Your download url is loading / ダウンロード URL を読み込んでいます
Your download url is loading / ダウンロード URL を読み込んでいます
Xilinx continues to be arguably the most efficient chip maker within the Valley with the introduction of the Versal HBM adaptive compute acceleration platform (ACAP), the most recent in its Versal processor portfolio.
Xilinx is called a FPGA firm, however the Versal line is a lot greater than that. Versal is the mashup of many alternative processor applied sciences into one die. In fact it has the FPGA programmable logic gates, but it surely additionally has Arm cores for functions and real-time processing, clever engines (AI blocks, DSPs), and I/O (PCIe Gen 5, CXL). The household ranges from the high-end Premium version to the Versal AI Edge processors.
The Versal HBM (excessive bandwidth reminiscence) sequence combines quick reminiscence, safe connectivity, and adaptable compute in a single platform. It’s constructed particularly to be used in information facilities, networking, and verticals like aerospace and protection.
IBM has been pushing laborious on being a aggressive menace in enterprise cloud, however is much behind the leaders like Amazon AWS, Microsoft Azure and Google Cloud. It’s newest technique to turn out to be extra related, along with shopping for RedHat for its cloud experience, is to develop a sequence of “straightforward on-ramp” Cloud Paks that it claims can considerably scale back the period of time needed for enterprises to be cloud-enabled. However is that this sufficient to alter the potential of IBM to compete in a extremely aggressive fashionable cloud surroundings?
On the coronary heart of the Versal HBM is HBM2e DRAM on the processor die, offering 820GB/s of throughput and 32GB of capability for 8X extra reminiscence bandwidth and 63% decrease energy than DDR5 implementations, which has not even been launched but.
HBM is an on-chip DRAM that runs a lot sooner than commonplace DRAM. That, mixed with the very fact it’s on the CPU die moderately than separate, bodily DRAM sticks, means a lot sooner reminiscence efficiency and little or no latency for compute-intensive, memory-bound functions.
It’s additionally an incredible networking processor, providing 5.6Tb/s of serial bandwidth with 112Gb/s PAM4 transceivers, 2.4Tb/s of scalable Ethernet bandwidth, 1.2Tb/s of line fee encryption throughput, 600Gb/s of Interlaken connectivity, and 1.5Tb/s of PCIe Gen5 bandwidth with built-in DMA, supporting each CCIX and CXL interconnects. That’s a variety of requirements, protocols, and information charges in a single package deal.
Xilinx says the Versal HBM processor can provide scalability and efficiency for 800G routers, switches, and safety home equipment in a single chip, whereas a conventional community processing unit (NPU) implementation of an 800G next-generation firewall would require a number of NPU units and DDR modules.
As a result of it’s a FPGA, the Versal HBM is an adaptive, heterogeneous compute platform, which might dynamically reconfigure itself in milliseconds to adapt to evolving algorithms and rising protocols, the corporate stated.
Xilinx has developer kits for each {hardware} and software program builders. {Hardware} builders can use Xilinx’s Vivado Design Suite to work with the Versal HBM units, whereas software program builders can use the Xilinx Vitis unified software program platform.
Although samples received’t be obtainable till the primary half of 2022, builders can begin prototyping on Versal Premium sequence units and analysis boards which can be obtainable and migrate later to the Versal HBM sequence after they change into obtainable.
