Nvidia/AMD Cuda/OpenCL on Boinc projects - which card to buy?

Message boards : GPUs : Nvidia/AMD Cuda/OpenCL on Boinc projects - which card to buy?
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · 6 · Next

AuthorMessage
Ian&Steve C.

Send message
Joined: 24 Dec 19
Posts: 156
United States
Message 102154 - Posted: 14 Dec 2020, 17:53:52 UTC - in response to Message 102131.  
Last modified: 14 Dec 2020, 18:27:50 UTC

Do you have a rough estimate of how much faster Cuda tends to be for Boinc projects? So I can add a factoring into my spreadsheet for choosing cards.
No, this is something you will need to check project by project yourself. This will vary.

Although some projects only have CUDA.
The only active project that seems to be CUDA only, is GPUGRID. most are OpenCL.

I thought Folding@Home was Cuda only?
Of course not, why do you think there are AMD cards contributing if that were the case?

It depends on the project, if I only did SP projects, I'd say there isn't much between the two cards (unless as you say Cuda coding is better). But I do DP projects. The Nvidias are half the speed of AMDs on those.
Again, it depends on the card. at the absolute max performance, Nvidia has AMD beat on DP also. look at the DP performance of a TitanV, its twice that of the best AMD card, the Radeon VII, and even on the openCL milkyway app it still outperforms the AMD Radeon VII. It's not simply that "AMD is better" it's that nvidia moves their higher performing DP cards to a higher price tier. their Geforce cards are marketed for gaming, not DP, so it makes no sense for them from a business standpoint to offer high DP performance to gamers who wont use it. guys who run BOINC and DC projects are such a small segment of the market that they really have no incentive to cater to us. AMD is going in the same direction and reducing the DP performance of their consumer gamer cards with each new generation. Even AMDs biggest and baddest consumer card RX6900XT runs DP at a mere 1:16 ratio only offering 1.44 TFlop, where their older Radeon VII at 1:4 pumps out 3.4 TFlop. But if you want ABSOLUTE performance, nvidia is still winning DP, and at better power efficiency. for someone running a BOINC DP project, and presumably running the system 24/7 for several years, the power efficiency gains eventually offset the initial product cost (titan V vs Radeon VII at current used-market pricing).

but again milkyway is the only project running DP full time. any other projects using it only do so for a small portion of the crunch time, and the SP performance is a much larger factor in task run times.
ID: 102154 · Report as offensive
Profile jay_e

Send message
Joined: 8 Mar 07
Posts: 112
United States
Message 102409 - Posted: 4 Jan 2021, 16:55:37 UTC

Greetings,
I have had a Radeon/ATI 7700 for many years. It had a good crunch/watt ratio.
A few years ago, I bought an Nvdia card.
Problems:
An additional power cable is required.
The cable provided is too short to reach the motherboard
The plug of the cable provided does not match the the sockets of the cables from the power supply.

I suggest taking a close-up photo of the plugs from your power supply.
Going to a store where the employees can open a box to verify pug-socket mating.

PS
Also check the capacity of your power supply. Need an upgrade?
Jay
ID: 102409 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 964
United Kingdom
Message 102411 - Posted: 4 Jan 2021, 18:58:26 UTC - in response to Message 102410.  

A lot may say 650W, but if you look on the label, you can only get 400W at 12V, and it's the 12V that the card uses.


The quoted power for the vast majority of PSUs is the TOTAL power, not the 12v - you certainly have to be careful and make sure that you meet both the total power (which includes the various 5v etc. supplies of the motherboard to function) and the demand from all peripherals (disks, GPUs, etc.) and that the individual supply lines are capable of supporting the connected things (there have been cases where the demand from say a GPU has exceeded the capacity of the selected supply from the PSU which is not conducive to a stable system.), however this is more of a problem when using high power-demand GPUs rather than low power ones.

Quite a peripheral manufacturer sites have fairly good calculators that can be used to predict what size PSU is required for a given system configuration (and amazingly enough many of them actually agree on the final size required!!!!)
ID: 102411 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 964
United Kingdom
Message 102437 - Posted: 5 Jan 2021, 18:30:37 UTC - in response to Message 102435.  

You mean "The quoted power for the vast majority of ****cheap rubbish**** PSUs is the TOTAL power, not the 12v". If you buy a Corsair, you get 12V at about 98% of the total power. Trust and Alpine (the second of which seems to have gone bust) explode often before the warranty expires, I had one Alpine go bang and make a lot of smoke when loading it at only 50% for 2 minutes. CIT manage ok, but they have a low 12V output, although they do what they say on the label, albeit with a lot of heat and fan noise as you approach maximum power. Buy a known make, then you don't have to check the 12V rating.

So you have no power left for the rest of the system - Corsair do make good PSUs, but are infamous for their optimistic rating of their 12v lines. OK, so you can get 98% of the total rated power on the 12v lines, but that leaves very little on the other lines for the rest of the system - let's say you are looking at a 1000w PSU that would mean 20w total available for the rest of the system - most modern CPUs draw much more than that on their own for example a random sample: A Ryzen 5 3600 that needs 65w on its own, that ignores the needs of memory, motherboard infrastructure and so on.... Learn to read ALL specs before you quote meaningless figures.

Your idea of running a secondary PSU is fair, but some of those numerous (very cheap) 12V PSUs are really of very bad quality and some are even knock-offs of "good" Chinese manufacturers, Meanwell are suffering a lot of this just now......
ID: 102437 · Report as offensive
Ian&Steve C.

