
Altera 推出了 Agilex 27 FPGA 和 SoC 系列,该系列专为 AI 基础设施、高速网络、6G 和卫星通信以及航空航天和国防应用而设计。该系列包括 14 种器件型号,每种均提供纯 FPGA 器件或集成 Arm 处理器的 FPGA SoC 版本。
新器件基于 Altera 的第三代 HyperFlex 架构,该公司声称其与上一代 Agilex 7 FPGA 相比,每瓦性能提升 1.8 倍。其他特性包括高达 1760 万个逻辑单元、第二代 AI Tensor DSP 模块、SoC FPGA 版本中的 Arm Cortex-A720/A520 处理器子系统、256 Gbps 收发器、PCIe 7.0、1.6TbE 以太网和 LPDDR6 内存支持。
Altera Agilex 27 FPGA 和 SoC 系列规格:
FPGA 架构 – 第三代 HyperFlex 架构,具有增强的自适应逻辑模块(ALM)
866K 至 1760 万个逻辑单元(LE)
274,190 至 5,683,872 个 ALM
与上一代 Agilex 7 相比,每瓦性能最高提升 1.8 倍
处理器(仅限 SoC 版本) – Armv9.2 硬处理器系统(HPS)
双核 Arm Cortex-A720 内核和双核 Cortex-A520 内核,CPU 速度高达 2.3 GHz
2MB 共享 L3 缓存和 512KB 具有 ECC 功能的片上 RAM
1,096 至 21,048 个第二代 AI Tensor DSP 模块
高达 1,211 TOPS INT8 性能
支持 INT8/9/16/32、BFLOAT16、FP16/32、TF32 和复数运算
FPGA 40Mb 至 822Mb M20K 内存
63Mb 至 1,096Mb M80K 内存
8Mb 至 158Mb MLAB 分布式内存
外部内存 – DDR5、LPDDR4、LPDDR5/5X、LPDDR6
3 至 18 个具有在线内存加密(IME)功能的硬内存控制器
高达 9.84 Tbps 的外部内存带宽
显示 – 基于 MIPI D-PHY 的显示连接;支持 LVDS
摄像头 – MIPI C-PHY 2.1 和 D-PHY 3.0,每通道高达 4.5 Gbps,用于高速摄像头和图像传感器连接
具有硬化 MAC、PCS 和 FEC 的版本最高支持 10x 800GbE 或 4x 1.6TbE
SoC 版本上的 3x TSN 以太网 MAC
PCIe 6.0 x8(在选定器件上最多支持 2 个实例)
PCIe 7.0 x8 / PCIe 6.0 x16(在选定器件上最多支持 6 个实例)
支持根端口 (RP)、端点 (EP)、分叉 (bifurcation) 以及完整性与数据加密 (IDE)
具有 GPIO 的高速和中速 I/O 组
高达 32x GT SerDes 收发器 @ 256 Gbps
高达 176x GX SerDes 收发器 @ 128 Gbps
GS SerDes 收发器,速率高达 32 Gbps
用于 Chiplet 连接的 UCIe 接口
其他 – DMA 控制器、定时器、看门狗和 GPIO
具有硬件信任根、安全启动和器件认证功能的安全管理器 (SDM)
支持符合 CNSA 2.0 的后量子密码技术 (PQC)
受保护的密钥存储和物理防篡改保护
AES/SM4 加密、MACsec、PCIe IDE 和在线内存加密
FIPS 140-3 Level 2 可认证选项
封装 – BGA 封装,尺寸从 27 × 27 mm 到 65 × 65 mm
Arm Cortex-A720 和 Arm Cortex-A520 HPS 框图
Agilex 27 系列采用了新的 HyperFlow 数据网络,连接 FPGA 结构、内存控制器、Arm 处理器、PCIe 和其他接口。它使用专用数据路径和 DMA 引擎来传输数据,而不占用 FPGA 逻辑资源,减少了路由拥塞,并为 AI、信号处理和其他任务留下了更多可用资源。该器件还采用第三代 HyperFlex 架构以提高性能和逻辑利用率,以及用于 AI 和信号处理中整数、浮点、向量和张量运算的第二代 AI Tensor DSP 模块。
对于机器人和物理 AI 应用,这些 FPGA 可以处理传感器同步、图像和信号预处理,以及向外部 CPU 或 GPU 传输数据。这使得时间关键的控制和安全操作可以与 AI 处理工作负载分开运行。Altera 还提供具有封装内内存的 Agilex 27 M 系列器件,以及通过 UCIe chiplets 集成数据转换器的 Direct RF 系列器件,不过这些版本的详细规格尚未公布。
Altera 的 Quartus Prime Pro Edition FPGA 开发软件支持 Agilex 27 系列。最新的 26.3 版本还增加了 QuartusIQ,这是一个 AI 助手,可协助进行 FPGA 设计、调试、查找文档,并基于当前项目提出改进建议。它可以直接在 Quartus GUI 中使用,也可以通过 Model Context Protocol (MCP) 连接外部 AI 工具。
此前,我们曾介绍过英特尔具有四核 Cortex-A53 处理器和高达 112 Gbps 收发器的 Agilex SoC FPGA 系列,以及近期推出的集成 64 Gsps RF 转换器的 Altera Agilex 9 Direct RF 系列 AGRW039 SoC FPGA。新的 Agilex 27 系列增加了逻辑容量和收发器速度,同时增加了 Arm Cortex-A720/A520 处理能力、更新的内存接口和 PCIe 7.0 支持。
该公司表示,Quartus Prime 对 Agilex 27 的设计支持将于 2026 年 11 月中旬向早期访问客户提供,工程样片预计于 2027 年下半年推出。QuartusIQ 已作为 Quartus Prime Pro Edition 26.3 的一部分提供。价格尚未公布。更多详情可在产品页面和官方公告中查阅。
本文来源:https://www.cnx-software.com/2026/11/05/altera-agilex-27-fpga-soc-family-features-up-to-17-6m-logic-elements-cortex-a720-a-520-cores-256-gbps-serdes-pcie-7-0/
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Altera has introduced the Agilex 27 FPGA and SoC family designed for AI infrastructure, high-speed networking, 6G and satellite communications, and aerospace and defense applications. The family includes 14 device models, each available as either an FPGA-only device or an FPGA SoC with integrated Arm processors
