so i've purchased an Nvidia Quadro M4000 for 800 bucks.
it's essentially a premium gtx 970 with 8gb vram, more bandwidth, that nvidia itself ensures receives the cream of the crop from the chips that are produced.
Aaaaand now to flash a custom bios to it to bring the clock speeds from 750mhz to 1200mhz.
i'll be using Maxwell 2 bios tweaker, gpu-z, NVflash.
anyone interested in this sort of thing?
the stuff in pic related is for the build its going into, dual xeons, 64gb ram 4 channel 16x4gb, ssd, water cooled, all that. gotta get the mahogany and glass to make the case. right now its inside an amd a6-3600 computer...
anyone interested?
>>53327788
Interested in what exactly, OP?
Fancy specs are one thing, but unless you actually plan on using it for something cool I'm not really interested.
No, gaming is not "something cool".
>>53327788
On /g/? No, we're only interested in shillposting and shitty destitute-tier computers.
>>53327788
>64gb ram 4 channel 16x4gb
...but there's 16 slots there in the pic... plenty of room for more
With the exception of watercooling and wooden case you pretty much described what i'd buy for a home linux workstation if I had the money for it.
From 8 year old laptop to a dual xeon upgrade would feel like a blast
>Aaaaand now to flash a custom bios to it to bring the clock speeds from 750mhz to 1200mhz.
Is there really a point? What kind of performance drop we're talking about with this (i assume factory) underclock? Was it done for stability reasons?
>>53327815
well, i do like to play a game now and then. three or four times a year i'll get a hankering and go play a game for a bit. so the overclock of a otherwise locked workstation card is useful for that.
but mostly i got it for applications that won't even allow themselves to start if they don't detect certain gpu's.
i'm specifically looking for a cracked or sample version of, ahem, "Presagis offers a number of sensor simulation products that are developed to deliver the realism and fidelity required to simulate real-world visual sensor displays. These extensible modules plug directly into the Vega Prime environment, and use the same synthetic environments as Vega Prime for correlated out-the-window, NVG, infrared, and radar views.
Using this toolkit approach, users can easily control and extend the parameters, as well as modify the results.
The available functionality includes the ability to create realistic NVG and infrared effects such as blooming, persistence, random temporal noise, and jitter. Vega Prime Radar offers multiple modes including DBS, RBGM, SAR, and ISAR."
for some robots stuff i'm developing. i'm interested in networked GPU accelerated drone group coordination. for reasons.
>>53327848
16 slots, each has a custom stepped 4gb ecc ram dimm. made for that motherboard and processors.
>>53327857
the watercooling will ensure that the xeons stay in their turboboost mode pretty much indefinitely, these are xeon E5-2670 v2's, which will turbo out to 3.5ghz for all 16 physical cores.
the mohagany is just because i love woodworking and jointing with hand tools and i want it to be nice and pretty and custom on my desk. it'll be kinda like a wall-case, mounted. and there's really no cases for sale that could do what i wanted with a motherboard that size, dual cpus, and watercooling.
the clocks, yes, quadros are underclocked so that there is essentially 0% chance of a failure, which could cost hundreds of thousands of dollars for a company...
>>53327943
>i'm interested in networked GPU
Well, coincidentally I'm currently working on a benchmark program for GPUDirect RDMA across a PCIe interconnect.
The bandwidth I'm getting is completely annihilating the bandwidth one would get with a single-rail InfiniBand link.
https://devblogs.nvidia.com/parallelforall/benchmarking-gpudirect-rdma-on-modern-server-platforms/
And latency-wise... Well I'm using PCIe all the way.
>>53327788
where do you live? asking for a black friend
>>53327997
>https://devblogs.nvidia.com/parallelforall/benchmarking-gpudirect-rdma-on-modern-server-platforms/
neat shit!
>>53328034
u realize i'm trying to build a robot army right
>>53327788
Seems cool but what are you going to use it for? If you are going to use pci passthrough for virtual machine gaymen on windows then I'm a bit jelly. I'd imagine it's going to sound like a hurricane though going by normal server hardware.
>>53328058
the 100% fan is loud, but not, like, kill yourself loud, like, sooth yourself to sleep loud on the card.
the noctua fans are relatively quiet for what they do.
>>Seems cool but what are you going to use it for?
robot army.
>>53328043
>neat shit!
Well that's the speeds I'm beating.
Will make a blogpost about it once I've refactored my program and replicated the tests on the same configurations.
We get close to 5.9 GB/s for host to remote GPU and GPU to remote host transfer, and 6.6 GB/s in the PCIe p2p example.
>>53328043
but where do you live? how good is your security? do you have home insurance?
>>53328088
you think you might be able to connect it to a series of say, security cameras, and use it for facial/weapon analysis?
lol, that's a Linus Tech Tips tier build.
>>53327788
I think you mean 7GB of ram if it's a 970 core.