Send message
Joined: 24 Dec 19
Posts: 156
United States
Message 102441 - Posted: 6 Jan 2021, 1:26:42 UTC

most of the modern/quality PSUs give you the rating for the 12V rail, then there's some extra to be expected for the other voltages.

for example here we can see with the EVGA 1600W, you get 1600W all on 12V, and an extra 120W for 5V and 3.3V.


it was the same situation on the Corsair units I checked.
ID: 102441 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 964
United Kingdom
Message 102443 - Posted: 6 Jan 2021, 8:18:53 UTC

I suggest you re-read the rating label that you have shown:
Combined power row gives "other voltages" = 120w, 12v=1599w, -12v =6w, +5vab = 15w
Output power (p max) = 1600W

Thus the MAXIMUM AVAILABLE POWER IS 1600W, provided one isn't pulling anything from any of the other lines, as soon as you start drawing anything from the other liens the allowable power from the 12v line drops so that the total power drawn is <=1600W.
ID: 102443 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 964
United Kingdom
Message 102444 - Posted: 6 Jan 2021, 8:45:47 UTC

I've bought batteries that are rated at 2500mAh capacity only to test them and find they're 300, not heavy enough to have that much lithium in them and are claiming to be 2.5 times better than anything Panasonic can make. I got those for free when I complained.


Well, those very light batteries, like all Lithium batteries, only contain a very small amount of lithium (about 1.5%, on an 18650 cell weight of 45g). Corners are cut by making the metal cases as thin as possible, reducing the amount of absorbed electrolyte, even having "half" the cell just a spacer (plastic tube with a bit of wire in it). In most cases these cheap cells can be identified quite rapidly by price, it really is "IF it looks too cheap to be true it isn't true" really applies. Sadly I've seen a few knock-off Panasonic cells, again at stupidly low prices - I asked one supplier the shipping weight for 1000 cells and got an impossibly low figure (about 25kg including packaging when it should have been about 47kg excluding packaging.)

And Panasonic currently manufacture a whole range of 18650 cells up to about 3400AHr - and their cells are among the best at a given rating.
ID: 102444 · Report as offensive
Ian&Steve C.

Send message
Joined: 24 Dec 19
Posts: 156
United States
Message 102446 - Posted: 6 Jan 2021, 13:57:00 UTC - in response to Message 102443.  
Last modified: 6 Jan 2021, 13:59:02 UTC

I suggest you re-read the rating label that you have shown:
Combined power row gives "other voltages" = 120w, 12v=1599w, -12v =6w, +5vab = 15w
Output power (p max) = 1600W

Thus the MAXIMUM AVAILABLE POWER IS 1600W, provided one isn't pulling anything from any of the other lines, as soon as you start drawing anything from the other liens the allowable power from the 12v line drops so that the total power drawn is <=1600W.


yes, and you still have the full 1600W available on 12v. that was my point.

and in a lot of cases, you could still pull >1600W since OCP is based on current on each rail separately. I have this exact PSU, and I have done it.
ID: 102446 · Report as offensive
robsmith
Volunteer tester
Help desk expert

Send message
Joined: 25 May 09
Posts: 964
United Kingdom
Message 102447 - Posted: 6 Jan 2021, 15:04:04 UTC - in response to Message 102446.  

WRONG - they will start to do some very strange things if you try to draw more than the Pmax - that is the absolute total that that PSU is capable of - why do you think there is a line on the label marked "Pmax"?

Having load banked a lot of PSUs from major manufactures over the years I can honestly say that very few actually work properly when one one of the rails is fully loaded. The most usual behaviour is that, when the 12v is near its maximum the 5v rail starts to go soft, and eventually that drops below its tolerable voltage range and shuts the whole PSU down. Saying that I've seen the +5v rail being capable of over double its rating before having any effect on the other outputs, and on the other hand I've seen PSUs from highly regarded manufacturers (including Corsair & EVGA) failing to supply the stated maximum power, even when all lines are inside their individual ratings.
Remember all these units are consumer grade and are built to very wide tolerances compared to real professional equipment, so while some will work beyond their ratings others will fall below (and very few of users have the equipment required to do a proper load-bank test, or can actually measure the power on the lines between the PSU and the attached equipment).
ID: 102447 · Report as offensive
Profile Dave

Send message
Joined: 28 Jun 10
Posts: 1377
United Kingdom
Message 102448 - Posted: 6 Jan 2021, 15:38:26 UTC

Remember all these units are consumer grade and are built to very wide tolerances compared to real professional equipment, so while some will work beyond their ratings others will fall below (and very few of users have the equipment required to do a proper load-bank test, or can actually measure the power on the lines between the PSU and the attached equipment).


Good point Rob. And that is why when I did the on-line building of my Ryzen machine, I ignored the message that the PSU was overspecified for the machine I was buying. In fact I got one that is overspecified even if I put in a second more powerful graphics card.
ID: 102448 · Report as offensive
Ian&Steve C.

Send message
Joined: 24 Dec 19
Posts: 156
United States
Message 102449 - Posted: 6 Jan 2021, 15:42:29 UTC - in response to Message 102447.  

well i've done it so... lol empirical evidence trumps theoretical.
ID: 102449 · Report as offensive
Previous · 1 · 2 · 3 · 4 · 5 · 6 · Next

Message boards : GPUs : Nvidia/AMD Cuda/OpenCL on Boinc projects - which card to buy?

Copyright © 2021 University of California. Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.