Google joins the custom CPU crowd with Arm-based Axion chips
(2024/04/09)
- Reference: 1712678407
- News link: https://www.theregister.co.uk/2024/04/09/google_custom_cpu/
- Source link:
Cloud Next Google lifted the veil on its long-rumored Arm CPU, dubbed Axion, during its Cloud Next event in Las Vegas on Tuesday.
The search giant has a history of building custom silicon going back to 2015. However, up to this point most of its focus has been on developing faster and more capable Tensor Processing Units (TPUs) to accelerate its internal machine learning workloads.
Now Google has [1]seen fit to build a custom CPU of its own. Specifically, Axion is based on the British chip designer's Neoverse V2 design, which uses the standard ARMv9 architecture and instruction set. However, beyond this, details like core count, process tech, cache or memory configuration remain unspecified. With that said, a rendering of the chip does imply a monolithic design.
[2]
Google is the latest US cloud provider to launch a custom Arm CPU – click to enlarge
We've asked Google for more information, but do know that Arm's Neoverse V2 architecture can support core configurations of up to 256, though Nvidia and AWS's V2-based designs have topped out 72 and 96 cores, respectively.
What Google has said is Axion will deliver about 30 percent better performance than "the fastest general purpose Arm-based instances" available today, which we suspect either refers to Google's Ampere VMs or possibly Amazon's Neoverse V1-based Graviton 3 instances.
[3]
Pitted against Intel and AMD's "comparable current-generation" processors, Axion is up to 50 percent faster and 60 percent more efficient, or so Google claims, rather ambiguously citing "Google Cloud internal data." So take these claims with a grain of salt.
[4]
[5]
Axion is designed to work in conjunction with Google's custom Titanium [6]microcontrollers , which essentially function as a data processing unit (DPU) to offload compute-intensive operations like networking, security, and storage IO, freeing up the processor to run customer workloads.
[7]SiFive is back with another 64-bit RISC-V dev board – hopefully
[8]Arm targets AI performance with latest Neoverse Compute Subsystems
[9]TSMC boss says one-trillion transistor GPU is possible by early 2030s
[10]Standardization could open door to third-party chiplets in AMD designs
As for where we'll see Axion deployed, Google says it's already begun transitioning many of its cloud and ad platform services, including BigTable, Spanner, BigQuery, Earth Engine, and YouTube ads to Arm systems in anticipation of migrating them to its in-house silicon.
For customer workloads, Google says the chip is ideally suited for web and app servers, microservices, databases, in-memory caches, data analytics, media processing, and CPU-based AI training and inferencing. Instances based on Axion are slated to go live later this year.
Google is the latest US cloud provider to roll its own CPUs. In fact, it's rather late to the party. Amazon's Graviton processors, which made their [11]debut at re:Invent in 2018, are now on their fourth generation. Meanwhile, Microsoft [12]revealed its own Arm chip, dubbed the Cobalt 100, last fall. ®
Get our [13]Tech Resources
[1] https://cloud.google.com/blog/products/compute/introducing-googles-new-arm-based-cpu/
[2] https://regmedia.co.uk/2024/04/09/google_axion_cpu.jpg
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZhW6h2NYAFNNiIQjz-G9CwAAAAQ&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZhW6h2NYAFNNiIQjz-G9CwAAAAQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZhW6h2NYAFNNiIQjz-G9CwAAAAQ&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://cloud.google.com/titanium?e=48754805&hl=en
[7] https://www.theregister.com/2024/04/09/sifive_riscv_hifive/
[8] https://www.theregister.com/2024/02/21/arm_neoverse_compute/
[9] https://www.theregister.com/2024/04/01/tsmc_one_trillion_transistor/
[10] https://www.theregister.com/2024/03/27/amd_chiplets_future/
[11] https://www.theregister.com/2018/11/27/amazon_aws_graviton_specs/
[12] https://www.nextplatform.com/2023/11/15/microsoft-holds-chip-makers-feet-to-the-fire-with-homegrown-cpu-and-ai-chips/
[13] https://whitepapers.theregister.com/
The search giant has a history of building custom silicon going back to 2015. However, up to this point most of its focus has been on developing faster and more capable Tensor Processing Units (TPUs) to accelerate its internal machine learning workloads.
