New Nvidia A2000 Workstation GPU

[Demmers]

Nvidia just announced the A2000 (workstation) GPU which will come out later this year https://www.nvidia.com/en-us/design-vis ... rtx-a2000/
I'm intrigued by this card, as i'm still yet to buy a dedicated GPU for my machine, but will only use it for folding. It specs out slightly less than a RTX 3060, BUT, consumes only 70W compared to 170W for the 3060! Apparently will be priced around $450/£325 (yeah sure).
I was originally thinking of getting the 3070 (the max of my budget and also the max of my mini-ITX case!), and I know this new card will obviously perform not as well, but I would love to know peoples thoughts on this from a rough guessed PPD/PPW point of view.

[toTOW]

Tesla T4 and its 70W TDP have the same performance on FAH as a 1070 ... does it answer your question ?

[Demmers]

Tesla T4 and its 70W TDP have the same performance on FAH as a 1070 ... does it answer your question ?

Good to know, thanks. I was just noting that the A2000 has 3328 CUDA cores, at just 70W. That's way more than the cards you just mentioned. But having said that, I have also noticed the base clock speed is a mere 562 MHz, so that would explain the low power draw (compared to 585 MHz for the Tesla T4 and 1506 MHz for the 1070 @ 150W). I'm just merely interested in how well this could perform as a dedicated FAH card, from an efficiency point of view, compared to the current gaming cards.

[MeeLee]

Tesla T4 and its 70W TDP have the same performance on FAH as a 1070 ... does it answer your question ?

Good to know, thanks. I was just noting that the A2000 has 3328 CUDA cores, at just 70W. That's way more than the cards you just mentioned. But having said that, I have also noticed the base clock speed is a mere 562 MHz, so that would explain the low power draw (compared to 585 MHz for the Tesla T4 and 1506 MHz for the 1070 @ 150W). I'm just merely interested in how well this could perform as a dedicated FAH card, from an efficiency point of view, compared to the current gaming cards.

~1/3rd to 1/4th the performance (based on 1/4th of the shader frequency), usually results in a much bigger PPD drop than that number.
Second, 1/3rd to 1/4th the performance, for 1/3rd the TDP, is lower efficiency.
The A2000 series is made for 4/8 bit AI calculations (like RT cores), which is useless for folding (which needs 32 bit).
There's a big race for AI, and Nvidia is cashing in on that technology.
So most of that performance will lie dormant.
It would be nice if one could increase the TDP to ~130-175W, and increase the core frequency closer to 1,6Ghz, if the hardware will allow it.

[Demmers]

~1/3rd to 1/4th the performance (based on 1/4th of the shader frequency), usually results in a much bigger PPD drop than that number.
Second, 1/3rd to 1/4th the performance, for 1/3rd the TDP, is lower efficiency.
The A2000 series is made for 4/8 bit AI calculations (like RT cores), which is useless for folding (which needs 32 bit).
There's a big race for AI, and Nvidia is cashing in on that technology.
So most of that performance will lie dormant.
It would be nice if one could increase the TDP to ~130-175W, and increase the core frequency closer to 1,6Ghz, if the hardware will allow it.

Well max boost is only recommended at 1.2GHz, so looks like i'll have to wait around to see if there's ever going to be a price drop back to near msrp and hopefully snag a 3070 then!

[toTOW]

The Tesla T4 was made for AI too, so if you use only tensor cores, you can get all its power, but as soon as you hit the SPs (Stream Processors or CUDA cores) for computations used by FAH, the more you use the SPs, the lower the performance, because of the TDP limiting everything.

For FAH, it's better to sustain higher clocks on a smaller number of SPs than to have more SPs at lower clocks.

[psaam0001]

My opinion: Give MeeLee's post (viewtopic.php?f=38&t=37343) a serious amount of consideration, before you make an investment in a "workstation" graphics card (or other product that is targeted towards those users whom are into AI).

Mind you, that many of the cards on that list will require the minimum of a 650-1000 watt power supply, depending on your CPU and system configuration. Please use the free tool at: https://www.evga.com/power-meter for guidance.

Paul

[Demmers]

My opinion: Give MeeLee's post (viewtopic.php?f=38&t=37343) a serious amount of consideration, before you make an investment in a "workstation" graphics card (or other product that is targeted towards those users whom are into AI).

Mind you, that many of the cards on that list will require the minimum of a 650-1000 watt power supply, depending on your CPU and system configuration. Please use the free tool at: https://www.evga.com/power-meter for guidance.

