Create an account on the HP Community to personalize your profile and ask a question
01-20-2018 03:36 PM
Hi, I hope someone who has more practice with those models can confirm an idea I got.
My understanding is that the z820 and Z840 can accomodate 3 PCI Express double slot if the machine has two CPUs, but only two of the cards would have x16 speed, the third only x8.
But I am not sure if the Z6 could accomodate the three double slot (even if two x16 and one x8), only two double slot x16 and one single slot x8, or only two double slot ones.
Power-wise, as far as I could see, the Z820 can have a 1125W power supply, and the same is valid for the z840, so they can accomodate even 717W of graphic cards, but the Z6 has a power supply of 1000 W, so I am afraid it could cope with a 575W graphic card configuration, but not with a 717W or a 612 W one.
I am trying to understand if I could manage with a z820, could do with a z6, or would need forcibly to go for the z840 - I don't really care about how many cores each CPU has, for that side, two 4-cores would be more than enough.
The idea is initially probably a Quadro 4000 (or a C2050) AND two dual slot Tesla M2090. The C2050 is 247 W, the two M2090 are 235 W, and the Quadro 4000 (single slot alternative to the the C2050) is 142 W - it would mean between 612W and 717W
I know these are PCI Express 2 cards, but there I am also considering a possible future scenario with PCI Express 3 configuration.
A different future scenario, with a Pascal-class P600 (single slot, 40W) or P2000 (also single slot, 75W), AND up to two Tesla P100 (dual slot, 250W), so a maximum total of 575W (if someone wonders why not a single Volta 100, huge price difference, a single volta 100 costs moch more than two tesla P100)
Thanks in advance for any help !
01-20-2018 03:50 PM
you are making some assumptions about the hp systems that will cause you possible problems
keep in ming that the base power draw of a hp workstation is based on it's configuration
IE- how many ram sticks, number of drives installed, single or dual cpu and if cpu is low or high wattage cpu
number of PCI-e add-in cards and what wattage the cards consume
also the hp power supply is a multi rail supply not a single rail, as such it is possible to overload a rail with exessive power draw while still being under the supply's total wattage spec
last, HP only supports specific quadro models in SLI non listed or consumer models will not work as hp does not have a lisence agreement with nvidia for non listed card models
while the system case/slots may be able to hold 3 video cards, you are only going to be able to use 2 at most in SLI assuming the are a hp supported model
01-21-2018 05:26 AM
Hi, thanks a lot for the fast reply, I definitively had not thought about the power rails, yes it can certainly create a mess !
I am sorry, I realised I created a misunderstanding on other stuff by not giving all the information I had already, my mistake.
I try to be more complete here.
Thank you for the link to the SLI certified, it is certainly useful, though at least as far as I know, I am talking about something a bit different.
I was thinking more in terms of a kind of Maximus sytem (quadro+tesla) but usable for different types of use (scientific, rendering and video editing, machine learning), thus requiring good performances both in dual and single precision (and seen some prices, half-precision can come in as well).
But, subject to confirmation (but that, I will have to see with Nvidia), the idea would be to link the "Pascal" cards (big investment, I may due it only end of the year and maybe stick with only one P100 and a Quadro) through NVlink and the related hardware adapter (well, doing it through the motherboard, requires a motherboard with that integrated, I am not buying a supercomputer or a power8 system :-)).
In that case, if it is not a power8 system with NVlink CPU-to-GPU, the cards communicate with each other using the NVlink, the CPU talks with them as normal PCI Ex cards - no need for SLI certification, the only requirement is from Nvidia driver side, to have the Quadro and Tesla(s) of the same generation.
An evolution of the system I have already, that is using very old stuff (nice, but not anymore supported by software I plan and will need to use 😞 ).
In the first instance (even if available, NVlink requires buying an adapter that is not free of charge), they would need talk through the bus, slower, but I can live with it.
The Quadro 4000 plus two Tesla 2090 is a configuration I am using in a very old Mac Pro, and you can be sure that I did the calculations before trying - but in that case, I could find out indications about Watts even for CPU, cards, memory (I had a configuration very diffused, and the limited number of memory slots helped), disk, and saw there was at least some margin to account for errors and ventilation (though it true I use a software to check the temperature and pump up the fan if needed, and I do not leave it on for more than some hours in the day).
Sorry for that confusion, I was not assuming that memory, CPU, other PCI Express and disks have no impact on the power consumption, I had already found some figures for that before.
The Xeon in the 820 and 840 I looked at are 135W each, the "Silver" that would be in the Z6 are 80W each - I understood that alone is enough to allow the power source wattage going down from 1125W (z820/z840) to 1000W (z6).
My real doubt for the z820 and z840 was the physical space with three double slots, though it seems I have found a partial answer.
It is for the z6 which I have more doubts, and not only physical space.
I saw HP supports officially on the z820 up to two C2075, requiring in addition a quadro of the same category (but as said, that comes from the Nvidia driver side, and it is logical, the idea is to visualise what you calculate 🙂 - basically, the same configuration I have in my old mac pro, and with the same Nvidia "Fermi" family and basically the same Watts consumption.
They also support up to two k20 with one k5000, same approach with more recent "Kepler" cards, but the k20 and k5000 are double slot, so means it can fit (though I have to check again if the wattage went down with those cards).
The z6 is where I have more doubts, power-wise, but even more, space wise - the specs I have seen list a number of slots, but it is not clear to me how many are really usable.
Also, while for the z820 and z840 the specs listed graphic cards and explicitly referred to the kind of Quadro+Tesla configuration I am planning, for the z6 they list more details of the characteristics of the possible factory-installed graphic cards, but they seem not to consider at all this type of combinations, despite the fact the z6 as far as I understand replaces the z820, and despite some indications about the target market.
01-21-2018 11:26 AM - edited 01-21-2018 11:40 AM
for power and space requirements, you can consider the compute C2075 cards and video cards to be one and the same
the C2075 cards are almost the same as a quadro video card with the only diffrence being no video outputs and custom driver software
so to be clear, yes a z820/840 with two cpu's installed will fit three of the cards, but hp does not condone doing so
you may have issues with power and the third card may not receive proper cooling
also you may not be aware, there are at least two variants of the C2075, one has a cooling shroud
and the other just has the heat sink. for what you are trying to do the cooling shroud variant is the one you must use as the cards will shut down if they get to hot (i think it's 70c)
last it's my understanding that even though the C2075 cards have 2 power inputs (6/8 pin) HP says they they will run fine with only one 6 pin power jack connected so the 1150 watt supply has two such connectors and hp sells a molex to 6 pin kit (which uses thicker gauge wires than nost adapters) for the offical config of video card + two C2075
i don't think the supply will tolerate a third card on the power rail. to overcome this you might consider a aux power supply that fits in the 5.25 drive enclosure.