>>53328131
ha, there is a completely different set of memory chips and controllers in the quadro version, it may even be possible to activate soft ECC with a bios flash. wouldn't care to though. useless for me. apparently it uses hynix. 7gb would be good anyways. its 8 tho
>>53328125
>you think you might be able to connect it to a series of say, security cameras, and use it for facial/weapon analysis?
Yes
We already use the interconnect in a cluster for doing real-time polyp detection in gastronomy videos.
And here's the really really really cool part: In addition just using PCIe interconnects, we have actually implemented a solution that allows a host to "borrow" devices from other hosts. With some nifty remapping of memory addresses using the IOMMU (as in Intel VT-d) and the NTB-functionality of the interconnect, we make remote devices appear as local. I can't go into details here as there is a potential for patenting, but it will be ready for market in a matter of weeks.
If you have a facial/weapon tracking software that runs on multiple GPUs, you can use the PCIe interconnects with NTB functionality and this "device borrowing/lending" technique and it will Just Work[tm] on all the GPUs in the cluster.
i'm going to have to go to my grampas house and help him take some stuff to the dump. be back shortly.
>>53328224
I obviously meant gastroscopy. Sorry, English is not my first language and I suck at medical terminology.
>>53328131
Nvidia would face some massive legal shit if people started failing simulations when they ram out of RAM. It's one thing if it's gamers who don't do anything of consequence, and a whole other deal when you're driving design software that might be responsible for the safety features of a car, or maybe even an aircraft. It's only like a 970 in the base specs.
>>53328284
*run out of RAM
>>53328284
Well, NVIDIAs intention is to use the 970 merely for prototyping and then for you to buy Quadros when you're ready for production.
>>53328205
>>53328284
The memory is definitely there, the trouble with the 970 core was that the last eighth of the memory cannot be accessed at the same bandwidth as the main lot which in turn reduced the total bandwidth as well.
So it was marketed as 224GB/s but in the end only performed around ~200GB/s and a separate ~20GB/s, but never combined. I am actually curious if this affects the quadro part as well.
>>53327943
>these are xeon E5-2670 v2's, which will turbo out to 3.5ghz for all 16 physical cores.
Don't they just have 10 physical cores per unit? Anyway are you paying for all this out of your own pocket? Is it new or used?
>>53328420
actually, they're v1's, now that i look. i bought the board with them included, the board supports v2bo but these are v1. max turbo is 3.3ghz and yes, 8 physical cores each. hmm. that's strange. i'm not pissed really because i got a great price, and i'm going to upgrade to 2687w v2 engineering samples at the first opportunity anyways, these processors were never meant to be a long term solution. i'll probably put them into a pair of node computers using Dante protocol for audio.
>>53327788
>it's essentially a premium gtx 970
Stop right there. The GPU architecture used in Quadros versus GeForce cards have changed since Kepler. They used to be pretty closely related, but ever since the Titan Black, Nvidia has been pushing consumer GeForce cards away from raw compute power to Jew the ever-loving fuck out of their Quadro cards.
The Titan X is an example of this: the original Titan and Titan Black had really good raw compute and FP32 capabilities, which made its $1000 price tag too appealing for professional consumers. With the subsequent Maxwell generation, Nvidia made the SMM on their GM200 GPU less FP32 heavy for better gaming performance, but weaker compute performance so that they can shill their Quadro and Tesla cards. So people who bought the Titan X basically paid extra for 12GB of RAM and not much else.
This goes for the rest of the Maxwell line-up, but also for the current Quadro line-up. They are much weaker for gaming than their price-equivalent consumer versions, but far better at FP32 and compute-heavy applications because their core architecture is different.
>>53328712
overclock ur underclocked quadro
lern driver settings w/nvidia inspector
???
profit
>>53328761
The core architecture between Quadros and GeForce Maxwells are different, so you're not getting the best bang-for-your buck in terms of gaming performance, even if you flash a GeForce Maxwell BIOS and overclock the hell out of it.
You're paying more for capabilities most consumers won't ever use, and effectively cut off the value of the card by disabling over half of the SMMs per core.
>>53328825
well, that's what i paid for, capabilities that most consumers will never use.
that being said, lets do this! if i ever really wanted to play games hard for some reason, i'd just get some amd card or someshit for a couple hundred bucks.
>>53328871
>well, that's what i paid for, capabilities that most consumers will never use.
>now I'm going to disable about 60% of the transistors on my GPU so it can act like a cheaper $300 consumer-grade GPU!
rebooting now, apparently the flash was successful.
i saved the ROM with gpu-z, opened with maxwell2 bios tweaker, decreased some of the temp maxes and increased the gpu clock, saved as mod.ROM, typed nvflash64 -6 mod.rom... and it said flash successful, reboot needed.
reenabled the adapter and rebooted just now. lets see what FPS i get...
what do you know, it fucking works.
i'm not going to go any farther with this in this ol HP beater i got for 20 bucks, because as efficient as this card is i do not trust the power supply.
but, /g/, now you know, that you can create custom bios for your maxwell 2 quadro cards to hype them up.
i just ramped mine up to 1139mhz, from its stock 750. people have gotten 970's up to 2500mhz. i have to learn more about the voltage and shit before i go farther.
