Is there any successful example of USRP RIO with PCIe adapter?

Question

Hi All,
       Can I ask who has ever had successful experience of achieving high IQ rate using USRP RIO with PCIe adapter? If so, what PC were you using?
       I am working on the USRP RIO. I tried to run the LabVIEW code for the USRP-2920 on the USRP 2950R but could not achieve a high IQ rate. The PC I am using has a mini-itx motherboard with H97 chipset, 16GB 1600MHz DDR3 memory, 256GB SSD, i7-4790K quad-core 4.0GHz CPU. Even if I connect the USRP-2920 to the PC, I could not achieve 20MS/s. When I was using a Macbook Pro with i7-3840QM, I could achieve 25MS/s IQ rate. But there is no way I installed a PCIe adapter to the laptop. 
       Thank you very much.

Answer 1

I have a benchmarking utility that I will attach here that can run through a number of IQ Rates and the number of samples in each Write call.  Here are the results when I ran with an USRP-2940R over a x4 MXIe (PCIe) link.  I configured to run a 2-channel continuous generation for 10 s for each test.  
Device connected over PCIe
Conducting benchmarks for continuous writes (CDB, 16-bit).
IQ Rate        363 726 1000 3000 6000 10000 100000
1000000      0      0       0       0        0         0           0
2500000      0      0       0       0        0         0           0
5000000      0      0       0       0        0         0           0
10000000    0      0       0       0        0         0           0
20000000    39    0      0       0        0          0           0
40000000    18    18    0       0        0          0           0
60000000    9      36    15     2        0          0           0
80000000    22    17    14     0        2          0           0
90000000    15    22    15     9        9          9           9
95000000    21    21     22    9        9          9           2
97500000    16    14     31    9        9          9           9
100000000  15    34     25    9        9          9          10
The formatting may be hard to figure out, but the small numbers are the number of underflows.  So 0 is what you want.  As you can see, I can sustain rates up to 80 MS/s for at least 10 s IF my write size is big enough.  That is, if I write at least 10,000 samples with each call to niUSRP Write, I don't see underflows.  That's two channels, so we're talking 640 MB/s over the bus.
Here are some tips to increase your Tx streaming throughput:
0) Don't use a continuous generation at all.  For many applications you don't need to stream continuously- a finite transmission (and you can loop a finite transmission) will do and you will basically not underflow in that case if you provide all the data up front.
1) As you can see from the chart, the bigger the data buffer you provide in each Write call, the faster you can stream.
2) Write sizes in multiples of the maximum packet size seem to work well.  That number is 2044 for the USRP-294x/5x series (although that may change in future driver releases).  Try sending bursts of 10220 samples.
3) If your application allows it, set a Start Trigger Time a little in the future.  Then start writing data before the device starts transmitting.  For example, set the Start Trigger Time to the (current device time + 1 second).  Then start writing data and you will have a second to pre-fill the on-device buffers.  This will substantially reduce the number of underflows.
4) Be sure to do your data processing out of your write loop, to keep the write loop filling the pipleline as quickly as possible.