AMD Virtex UltraScale+ HBM Devices
Learn about the memory bandwidth and capabilities of AMD’s 16 nm Virtex UltraScale+ FPGAs with CCIX technology and High Bandwidth Memory (HBM) in this video.
Benefits of the AMD Virtex UltraScale+ HBM Product
With up to 500 Mb of total on-chip integrated memory and up to 16 GB of high-bandwidth memory (HBM) Gen2 integrated in-package for 460 GB/s of memory bandwidth, AMD Virtex UltraScale+ HBM FPGAs offer the largest on-chip memory density.
For workloads that process large datasets from AI inference, video transcoding, next-generation firewalls, search applications, and data warehouses, an innovative embedded HBM controller and ground-breaking integration allow for maximum bandwidth, effective routing and logic utilisation, and optimised power efficiency.
Important attributes and advantages
Large-Bandwidth Storage
20 times the bandwidth of a DDR4 DIMM is provided by 460 GB/s of HBM bandwidth. For the maximum useable HBM bandwidth, extended AXI ports and an integrated port switch reduce design size, complexity, and time to market while enabling access from any port to any address.
SerDes Multi-Terabit Bandwidth
High throughput systems can have a bandwidth of up to 3.1 Tb/s SerDes. Support for the newest optics and compatibility with current infrastructure are made possible by support for 32.75G NRZ and 58 Gb/s PAM4 transceivers.
Integrated Fast Internet Access
Faster design cycles and reduced power consumption are made possible by integrated hard IP for 100G Ethernet with RS-FEC, 150G Interlaken, and PCIe Gen4.
Low Bit/Power
Low OpEx for power-sensitive use scenarios is made possible by 4-6X lower power per bit (~7 pJ/bit) when compared to a discrete memory solution.
Gen4x8 PCIe with CCIX
For a comprehensive end-to-end solution for multiple 100G ports, cache coherent computation employing CCIX ports is appropriate.
Reduction of Footprints
These FPGAs allow for the construction of compact systems without sacrificing speed by removing the need for I/O to external packaged memories and offering large amounts of on-chip memory.
Applications
Investigate New Opportunities with Virtex UltraScale+ HBM FPGAs
Network and Security Acceleration
Many network and security applications greatly benefit from the marriage of HBM with high-end programmable circuitry. The Virtex UltraScale+ HBM FPGA-based network accelerator card can achieve the same solution in one device with the same memory, less power, and faster performance, as opposed to requiring several DIMMs on the board. Compared to commercially available TCAMs, a solution utilizing a Virtex UltraScale+ HBM FPGA with HBM stacks offers up to a 5X higher look-up rate due to HBM bandwidth and 80X more search entries. By simplifying the PCB and lowering the complexity of the memory subsystem, an HBM implementation offers a simpler and lower risk design flow in addition to these intrinsic advantages for the final product.
Data Analytics and Database Acceleration
Database infrastructure is under stress from the proliferation of data in contemporary applications, since power limits stay constant while bandwidth demands rise quickly. Performance and capacity for a wide range of database and data analytics workloads, including as lookup, query, compaction, and compression, are provided by up to 28 TOPs (INT8) and 460 GB/s of memory bandwidth. Additionally, AMD offers a wide range of development tools, including the Vitis unified software platform, the Vivado Design Suite, and pre-optimized partner solutions, allowing users to select the best option for their needs.
Adaptable Artificial Intelligence
The configurable flexibility of Virtex UltraScale+ HBM FPGAs makes them ideal for the constantly changing machine learning (ML) and artificial intelligence (AI) architectures. This FPGA uses chip-on-wafer-on-substrate (CoWoS), which AMD invented, to access HBM memory with hundreds of signals. AI inference applications may process big datasets as quickly as possible while using less power with the integrated HBM. Virtex UltraScale+ HBM FPGAs have more than enough memory to store and retrieve a whole model without requiring external memory sources with their 460 Gb/s memory bandwidth.
Video and Image Processing
Nowadays, the majority of internet traffic is from video streaming. A steady stream of excellent films and photos is necessary for social media, gaming, marketing, and broadcasting. Both the user experience and the data centre infrastructure must be managed by service providers. FPGA-based video acceleration helps users provide reliable, latency-critical services at a reasonable cost while also streamlining infrastructure. For compute-intensive video and image processing applications, Virtex UltraScale+ HBM FPGAs provide a range of form factors and serial interfaces, including 58G PAM4.
Product Table
XCVU31P | XCVU33P | XCVU35P | XCVU37P | XCVU45P | XCVU47P | XCVU57P | |
---|---|---|---|---|---|---|---|
System Logic Cells (K) | 962 | 962 | 1,907 | 2,852 | 1,907 | 2,852 | 2,852 |
DSP Slices | 2,880 | 2,880 | 5,952 | 9,024 | 5,952 | 9,024 | 9,024 |
HBM DRAM (GB) | 4 | 8 | 8 | 8 | 16 | 16 | 16 |
HBM AXI Interfaces | 32 | 32 | 32 | 32 | 32 | 32 | 32 |
PCIe Gen3x16/ Gen4x8/CCIX | 4 | 4 | 4 | 4 | 4 | 4 | 4 |
GTY/GTM Transceivers (32.75/58 Gb/s) | 32/0 | 32/0 | 64/0 | 96/0 | 64/0 | 96/0 | 32/32 |
I/O | 208 | 208 | 416 | 624 | 416 | 624 | 624 |
Get Started
Make use of the tried-and-true hardware, software support, tools, design samples, and documentation included in the kit to speed up your design process and reach market readiness.
Virtex UltraScale+ HBM FPGA VCU128 Evaluation Kit
The AMD Virtex UltraScale+ VU37P HBM FPGA, which uses stacked silicon interconnect to add an HBM die next to the FPGA die on the package substrate, is incorporated into the VCU128 board.
Assistance & Materials
Virtex UltraScale+ HBM FPGA Product Brief
Up to 16 GB of high-bandwidth memory (HBM Gen2) can be integrated into Virtex UltraScale+ HBM FPGAs at 460 GB/s bandwidth and a very low power consumption of about 7 pJ/bit. An integrated HBM controller and switch minimises R&D time and reduces logic space by 250 K LUTs.