>>53329215
>people have gotten 970's up to 2500mhz. i have to learn more about the voltage and shit before i go farther.
You literally have no idea what you're getting yourself into.
Maxwell does not respond well to voltage increases when overclocking. Quadros are more akin to Kepler than they are Maxwell. And 2500MHz was most likely LN2 overclocking.
>>53329294
>>You literally have no idea what you're getting yourself into
so what. nobody ever really does before they get into it.
>>i have to learn more
>>HURR NO IDEA
how about you offer some of your infinite wisdom, oh wise one?
>>53329376
Give up on the 2500MHz ambition, that will NEVER happen on air. You should consider yourself very lucky to have even reached 1100MHz with a stock cooler.
Quadro's GPU design is different from regular Maxwells. I'm not sure if they respond to voltage increases like Kepler did, but I wouldn't be surprised if the amount of voltage needed exceeds that of an equivalent Maxwell. In fact, this Quadro might be just like Kepler in regards to overclocking due to its architectural similarity.
Power delivery on Quadros is very much gimped compared to consumer GeForce cards since they were never designed to be overclocked, so you're going to have a very narrow window to toy around with voltages and frequencies. You might not be able to reach 1200MHz because you can't feed enough power to the GPU for stable performance.
Plus, that blower cooler won't be strong enough to keep the MOSFETs and GPU cool for long-term overclocking. The only way around this would be to find a custom aftermarket air cooler and hope for the best, or to have a 3000RPM Delta fan blowing straight into the front part of the card. I don't think water-blocks are made for Quadro cards anymore.
>>53329479
not aiming for 2500mhz.
here's what i'm seeing right now, its at 962mV at 1139mhz, and its sitting at 50C. that seems to indicate headroom? anyways, looking to learn more.
>>53329660
>>53329479
whoops
>>53329675
>>53329660
From the looks of it, you're more likely to hit a power limit rather than the thermal limit of the GPU. Quadros were never meant to be overclocked, so their power delivery is intentionally anemic.
>>53329715
i see. so, where can i monitor its total consumption, and what am i looking for in specs or technical sheets on this card that would tell me its operating ranges? what components on the card itself could i look up and learn about to see what the hardware itself is capable of? and what happens when i hit the power limit? it'll have a 1000 watt evga supernova g2 powering it.
>>53329782
>where can i monitor its total consumption
There's no real program that can do that, and you want to look at the voltage for that, not wattage. You can do back-of-the-hand math to determine the wattage based on the voltage used.
>what am i looking for in specs or technical sheets on this card that would tell me its operating ranges
You're up shit's creek without a paddle here. Maybe another Nvidia enthusiast forum might know.
>what components on the card itself could i look up and learn about to see what the hardware itself is capable of
Try looking for pictures of the board without the cooler and shroud attached. You want to look for the power delivery the card has, mostly by counting the number of chokes and/or MOSFETs it has. The lower, the worse it is. Also, look at how many PCI 6 or 8-pin power connectors it has. That's the absolute limit to how much power the card can draw from the system.
>what happens when i hit the power limit
You either lose stability when you try to over-reach or the card will throttle down noticeably on the voltage/GPU usage chart (look for a plateau effect).
>>53329867
ok. so 1.2 is consider a upper mV limit for 970s. i'm drawing .962 at 1139mhz/50c. its TDP is 120w. i can look for 'chokes' and mosfets on the board to identify and assess their capacities. it has a single 6 pin.
stability loss looks like what? i can look that up on google. as for throttling, i can make sure that any custom BIOS i flash has a fall-back when it reaches mV-maximums.
this is a alright place to start.
>>53329968
>so 1.2 is consider a upper mV limit for 970s. i'm drawing .962 at 1139mhz/50c
That's still well over the recommended amount for Quadros. Again, they're not entirely the same GPU.
>its TDP is 120w
Irrelevant.
>it has a single 6 pin.
The most power it can draw is 150W. That's the absolute limit. The thing about the MOSFETs and chokes is only a general indication of how good the card's power delivery is - it is NOT a direct indicator of how far you can take the GPU. You're limited by its 150W power limit.
>stability loss looks like what
Artifacting or other graphical errors, but also outright crashes. You might even end up with uncorrectable crash that would kill the card because you're not using the intended BIOS. I have seen cards turn into $500 paperweight after their BIOS was flashed and pushed to moderate limits. And reflashing BIOS does not work for these crashes.