The new devices rely on Altera’s third-generation HyperFlex architecture, which the company claims delivers 1.8 times the performance per Watt compared to the previous-generation Agilex 7 FPGAs. Other features include up to 17.6 million logic elements, second-generation AI Tensor DSP blocks, an Arm Cortex-A720/A520 processor subsystem in the SoC FPGA variants, 256 Gbps transceivers, PCIe 7.0, 1.6TbE Ethernet, and LPDDR6 memory support.
Altera Agilex 27 FPGA and SoC family specifications:
FPGA architecture – 3rd-generation HyperFlex architecture with enhanced adaptive logic modules (ALMs) 866K to 17.6 million logic elements (LEs) 274,190 to 5,683,872 ALMs Up to 1.8× better performance per watt compared to the previous-generation Agilex 7
Up to 1.8× better performance per watt compared to the previous-generation Agilex 7
Processor (SoC variants only) – Armv9.2 Hard Processor System (HPS) Dual Arm Cortex-A720 cores and dual Cortex-A520 cores, with CPU speeds up to 2.3 GHz 2MB shared L3 cache and 512KB on-chip RAM with ECC
Dual Arm Cortex-A720 cores and dual Cortex-A520 cores, with CPU speeds up to 2.3 GHz
2MB shared L3 cache and 512KB on-chip RAM with ECC
AI and DSP acceleration 1,096 to 21,048 second-generation AI Tensor DSP blocks Up to 1,211 TOPS INT8 performance Supports INT8/9/16/32, BFLOAT16, FP16/32, TF32, and complex arithmetic Fixed-point, floating-point, tensor, and vector processing modes
1,096 to 21,048 second-generation AI Tensor DSP blocks
Supports INT8/9/16/32, BFLOAT16, FP16/32, TF32, and complex arithmetic
Fixed-point, floating-point, tensor, and vector processing modes
Memory FPGA 40Mb to 822Mb M20K memory 63Mb to 1,096Mb M80K memory 8Mb to 158Mb MLAB distributed memory External memory – DDR5, LPDDR4, LPDDR5/5X, LPDDR6 3 to 18 hard memory controllers with inline memory encryption (IME) Up to 9.84 Tbps external memory bandwidth Memory-in-package options on M-Series devices
FPGA 40Mb to 822Mb M20K memory 63Mb to 1,096Mb M80K memory 8Mb to 158Mb MLAB distributed memory
External memory – DDR5, LPDDR4, LPDDR5/5X, LPDDR6
3 to 18 hard memory controllers with inline memory encryption (IME)
Memory-in-package options on M-Series devices
Display – MIPI D-PHY-based display connectivity; LVDS support
Camera – MIPI C-PHY 2.1 and D-PHY 3.0, up to 4.5 Gbps per lane, for high-speed camera and image sensor connectivity
Networking Up to 10x 800GbE or 4x 1.6TbE with hardened MAC, PCS, and FEC Hardware MACsec support 3x TSN Ethernet MACs on SoC variants
Up to 10x 800GbE or 4x 1.6TbE with hardened MAC, PCS, and FEC

Expansion PCIe 6.0 x8 (up to 2 instances on selected devices) PCIe 7.0 x8 / PCIe 6.0 x16 (up to 6 instances on selected devices) Root Port (RP), Endpoint (EP), bifurcation, and Integrity and Data Encryption (IDE) support High-speed and medium-speed I/O banks with GPIO High-speed transceivers Up to 32x GT SerDes transceivers @ 256 Gbps Up to 176x GX SerDes transceivers @ 128 Gbps GS SerDes transceivers @ up to 32 Gbps NRZ and PAM4 signaling UCIe interface for chiplet connectivity
PCIe 6.0 x8 (up to 2 instances on selected devices)
PCIe 7.0 x8 / PCIe 6.0 x16 (up to 6 instances on selected devices)
Root Port (RP), Endpoint (EP), bifurcation, and Integrity and Data Encryption (IDE) support
High-speed and medium-speed I/O banks with GPIO
