AMD K12: +20% IPC over Zen1, Jim Keller's mythical ARM server CPU from AMD скачать в хорошем качестве

AMD K12: +20% IPC over Zen1, Jim Keller's mythical ARM server CPU from AMD

Rare info about architecture of mythical (canceled) K12 processor core from AMD. Info is based on non-public presentation slides for server customers. The K12 was based on new and modern 64-bit ARMv8 instruction set and was wider core with higher IPC approximately 20% over AMD Zen 1 (x86) sister core. Both cores was aiming for similar core area. Simple RISC decode in compare to complex x86 allowed to spare transistors and use them for more execution units inside the K12 core. Resulting in two additional pipes in integer back-end core. One was 5th pipe with separated Branch Unit to standard 4xALU pipes. Second was 3rd memory unit. This 8-wide integer core (ARM K12) in compare to 6-wide x86 Zen 1 was theoretically able to provide up to +33% IPC (uops/cycle). Effectively demonstrating higher performance per transistor - typical ARM ISA advantage. As x86-64 ISA needs to keep backward compatibility with 1990 chips still, the new (2012) ARM 64-bit ARMv8 instruction set was clean sheet architecture aimed for high performance computing. ARM released in 2016 SVE SIMD extension supporting HW FPUs up to 2048-bit wide for vector and matrix computing. ARM instruction set is way superior over x86-64 since 2016 and even more breakthrough extensions for servers and HP computing was introduced (SVE2 and TME in 2019). Cancelling K12 core after tape out was clearly big mistake in long term time frame. Father of this K12 core was Jim Keller who came into AMD from Apple and he was one of few people who was aware of outdated, proprietary and stagnating x86 and new rising ARM star. Apple's ARM core are the most advanced architecture on the world right now and Jim Keller was the father of those early Apple ARM cores (A7, A8, A9, A10). Funny thing is that back at 2015 ẗhe canceling of K12 looked like a right thing to do. But from 2020 perspective it looks like horrible mistake. Well, that's why world need visioners like Jim Keller. /*******************/ 2021/8/22 UPDATE: /*******************/ Today's Zen 3 architecture is basically copy of K12. That's horrible situation for AMD as they could have performance of Zen 3 on market back in 2017 as K12. Since cancelling of K12 in 2015 they spend 5 years at reworking K12 to x86 Zen 3. Here you can see how much time and effort it takes to compensate all disadvantages of such a old ISA like x86-64 for similar IPC. And Zen 3 has still only 4x decoders, while Apple M1 has 8x decoders and even cheap license ARM's Cortex X1 core has 5x decoders (X1 has about +6% IPC above Zen 3). Terrible. Another thing is that back in 2015 AMD knew the 2048-bit SVE vectors are in development while x86 has no alternative to it. SVE was officially announced in 2016 but it's clear that SVE spec was finished in 2015. First SVE capable CPU - Fujitsu A64FX CPU for Japanese super computer Fugaku (today's fastest SC on the world despite having no GPU) - showed up at the end of 2020. All new ARM CPUs from Apple, Nuvia-Qualcomm, Ampere and ARM will have 2048-bit SVE2 vector support since 2022. ALL. Even little cores, thanks to SVE scalability down to 128-bit HW length while still being able to run any SW written for 2048-bit vectors. Dynamically allocated SVE against ancient static allocated x86's AVX512. Instead of developing K13 with SVE vectors AMD did reworked K12 for old and already dead x86. Unbelievable example of wasting time and resources. Further more, as Zen 3 has only 4xINT + 1xBranch exe units for scalar core, the next Zen 4 will have similar narrow setup. Zen 4 might beef up Branch unit for INT capability to have total 5x INT similar to Intel's Golden Cove. But only 5 port for scalar in 2022 sounds like a bad joke. ARM's Cortex A77 has 4+2 since 2019 and Apple has monstrous 6+2 since 2017. Next year Cortex X3 based on Makalu uarch will be similar to Apple 6+2 and Apple will have new 12x port uarch this year. Best x86 CPUs like Zen4 and Golden Cove with narron 5x port will be massacred by cheap license 8x port Makalu Cortex X3 core (and its server version Neoverse V2). And chainsaw massacred by new Apple's and Nuvia's 12x port uarchs (with almost double IPC). No wonder that Google and MS are preparing their own server and mobile CPUs based on new license ARMs. x86 is really dead. Jim Keller was absolutely right in everything. Only it takes 10 years until the truth will get to surface. Too bad that AMD will face financial troubles again thanks to such a terrible CEO like Lisa Su. But all bad things lead to something good finally. Maybe Jim Keller will return into AMD for the 3rd time to save them :-) This time I hope Jim will do that like Steve Jobs did when forcing Apple to buy his Next, similarly AMD to buy Jim's Tenstorrent so Jim Keller will become biggest AMD share holder. Legendary Jim Keller as CEO and owner of AMD while acquiring best AI/ML architecture on market. That would be cool.