor Computing Systems' 4- and 8-core processors provide unpaired cores, such that one SMT4 core per slice is fused off. This allows each of the SMT4 cores to utilize the full cache of the slice exclusively, increasing performance for these ST-focused processors.
Contents
Process
POWER9 is fabricated using the GlobalFoundries 14HP (High Performance) process. This is distinct from the GlobalFoundries 14LPP (Low Power) process used by other GF 14nm customers, and is believed to be an IBM-specific process using ex-IBM Microelectronics intellectual property. The process is also used for the CPUs in IBM's z14 mainframes. A detailed discussion of 14HP and how it differs from 14LPP is available here.
Chips
There are three known silicon masks of POWER9:
- Nimbus (POWER9 Scale Out)
- Cumulus (POWER9 Scale Up)
- Axone (POWER9′ ("POWER9 Prime"), aka POWER9 with Advanced I/O)
Nimbus is the “scale out” variant and uses direct-attach DDR4 memory. Cumulus is the “scale up” version and uses Centaur memory buffers, allowing larger amounts of memory to be attached to a system.
Chips can be fused as SMT4 or SMT8 during manufacturing. The SMT8 variant essentially fuses each pair of cores into one “core”, halving the core count while doubling the number of threads per core. SMT4 variants are intended for PowerNV platforms running Linux, and SMT8 variants are intended for use with IBM's PowerVM hypervisor which can run Linux, AIX or IBM i.[1]
Modules
Chip | Module | Memory Channels | XBUS Lanes | PCIe Lanes | OpenCAPI Lanes | Socket |
---|---|---|---|---|---|---|
Nimbus | Sforza | 4 | 1 | 48 | 0 | LGA 2601 |
Monza | 8 | 1 | 34 | 48 | LGA 3899 | |
LaGrange | 8 | 2 | 42 | 16 | LGA 3899 | |
Cumulus | (unknown) | (memory attached via Centaurs) | (unknown) | (unknown) | (unknown) | ? |
Axon | (unknown) | (memory attached via OMI) | Up to 3 | Up to 48 | Up to 48 | ? |
XBUS is used for inter-processor communication on dual-socket system
Nimbus
Nimbus chips are available in three different modules: Sforza, Monza, and LaGrange. Each module uses the same silicon mask but is packaged differently, exposing different I/O functionality to the host platform, allowing purpose-built systems to be constructed in addition to more general-purpose computers.
Sforza is the most flexible of these packages, providing PCIe 4.0 lanes as the main I/O resource, and is what Talos™ II uses for maximal similarity to existing desktop, workstation, and server systems.
Monza modules offer the most OpenCAPI/NVLink bandwidth and are used in IBM's AC922 (Witherspoon) systems, such as those used by the Sierra and Summit supercomputers.
LaGrange modules offer increased XBus bandwidth between processor sockets and are used by the Google/Rackspace Zaius motherboard used in the Barreleye G2 system.[2]
Part numbers for different POWER9 Sforza SKUs can be found on page 58 of the datasheet. These part numbers are printed on the surface of the CPU module and can be used to determine the type of the CPU.
Several revisions of the Nimbus mask have been issued:
- DD2.1 was the final preproduction revision before GA. It has errata preventing the use of hardware virtualization, but DD2.1 Sforza can be used in e.g. the Talos II if this functionality is not needed.
- DD2.2 is the first GA revision of Nimbus. DD2.2 Sforza is sold at raptorcs.com.
- DD2.3 is an updated revision of Nimbus. They added Ultravisor functionality and Hardware watchpoint support, also improved Meltdown and Spectre mitigations. DD2.3 parts are sold at raptorcs.com under the designation POWER9 v2.
Cumulus
Little is known about Cumulus chips at this time; as Scale Up chips, they will trade some I/O bandwidth for support for more than two sockets.[3]
Axone
Branded POWER9′ ("POWER9 Prime"), also known as POWER9 with Advanced I/O. Newly announced in August 2019. Uses serial memory attachment via OMI, an evolution from the Centaur.
References
- ↑ Stuecheli, Jeff. POWER9. Presentation for AIX VUG. (video download, slides, timemarks)
- ↑ Gangidi, Adi Zaius/Barreleye G2 Server Development Update. 2017-11-13
- ↑ Morgan, Timothy Prickett. POWER9 to the People. 2017-12-05
Resources
- POWER9 CPU and Platform Documentation
- POWER9 Hardware Compatibility List
- Basic POWER9 overview presentation
- Power ISA version 3.0B - implemented by POWER9