Wiki Article
Google Axion
Nguồn dữ liệu từ Wikipedia, hiển thị bởi DefZone.Net
| Launched | October 2024 |
|---|---|
| Designed by | |
| Manufactured by | |
| Fabrication process | |
| Platform(s) | |
Instructions and architecture | |
| Instructions set | Arm |
| Instructions | AArch64 (Armv9-A) |
| Extensions | |
| Core architecture |
|
Cores | |
| Core count | Up to
|
| L1 cache | 128 KB per core (C4A):
|
| L2 cache | 2 MB per core (C4A) |
| L3 cache | 80 MB shared system cache (C4A) |
Memory support | |
| Type | DDR5 |
| Maximum capacity |
|
Google Axion is a family of custom Arm-based server microprocessors designed by Google for its Google Cloud data centers.[2] Announced in April 2024 at Google Cloud Next, Axion is Google's first custom general-purpose data center CPU, following earlier application-specific processors such as the Tensor Processing Unit (TPU) for machine learning and Google Tensor for mobile devices.[3]
Built on Arm Neoverse microarchitectures, Axion was developed to compete with other hyperscaler-designed Arm server processors, including Amazon Web Services' AWS Graviton and Microsoft's Microsoft Azure Cobalt.[2][3] Following an initial preview period, the first Axion-powered virtual machines (the C4A series) entered general availability in October 2024, followed by bare-metal (C4A Metal) and general-purpose scale-out (N4A) configurations in late 2025.[4][5]
History and background
[edit]Prior to introducing Axion, Google developed several custom application-specific integrated circuits (ASICs) for its cloud infrastructure, beginning with the Tensor Processing Unit (TPU) in 2015 and Video Coding Units (VCUs) for YouTube transcoding.[3] For Arm-based general-purpose cloud instances, Google previously relied on third-party merchant silicon, introducing the Tau T2A instance family in 2022 powered by Ampere Computing's Altra processors.[4]
Google officially announced the Axion processor on April 9, 2024, during its Google Cloud Next conference.[2] Industry analysts noted that the release aligned Google with rival cloud providers Amazon Web Services (which launched its first AWS Graviton chip in 2018) and Microsoft (which announced its Arm-based Cobalt 100 CPU in late 2023) in shifting internal workloads and customer virtual machines away from third-party x86 processors.[2][3]
In October 2024, the first Axion-based virtual machine series, C4A, reached general availability across several North American, European, and Asian cloud regions.[6] In November 2025, Google expanded the Axion lineup with the introduction of N4A virtual machines for scale-out workloads and C4A Metal bare-metal instances.[5]
Architecture
[edit]The first-generation Axion processor powering the C4A series is based on the Arm Neoverse V2 core design and implements the 64-bit Armv9-A instruction set architecture.[3] Subsequent Axion processors deployed in Google's N4A instance family utilize the Arm Neoverse N3 core.[7] According to semiconductor research firm TrendForce, Axion is manufactured by TSMC using its 3 nm (N3) fabrication process.[1] During the physical design of the chip, engineers utilized AlphaChip, a reinforcement learning floorplanning system developed by Google DeepMind, to lay out several of the processor's circuit blocks.[8]
Because Armv9-A eliminates hardware support for 32-bit execution at the core level, Axion operates exclusively as a 64-bit (aarch64) processor.[4] On the Neoverse V2-based C4A platform, each core incorporates Scalable Vector Extension 2 (SVE2) with quad 128-bit SIMD execution units, cryptographic acceleration extensions (AES and SHA), 64 KB of L1 instruction cache, 64 KB of L1 data cache, and 2 MB of private L2 cache, backed by an 80 MB shared system L3 cache and DDR5 memory.[4][5]
Titanium offload system
