01-15-2019 11:45 PM
If possible please do reply inline.
Hello XILINX TechSupport , @kvasantr ,
Please note I've been able to successfully bring up the VCU1525 on a DELL PowerEdge Server R730 and run through successfully all steps in the U200 Getting started guide.
As really I am a newbie (into hardware picking up ) into this my next steps are
Evaluate XDMA versus QDMA versus use PCIE block directly for the VCU1525 based on below considerations.
1. Performace and stress tests with regards to throughput and latency.
2. Different in functionality and under which situations is each better.
3. Resource and power usage.
4. Timing analysis and analysis of various reports.
Its an interesting exercise which we'd love to do and i believe would be of good strategic importance to XILINX as well.
With regards to the same I request XILINX for below:
1. Guide us towards or send us some easy to read, succinct, end to end documentation to be able to implement each IP (XDMA,QDMA and PCIE Integrated block) and run above tests. I have the product IP documentation but to be honest its so hard to interpret it even after going through PCIExpress specification.
Also what scripts can I use or do I have to create my own test harnest (C++ on host pushing data for H2C transfer , Remote machine via network interface pushing data for C2H transfer ) ?
2. Possibly let us know XDMA , QDMA are suited for what sort of applications.
3. Guide us on what to extract from various implementation reports.
I understand resources are scarce but its an important exercise , we always do such exercises and many vendors(cannot take names) who work with us have immensely benefited. Please assign a capable resource with whom I can work to mature your QDMA IP.
01-16-2019 10:16 AM
At this juncture, the VCU1525 and U200 Alveo Data Center Accelerator Card are both designed to be used in the SDAccel environment with the Xilinx provide Shells. Currently, our shells incorporate specifically an XDMA PCIe core, with some enhancements to better interoperate in that environment. While there are other shells in work that may incorporate other version of our PCIe IP, there are none available at this time.
If your goal is specifically to do a direct IP comparison of the Ultrascale+ Integrated Block for PCI Express IP, the XDMA for PCI Express, and the QDMA for PCI Express, the VCU118 Development card would likely be a better choice at this junction. I will comment that the full infrastructure you are looking for with each of these cores may have varied availability.
For example, the XDMA and QDMA cores both come with Evaluation drivers and performance tests, that can be run specifically with their example designs. Each Product Guide (PG195 and PG302 respectively) provide guidance on how to create those designs. The UltraScale+ Integrated block (PG213) also has a very basic example design, which can accept single DWORD requests from the upstream device, and respond, but for a true head-to-head comparison would require some significant enhancement on the part of the user. Additionally, it works with the standard PCIe drivers, but we don't provide a test environment specifically for the general use case. There is a Technical Reference Design that does do performance testing targeted at specific development boards (like the VCU118).
For resources in differentiating the cores, each guide opens with a Chapter on the various functionalities and some common use cases. https://www.xilinx.com/products/technology/pci-express.html provides links to all the collateral for our PCI Express cores, including Videos on implementing the cores in performance test configurations where available.
01-16-2019 08:38 PM
OK so the reality check is that I have the VCU1525 -- so unfortunate was duped by the XILINX India distributor into VCU1525 by with helding information about discontinuation or being ignorant about it. Want to escalate the same figuring out how I can.
I cannot purchase the VCU118 unless you loan me one and I can happily do the IP comparison for free for XILINX.
Yes the goal is to do a IP comparison. I believe VCU1525 works with Vivado 2018.2. Shell concept has just entered since 2018.3 ?? Correct. Why do you then say its not possible ?
Anyways XILINX may have done some tests as well, you can share those numbers. Some comparison documentation might already exist please share the same.
If nothing exists well work with me(us at iRage we are a boutique prop trading firm) and we can help you in this very important activity. Its pretty justified for this extended partnership as its WIN-WIN for both.
01-17-2019 08:38 AM
VCU1525 was delivered in 2017.4 with a Shell - as part of the SDAccel suite - up to the 2018.2 release. It was only rolled into the XRT (Xilinx Runtime) environment as the load mechanism in 2018.3 (XRT was produced as part of 2018.2.2). That is the supported way to load the card and utilize it.
02-18-2019 10:04 PM
We have VCU1525. Please kindly give your response inline to the points mentioned if possible.
We have zeroed in on either of below:
1. PCIE Integrated block: Believe its lower latency than XDMA. But more work on host side to get DMA up and running.
2. QDMA: Theoritically seems like although it has overhead of AXI it may still be faster than PCIE Integrated block.
Please guide us
1. More on a theoritical comparison of either based on your understanding.
2. Some more documents , presentations etc related to QDMA will help.
3. Plus some example , starter code.
02-19-2019 08:23 AM
Since you are looking at doing a custom design on VCU1525 vs. using a Shell, I would recommend you post your PCIe comparison questions to the PCIe specific forum.
02-19-2019 07:38 PM