Design Flow Assistant
Interactive guide to help you create a development strategy.
Versal HBM Series
Hyper Integration of Fast Memory, Secure Data, and Adaptive Compute
Product Advantages
The Versal HBM series features heterogeneous integration of fast memory, secure connectivity, and adaptive compute to eliminate processing and memory bottlenecks for memory-bound, compute-intensive workloads such as machine learning, database acceleration, next-generation firewalls, and advanced network testers. It is built from the ground up to adapt to continually evolving algorithms, protocols, and data rates. Watch the overview video for more details.
With the integration of HBM2e DRAM, the Versal HBM series delivers up to 6X more bandwidth at 65% lower power per bit vs. the Versal Premium series*. Built on a production-proven Versal Premium adaptive SoC foundation, the Versal HBM series integrates an extensive set of networked, multi-terabit, power-optimized connectivity cores, and 112 Gb/s PAM4 transceivers to adapt to emerging network protocols and modules. While doubling the transceiver speed, the Versal HBM series secures every layer of the network infrastructure with built-in encryption engines. With the programmable network on chip (NoC), up to 2.2 Tb/s of on-chip connectivity alleviates routing congestion among architectural components. In addition, the Versal HBM series offers twice the logic density of the previous generation HBM solution to maximize performance for constantly evolving algorithms and protocols.
*Based on AMD internal analysis in May 2023, comparing a single Versal HBM VH1542 device with in-package HBM2E to a Versal Premium VP1502 device implementation with four LPDDR4-4266 components. Assuming sequential memory accesses with 40% read/write transactions. Power calculation generated using AMD Power Design Manager and a third-party system power calculator. Configurations may vary, yielding different results. (VER-013)
Key Features
- Integrated HBM2e
- Scalable Serial Bandwidth
- Power-Optimized, Networked IP
- Programmable Logic
- DSP Engines
- Processing System
Integrated HBM2e
Integration of HBM2e technology delivers up to 819 GB/s memory bandwidth and 32 GB capacity to minimize power, area, and latency for compute-intensive applications. Compared to commodity memory solutions (DDR5), the Versal HBM adaptive SoC delivers up to 6X more bandwidth at 65% lower power per bit vs. the Versal Premium series1 by placing stacked memory immediately adjacent to the compute fabric. Integrated HBM is globally accessible from anywhere on the device by the programmable NoC. With an integrated memory controller and enhanced hardened switch function, any memory location is accessible from any port.
1. Based on AMD internal analysis in May 2023, comparing a single Versal HBM VH1542 device with in-package HBM2E to a Versal Premium VP1502 device implementation with four LPDDR4-4266 components. Assuming sequential memory accesses with 40% read/write transactions. Power calculation generated using AMD Power Design Manager and a third-party system power calculator. Configurations may vary, yielding different results. (VER-013)
Versal HBM Series Product Brief
The Versal HBM series enables the convergence of fast memory, adaptable compute, and secure connectivity in a single platform.
Applications & Industries
Machine Learning Acceleration
Artificial intelligence and machine learning (AI/ML) evolve rapidly; complex algorithms need to process massive amounts of data, requiring enormous memory bandwidth. In the traditional compute architecture, when multiple CPU cores work simultaneously, the system stalls as data cannot move fast enough from external memory and eventually reaches the limit. In contrast, the Versal HBM series provides both massive parallel processing capability via Adaptable Engines and Intelligent Engines and enormous memory bandwidth via integrated HBM. As a result, the Versal HBM series enables accurate and faster data insights for many AI/ML processes such as Cosine Similarity and Louvain Modularity. With an extensive set of Vitis™ unified software platform's performance-optimized libraries, the Versal HBM adaptive SoC based solution can deliver higher AI/ML performance and efficiency for fast-evolving AI for data centers and the cloud.
Compute Pre-Processing and Buffering
Pre-processing data is critical to achieve the best results from fixed-function compute devices. The size of datasets for real-world ML models can easily surpass the terabytes. Hence, the target accelerator needs large-scale pre-data processing frameworks to process these datasets efficiently. With the Adaptable Engines and 819 GB/s of HBM bandwidth, the Versal HBM series removes unwanted data, transforms selected data, and augments data to create powerful predictive inputs for the target accelerator. Equipped with high-speed 112G PAM4 transceivers, the Versal HBM series maximizes throughput and system performance with low latency.
Next-Generation Firewall
Network operators desire uninterrupted, intelligent management, and robust network availability to secure data and avoid attacks on enterprise networks.
The Versal HBM series enables unmatched scalability for implementing multi-layer network security from physical and data link layers to VPNs to transport layer security levels with 10s of millions of concurrent sessions with customized policies and controls. Moreover, multiple 400G integrated High-Speed Crypto (HSC) Engines allow the system to maintain line-rate throughput and low latency without compromising the performance. With 32G HBM, next-generation firewalls can manage multiple look-up tables without accessing external memories to buffer and reorder network flows. 112G PAM4 transceivers enable support for the latest optical standards and protocols for scalability to higher throughput needed by next-generation firewalls. Adaptable Engines enable ML algorithms to modernize security architecture against emerging threats.
