Accelerate machine learning with the Xilinx ALVEO A-U200-P64G-PQ-G
Bitcoin mining experienced a severe acceleration with the introduction of dedicated accelerator ICs, most of which achieved effiency ratings miles above normal GPUs. Xilinx now introduces a series of FPGA-based accelerators intended for data center and machine learning jobs.
From a principal point of view, both ALVEO U200 and ALVEO U250 are a set of FPGAs which connect to the workstation via a PCIe link. In addition to that, a MicroUSB and a network interface are added to the accelerator cards, allowing data center administrators effortless access to the various configuration interfaces.
In addition to that, a wide variety of SDKs are provided: they allow developers to create kernel elements optimized for the needs of the ALVEO’s internal architecture.
Memory access traditionally was a bottleneck for FPGA accelerator cards. XILINX solves this problem by providing local memory. A total of four slots take one DDR4 DIMM each, XILINX officially permits only the use of the MTA18ASF2G72PZ-2G3B1 16GB unit produced by Micron.
Local DDR4 memory reduces bus contention
Keep in mind that Xilinx offers a total of four versions of the card. A-U200-P64G-PQ-G and A-U250-A64G-PQ-G are a bit heavier, but come with active cooling. A-U200-P64G-PQ-G and A-U250-P64G-PQ-G are a bit lighter, but are passively cooled and thus require airflow from the server blade.
Cooling the passive Alveos can become a hassle
Given that the Datacenter Accelerometers are just starting to show up on the market, distributor margins are extremely wide. As of this writing, prices – as shown in the figure – range from 9000€ to 11000€ – definitely do run an oemsecrets search before committing!