Now Google has [1]seen fit to build a custom CPU of its own. Specifically, Axion is based on the British chip designer's Neoverse V2 design, which uses the standard ARMv9 architecture and instruction set. However, beyond this, details like core count, process tech, cache or memory configuration remain unspecified. With that said, a rendering of the chip does imply a monolithic design.
[2]
Google is the latest US cloud provider to launch a custom Arm CPU – click to enlarge
We've asked Google for more information, but do know that Arm's Neoverse V2 architecture can support core configurations of up to 256, though Nvidia and AWS's V2-based designs have topped out 72 and 96 cores, respectively.
What Google has said is Axion will deliver about 30 percent better performance than "the fastest general purpose Arm-based instances" available today, which we suspect either refers to Google's Ampere VMs or possibly Amazon's Neoverse V1-based Graviton 3 instances.
[3]
Pitted against Intel and AMD's "comparable current-generation" processors, Axion is up to 50 percent faster and 60 percent more efficient, or so Google claims, rather ambiguously citing "Google Cloud internal data." So take these claims with a grain of salt.
[4]
[5]
Axion is designed to work in conjunction with Google's custom Titanium [6]microcontrollers , which essentially function as a data processing unit (DPU) to offload compute-intensive operations like networking, security, and storage IO, freeing up the processor to run customer workloads.
[7]SiFive is back with another 64-bit RISC-V dev board – hopefully
[8]Arm targets AI performance with latest Neoverse Compute Subsystems
[9]TSMC boss says one-trillion transistor GPU is possible by early 2030s
[10]Standardization could open door to third-party chiplets in AMD designs
As for where we'll see Axion deployed, Google says it's already begun transitioning many of its cloud and ad platform services, including BigTable, Spanner, BigQuery, Earth Engine, and YouTube ads to Arm systems in anticipation of migrating them to its in-house silicon.
For customer workloads, Google says the chip is ideally suited for web and app servers, microservices, databases, in-memory caches, data analytics, media processing, and CPU-based AI training and inferencing. Instances based on Axion are slated to go live later this year.
Google is the latest US cloud provider to roll its own CPUs. In fact, it's rather late to the party. Amazon's Graviton processors, which made their [11]debut at re:Invent in 2018, are now on their fourth generation. Meanwhile, Microsoft [12]revealed its own Arm chip, dubbed the Cobalt 100, last fall. ®
Get our [13]Tech Resources
[1] https://cloud.google.com/blog/products/compute/introducing-googles-new-arm-based-cpu/
[2] https://regmedia.co.uk/2024/04/09/google_axion_cpu.jpg
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZhW6h2NYAFNNiIQjz-G9CwAAAAQ&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZhW6h2NYAFNNiIQjz-G9CwAAAAQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZhW6h2NYAFNNiIQjz-G9CwAAAAQ&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://cloud.google.com/titanium?e=48754805&hl=en
[7] https://www.theregister.com/2024/04/09/sifive_riscv_hifive/
[8] https://www.theregister.com/2024/02/21/arm_neoverse_compute/
[9] https://www.theregister.com/2024/04/01/tsmc_one_trillion_transistor/
[10] https://www.theregister.com/2024/03/27/amd_chiplets_future/
[11] https://www.theregister.com/2018/11/27/amazon_aws_graviton_specs/
[12] https://www.nextplatform.com/2023/11/15/microsoft-holds-chip-makers-feet-to-the-fire-with-homegrown-cpu-and-ai-chips/
[13] https://whitepapers.theregister.com/