Create an account on the HP Community to personalize your profile and ask a question
01-28-2019 12:47 AM
Plan to buy a HP workstatio Z4 G4 with 1000W chassis. I have NVIDIA Quadro K6000 and NVIDIA Tesla K40 with me so I do not want to buy any video cards. The option is no integrated video card.
Can somebody from technical side tell me if Z4 G4 can support NVIDIA Quadro K6000 and NVIDIA Tesla K40 ?
Thank you very much!
01-28-2019 06:44 AM - edited 01-28-2019 06:56 AM
Using a Quadro K6000 + Tesla K40 is not practical in a Z4 G4. The K6000 TDP = 225W and the Tesla K40 = 235W. These are both Kepler which is important for a Quadro + Tesla configuration and the 1000W PSU should support the power requirements, but the K40 was designed for a server environment with extremely high air flow and dedicated expansion slot air cooling. If there is a waterblock for the K40 or one could be adapted and if that could be tied into an open loop water cooling system- that would have to be an external unit. Pending for use with a z620, I have an Alphacool Eiswand external unit which is uprated and that was $314 for the raidator /reservoir / pump/fans + $167 for the CPU waterblock, connector and tubing.
How are you using the K6000 and K40 currently?
Considering the high degree of impractility and probabl high the cost of an external cooler- at least $400 and the number of generations that Kepler is obsolete, consider three alternateves: one, a pair of Quadro K6000's- only one needed to purchase- so as to avoid the cooling problem, two, a pair of Quadro P4000 8GB, and three, the recently released NVIDIA Quadro RTX 4000 8GB.
The K6000 has 12GB memory, 2,880 CUDA cores with single precision of 5.2TFLOPS and double preicision of 1.7TFLOPS. The Passmark average 3D mark is 7598. That is not a large improvement on the Quadro P2000 5GB which is about $430 new and has a 3D average of 7435. Of course, the K6000 has more than 2X the memory of teh P2000, which is significant .
The K40 has 12GB of memory, 2,880 CUDA cores with single precision of 4.29 TFLOPS and double preicision of 1.43 TFLOPS
In my view, the K6000 is currently valued at about $600 used and the K40 at $200-250, which reflects the important limitation in installation due to the active/passive cooling.
However, unless there is a current cooling colution for the K40, as the K6000 and K40 are worth ~$800, consider alternatives.
The Quadro P4000 has 8GB memory with a 243GB/s bandwidth, is only 105W , 1792 CUDA cores, supports 5K resolution.
the Passmark average 3D = 10434. < Compare to 7598 for the K6000. P4000's used are currently costing about $550.
The Quadro RTX 4000 - which is expected to cost about $900 has 12GB FP32 performance of 7.1 TFLOPS, 2.304 CUDA cores, and TDP of 160W. There are no benchmarks on the RTX 4000 as of yet, but it anticipated to be very fast- at least GTX 1070 Ti level and of couse has the ray tracing and tensor core technology. One of those may provide sufficient perfroamne and I think today a single GPU- if possible, is the better solution.
At this point, the best practicaillity and cost/performnace solution appears to be using two Quadro P4000's. that will have 16GB of Memory, 3,582 CUDA cores for a cost less than the K6000 + K40 + cooler toal installation value. that should have a very good compute performance due to the higher clock speed and large memory bandwidth. It would be important to understand the P4000 in SLI to the Z4, so that full benefit could be relaized in the intended applications. Important : The P4000 may be connected in SLI- and HP supplies an SLI- bridge, but I don't know if the Z4 G4 supports thw P4000 in SLI. The way in which any two GP's are used should be thoroughly researched in the context of the priority applications.
What kind of applications will the Z4 be running?
HP z620_2 (2017) (R7) > Xeon E5-1680 v2 (8-core@ 4.3GHz) / z420 Liquid Cooling / 64GB DDR3-1866 ECC Reg / Quadro P2000 5GB _ GTX 1070 Ti 8GB / HP Z Turbo Drive M.2 256GB AHCI + Samsung 970 EVO M.2 NVMe 500GB + HGST 7K6000 4TB > HP OEM Windows 7 Prof.’l 64-bit
01-28-2019 08:06 PM
Thanks a lot for your time and patience to write this long reply with professional inputs and comments.
I am planning to use my workstation for some CAD and EM simulation work. I have those GPU cards already so I do not want to spend extra hundreds of dollars for a new card.
So in this case, if I use only NVIDIA Quadro K6000, I should be ok to plug it into Z4 and make it work, right?
01-29-2019 09:53 AM
Yes, the Quadro K6000 is a very capable card. My local linear accelerator where I did a small project four years ago was using that GPU to design the amazingly precise supercooled acceleratoe modules. The 12GB of memory is an asset in simulations and of course, they run simulations of the experiments constantly, albeir those were run then on 11 parallel dual Xeon systems each with 3 GPU's compute modules, that I think then were Tesla K20X .
My point in the previous reply was to suggest that if the K40 can not be used because of the cooling problem, then consolidating the value of the two cards- say, $800, might be applied to a single more modern card, or sort of to 1.5 cards. The idea is to optimize the function of the Z4 which has a far faster memory and disk system than the systems comtemporary to the K6000 of 2013. Computers do have some conformity to the idea that " a chain is only as strong as it's weakest link" and I think that for a bew Z-series, a 6-year old K6000- as great as it was new- is not extracting full use of a new high performance system.
The Passmark 3D mark is reflective of computational power per unit time (calculation of polygon positions and attricbutes) , so even though the K6000 has more memory and CUDA cores, the clock speed is slower and the interchange with the Z4 memory will have something of a bottleneck. In the future too, it might be possible to use a pair of P4000's in SLI for 16GB of memory.
This kind of change can be surprisingly economical, but less so as the obsolotet technology depreciates. For example, a K6000 which has a value of up to $600+ has a Passmark 3D of about 7500-8000 in z420's and z620's, whereas a used Quadro P4000 8GB - used $500-550 can have a 3D in a Z4 up to 13085. It is telling in my view that there are no Z4, Z6, or Z8's using the K6000. The Tesla K40 comprises about half the value of a potential 2nd P4000 now and by August a P4000 may cost $350 or so. I just sold last week, a Quadro P2000 before it depreciates further in anticipation of whatever the "Quadro RTX 2000" turns out to be. If NVIDIA holds the usual series perormance progression, the RTX 2000 should be as fast of faster than a P4000. The recent Quadro RTX 4000 is said to be almost as fast as a GTX 1080, but there are no published benchmarks as of yet..
Of course, if you can't interrupt work and fuss with buying and sellling, that is completely understandable.
It may worth considering contacting the maker of the EM software and asking fro their compoennt recommendations. I notice that for example, Keysight EMPro has phone and email contacts for questions.