High-speed transceivers Up to 32x GT SerDes transceivers @ 256 Gbps Up to 176x GX SerDes transceivers @ 128 Gbps GS SerDes transceivers @ up to 32 Gbps NRZ and PAM4 signaling UCIe interface for chiplet connectivity
Misc – DMA controllers, timers, watchdogs, and GPIO
Security Secure Device Manager (SDM) with hardware root of trust, secure boot, and device attestation Post-quantum cryptography (PQC) support aligned with CNSA 2.0 Protected key storage and physical anti-tamper protection AES/SM4 encryption, MACsec, PCIe IDE, and inline memory encryption FIPS 140-3 Level 2 certifiable options
Secure Device Manager (SDM) with hardware root of trust, secure boot, and device attestation
Post-quantum cryptography (PQC) support aligned with CNSA 2.0
Protected key storage and physical anti-tamper protection
AES/SM4 encryption, MACsec, PCIe IDE, and inline memory encryption
Package – BGA packages from 27 × 27 mm to 65 × 65 mm
Arm Cortex-A720 and Arm Cortex-A520 HPS Block Diagram Agilex 27 Block Diagram
The Agilex 27 family features a new HyperFlow data network that connects the FPGA fabric, memory controllers, Arm processors, PCIe, and other interfaces. It uses dedicated data paths and DMA engines to transfer data without using FPGA logic resources, reducing routing congestion and leaving more resources available for AI, signal processing, and other tasks. The devices also feature third-generation HyperFlex architecture for better performance and logic utilization, along with second-generation AI Tensor DSP blocks for integer, floating-point, vector, and tensor operations used in AI and signal processing.
For robotics and physical AI applications, the FPGAs can handle sensor synchronization, image and signal preprocessing, and data transfer to external CPUs or GPUs. This allows time-critical control and safety operations to run separately from AI processing workloads. Altera also offers Agilex 27 M-Series devices with in-package memory and Direct RF-Series devices with integrated data converters through UCIe chiplets, although detailed specifications for these variants are not yet available.
Altera’s Quartus Prime Pro Edition FPGA development software supports the Agilex 27 family. The latest version 26.3 also adds QuartusIQ, an AI assistant that helps with FPGA design, debugging, finding documentation, and suggesting improvements based on the current project. It can be used directly from the Quartus GUI and can also connect to external AI tools through the Model Context Protocol (MCP).
Previously, we wrote about Intel’s Agilex SoC FPGA family with quad-core Cortex-A53 processors and up to 112 Gbps transceivers, as well as the more recent Altera Agilex 9 Direct RF-Series AGRW039 SoC FPGA with integrated 64 Gsps RF converters. The new Agilex 27 family increases logic capacity and transceiver speeds while adding Arm Cortex-A720/A520 processing, newer memory interfaces, and PCIe 7.0 support.
The company says Quartus Prime design support for Agilex 27 will be available to early-access customers in mid-November 2026, with engineering samples expected in the second half of 2027. QuartusIQ is already available as part of Quartus Prime Pro Edition 26.3. Pricing has not been announced. More details are available on the product page and in the official announcement .