Application Performance Test Equipment
As data center, cloud, and AI networks have begun gearing up for 800G optical connectivity, many data center networking and cloud providers need to leverage bleeding-edge test equipment to ensure interoperability and robust network infrastructure for compute-intensive applications.
112G PAM4 transceivers in Versal HBM devices are one of the most important building blocks for data center networking and cloud providers to build networks that can adapt to emerging protocols and interoperability with optics. Dedicated channelized multirate Ethernet cores feature individually accessible HSC, MAC, PCS, and FEC blocks alongside 32G HBM, and programmable NoC to implement the most complex test logic for massive traffic buffering, efficient data movement, intelligent data-flow control, tracking, and reporting for L4-L7 test equipment.
Product Specifications
Memory Features
| VH1522 | VH1542 | VH1582 | VH1742 | VH1782 | |
|---|---|---|---|---|---|
| HBM DRAM (GB) | 8 | 16 | 32 | 16 | 32 |
| Total Block RAM (Mb) | 89 | 89 | 89 | 132 | 132 |
| UltraRAM (Mb) | 366 | 366 | 366 | 541 | 541 |
| Total PL Memory (Mb) | 509 | 509 | 509 | 752 | 752 |
DSP Engines Features
| VH1522 | VH1542 | VH1582 | VH1742 | VH1782 | |
|---|---|---|---|---|---|
| DSP Engines | 7,392 | 7,392 | 7,392 | 10,848 | 10,848 |
Programmable Logic Features
| VH1522 | VH1542 | VH1582 | VH1742 | VH1782 | |
|---|---|---|---|---|---|
| System Logic Cells (K) | 3,837 | 3,837 | 3,837 | 5,631 | 5,631 |
| LUTs | 1,753,984 | 1,753,984 | 1,753,984 | 2,574,208 | 2,574,208 |
Processing Subsystem Features
| VH1522 | VH1542 | VH1582 | VH1742 | VH1782 | |
|---|---|---|---|---|---|
| Application Processing Unit | Dual-core Arm® Cortex®-A72, 48 KB/32 KB L1 Cache w/ parity & ECC; 1 MB L2 Cache w/ ECC | ||||
| Real-Time Processing Unit | Dual-core Arm Cortex-R5F, 32 KB/32 KB L1 Cache, and 256 KB TCM w/ECC | ||||
| Memory | 256 KB On-Chip Memory w/ECC | ||||
| Connectivity | Ethernet (x2); UART (x2); CAN-FD (x2); USB 2.0 (x1); SPI (x2); I2C (x2) | ||||
Platform Features
| VH1522 | VH1542 | VH1582 | VH1742 | VH1782 | |
|---|---|---|---|---|---|
| GTYP Transceivers (32.75 Gb/s) | 681 | 681 | 681 | 681 | 681 |
| GTM Transceivers (56G (112G)) | 20 (10) | 20 (10) | 20 (10) | 60 (30) | 60 (30) |
| PCIe® w/ DMA (CPM5) | 2 x Gen5x8 | 2 x Gen5x8 | 2 x Gen5x8 | 2 x Gen5x8 | 2 x Gen5x8 |
| PCI Express (PLPCIE5) | 8 x Gen5x4 | 8 x Gen5x4 | 8 x Gen5x4 | 8 x Gen5x4 | 8 x Gen5x4 |
| 400G High-Speed Crypto Engines | 2 | 2 | 2 | 3 | 3 |
| 100G Multirate Ethernet MAC | 4 | 4 | 4 | 6 | 6 |
| 600G Ethernet MAC | 1 | 1 | 1 | 3 | 3 |
| 600G Interlaken | 0 | 0 | 0 | 1 | 1 |
1. 16 GTYP transceivers are dedicated to CPM5 for PCI Express use.
For All Developers
AMD provides a leading software development environment for designing with adaptive SoCs and FPGAs—this includes tools (compilers, simulators, etc.), IP, and solutions.
This environment can reduce development time while allowing developers to achieve high performance per watt. AMD adaptive SoCs & FPGA design tools enable all types of developers from AI scientists, application and algorithm engineers, embedded software developers, and traditional hardware developers to use AMD adaptive computing solutions.
Get Started
Jump-start your design cycle and achieve fast time-to-market with the proven hardware, software support, tools, design examples, and documentation available for the kit.
Start Developing on the Versal HBM Series VHK158 Evaluation Kit
Start evaluating Versal HBM series capabilities today with the VHK158 evaluation kit featuring the VH1582 device. Leveraging integrated HBM, this platform is ideal for developing compute-intensive, memory-bound applications. Jump-start your design cycle and achieve fast time-to-market with the proven hardware, software support, tools, design examples, and documentation available for the kit.
Resources
Stay Informed
Join the Versal notification list and be the first to receive updates.