[edit]Axion instances are paired with Google's custom Titanium system, an architecture of purpose-built microcontroller data processing units (DPUs) and tiered hardware offloads comparable to AWS Nitro.[3] Titanium offloads hypervisor management, network processing, security operations, and block storage I/O (such as Google's Hyperdisk storage) from the main Axion CPU cores so that host compute resources remain dedicated to customer workloads.[3][9]
Cloud instance families
[edit]Axion processors are not sold commercially as standalone hardware; they are accessed exclusively through Google Cloud services such as Google Compute Engine, Google Kubernetes Engine, and AlloyDB.[6][9]
| Series | Type | Core architecture | Max vCPUs | Max memory (DDR5) | Max network bandwidth | Availability |
|---|---|---|---|---|---|---|
| C4A | Compute-optimized VM | Arm Neoverse V2 | 72 | 576 GB | 100 Gbps | October 2024[6][9] |
| C4A Metal | Bare-metal server | Arm Neoverse V2 | 96 | 768 GB | 100 Gbps | November 2025 (Preview)[5] |
| N4A | General-purpose scale-out VM | Arm Neoverse N3 | 64 | 512 GB | 50 Gbps | Late 2025 / Early 2026[10][7] |
Performance and reception
[edit]At launch, Google stated that Axion delivered up to 30% higher performance than the fastest general-purpose Arm instances then available in the cloud, and up to 50% better performance and 60% better energy efficiency than comparable x86-based virtual machines.[2]
In independent Linux benchmarking conducted by Phoronix in October 2024, Axion-powered C4A instances demonstrated substantial generational gains over Google's prior Ampere Altra-based T2A instances and delivered competitive raw performance and higher performance-per-dollar against Intel Xeon Emerald Rapids (C4) instances across web server, compilation, and database workloads.[4] However, Phoronix noted that for specific heavy AVX-512 workloads, Intel's x86 processors retained an advantage over Axion's quad 128-bit SVE2 implementation.[4]
Subsequent independent testing of the lower-cost N4A instances in February 2026 against AMD EPYC Turin (N4D) and Intel Xeon (N4) instances found that Axion led in energy efficiency and cost-effectiveness across a majority of multi-threaded cloud and cryptography workloads.[10]
See also
[edit]References
[edit]- ^ a b c "Google's Axion CPU Reportedly Built on TSMC's 3nm, Set to Drive Foundry's Data Center Revenue Growth". TrendForce. October 21, 2025. Retrieved October 10, 2026.
- ^ a b c d e Shilov, Anton (April 10, 2024). "Google unveils Axion CPU for datacenters: claims up to 50% better performance than x86 processors". Tom's Hardware. Retrieved October 10, 2026.
- ^ a b c d e f g Kennedy, Patrick (April 10, 2024). "Google Axion for Arm Cloud Compute". ServeTheHome. Retrieved October 10, 2026.
- ^ a b c d e f Larabel, Michael (October 30, 2024). "Benchmarks Of Google's Axion Arm-based CPU: Competitive Performance & Compelling Value". Phoronix. Retrieved October 10, 2026.
- ^ a b c d Shilov, Anton (November 6, 2025). "Google deploys new Axion CPUs and seventh-gen Ironwood TPU". Tom's Hardware. Retrieved October 10, 2026.
- ^ a b c Losio, Renato (November 24, 2024). "First Google Axion Processor Now Available: Claims Best Performance in Cloud Market". InfoQ. Retrieved October 10, 2026.
- ^ a b "Arm VMs on Compute". Google Cloud Documentation. Retrieved October 10, 2026.
- ^ Moss, Sebastian (November 18, 2024). "Google DeepMind publishes paper refuting criticism of AI chip design platform AlphaChip". Data Center Dynamics. Retrieved October 10, 2026.
- ^ a b c "Google Axion processors". Google Cloud. Retrieved October 10, 2026.
- ^ a b Larabel, Michael (February 24, 2026). "Google Cloud N4 Series Benchmarks: Google Axion vs. Intel Xeon vs. AMD EPYC Performance". Phoronix. Retrieved October 10, 2026.