Last 12 months IBM introduced particulars about its new Power10 household of processors: eight threads per core, 15 cores per chip, and two chips per socket, with a brand new core microarchitecture, all constructed on Samsung’s 7nm course of with EUV. New applied sciences equivalent to PCIe 5.zero for add-in playing cards, PowerAXON for chip-to-chip interconnect, and OpenCAPI for super-wide reminiscence assist made Power10 sound like a beast, however the query was all the time about time to market – when might clients get one? Today IBM’s Power10 E1080 Servers are being introduced, aimed squarely on the cloud market.
Power10: A Brief Summary
IBM’s Power collection of processors have been a steadfast development over the past couple of many years, typically utilizing some secret manufacturing course of to eke out that specialist frequency proper on the bleeding edge. The new Power10 processor can also be constructed for efficiency, with the 602mm2 16 core silicon die working at over four GHz with eight threads per core. For yield causes one core per chip is disabled, however a full 16 socket system can run to 1920 logical threads.
IBM constructed Power10 to be both in a single chip module (SCM) with one piece of silicon, or a dual-chip module (DCM) with two items of silicon. Where the chip actually shines is within the two multi-protocol connectivity interfaces across the edges of the silicon.
PowerAXON on the corners and OMI (OpenCAPI Memory Interface) sound like wonderful versatile interfaces. Running at 1 TB/sec every, the PowerAXON can be utilized for chip-to-chip communication, storage, common DRAM, ASICs/FPGA connections, and clustered reminiscence. The OMI can be utilized for storage additionally, or foremost DRAM, or for high-bandwidth GDDR/HBM. Together, these applied sciences permit for as much as eight TB per system, or 2048 TB of addressable reminiscence throughout a networked cluster of programs. There’s additionally PCIe 5.zero x32 for add-in playing cards.
IBM compares Power10 towards Power9: +20% single thread enchancment, +30% per core enchancment, and an general 3x efficiency per watt towards the earlier 14nm processor. Also bundled is a brand new AI compute layer supporting 4 512-bit matrix engines and eight 128-bit SIMD engines per core, offering 20x or extra INT8 efficiency per socket.
For extra slides and particulars, test our Live Blog from final 12 months’s Hot Chips.
IBM Power10 E1080
The acknowledged E1080 design is an eight-socket system, supporting clear reminiscence encryption, 2.5x sooner encryption, and new RAS options for superior restoration, self-healing, and diagnostics. IBM’s supplies deal with one specific benchmark: the way it beats different choices in a two-tier SAP Hana SD normal benchmark with solely half the sockets, or an Oracle benchmark requiring solely 20% of the facility and variety of Oracle licenses to attain the identical consequence.
Multiple instances within the launch IBM mentions ‘instant scaling, pay per use consumption’, particularly because it pertains to Red Hat’s OpenShift expertise within the cloud. This all pertains to IBM’s ‘Hybrid Cloud’ technique, which is supposed to imply {that a} enterprise runs some non-public inner cloud assets whereas additionally utilizing some cloud server supplier ‘public’ assets, and it’s the general public ingredient that relates right down to the price. IBM runs its personal cloud service, to which Power10 will probably be a component.
IBM was comparatively mild on particulars about actual SKUs to be supplied, reminiscence choices, whether or not these will probably be out there for direct buy and deployment. A whole lot of discussions went into the brand new AI accelerators with ONNX frameworks in addition to the working system assist for enterprise options equivalent to side-channel assaults, intrusion detection, compliance reporting, and full-stack encryption with assist for ‘quantum-safe cryptography’.
IBM is taking orders now with shipments anticipated to start earlier than the top of the…