Paul

Thanks Paul. I had already seen that list earlier, hence one of the reasons I had already decided on a 3070. I have a 650W PSU and the Ryzen 3400G already, so I know running the 3070 shouldn't be a problem. I was just unaware of the factors mentioned earlier involving clock speed/cores.

[MeeLee]

The Tesla T4 was made for AI too, so if you use only tensor cores, you can get all its power, but as soon as you hit the SPs (Stream Processors or CUDA cores) for computations used by FAH, the more you use the SPs, the lower the performance, because of the TDP limiting everything.

For FAH, it's better to sustain higher clocks on a smaller number of SPs than to have more SPs at lower clocks.

On the other hand, more shaders, running at lower frequencies, are much more power efficient than fewer higher frequency shaders.
Though, the difference is very minimal nowadays, as processors are now manufactured smaller than the 14/10nm of a few years ago.

[psaam0001]

After getting my Ryzen 3 3200G system back to folding decently (following a UEFI/BIOS update), I am not going to tempt fate with adding an NVidia card again--until funds are available to upgrade it to a 3950X or 5950X 16-core Ryzen 9.

Paul

[Frontiers]

and 1506 MHz for the 1070 @ 150W).

Nope, mine GTX 1070 at 1506 MHz in Windows have 80W powerdraw, because voltage drops significantly and power draw depends on voltage with power of two and depends on clocks linearly with power of one.
Gated to 150W GTX 1070 will have clocks between 1950 and 2000 MHz, gated to 130W will have ~1900 MHz.

Tesla T4 have passive radiator for using with row of rack turbines at 4-6k rpm, it wouldn't work in regular case with low pressure front fans at 700-1000 rpms.
A2000 will have PPD close to GTX 1060 6GB @ core_22, 1066 draws around 110 watts, but cost used half of A2000.

My opinion: Give MeeLee's post (viewtopic.php?f=38&t=37343) a serious amount of consideration

With 3070ti at 1st place, which draws close to 3080 - 70-80 watts more than 3070, but fold close to 3070, and sells with cooler like from 3070 but not form 3080 - that's BS.
From trio of 3070, 3080 and 3070ti - 3070ti have highest chances of frying it GDDR6x, because of more budget cooler with very hot VRAM positioned close to hot chip.
Basically both 3080 and 3070ti - for 24/7 compute loads should be used only with watercooling, for not burning their hot VRAM.

[MeeLee]

My opinion: Give MeeLee's post (viewtopic.php?f=38&t=37343) a serious amount of consideration

With 3070ti at 1st place, which draws close to 3080 - 70-80 watts more than 3070, but fold close to 3070, and sells with cooler like from 3070 but not form 3080 - that's BS.
From trio of 3070, 3080 and 3070ti - 3070ti have highest chances of frying it GDDR6x, because of more budget cooler with very hot VRAM positioned close to hot chip.
Basically both 3080 and 3070ti - for 24/7 compute loads should be used only with watercooling, for not burning their hot VRAM.

TDP can be adjusted easily on Nvidia GPUs.
Lowering TDP, lowers temperature of the cooler, which in turn allows VRAMs to run cooler.
FAH uses anywhere between 500MB to 1GB of VRAM. All other modules are passive, which means between 75 to 90% of the VRAM is unused, and cool.
some models did come out where the manufacturer neglected to remove the thermal paste stickers, so the VRAM runs hot.
If your VRAm runs hot, you can either return your GPU and expect a replacement within warranty, within a good 2 to 3 months.
But if you don't want to wait so long, all you have to do, is remove the shroud, heatsink, and remove the thermal pad stickers.

[psaam0001]

Right now, I put that Zotac GTX 1650 back into my Windows 10 machine (now that I massaged a few UEFI settings to get the AMD Ryzen 3's iGPU running smoother), without the noisy fan. I'm still waiting for an e-mail from the folks at gpufanreplacement.com to give me an estimate on the replacement pair of fans. If it's >$50.00 USD, I'm just going to put the working fan I'm salvaging on the GT 1030 from Gigabyte (that also needed a single replacement fan).

Paul

[MeeLee]

Right now, I put that Zotac GTX 1650 back into my Windows 10 machine (now that I massaged a few UEFI settings to get the AMD Ryzen 3's iGPU running smoother), without the noisy fan. I'm still waiting for an e-mail from the folks at gpufanreplacement.com to give me an estimate on the replacement pair of fans. If it's >$50.00 USD, I'm just going to put the working fan I'm salvaging on the GT 1030 from Gigabyte (that also needed a single replacement fan).

Paul

I know Asus and Gigabyte won't replace the fan.
They will charge for 'repairs' which for Gigabyte was $150 (triple fan), and Asus was $700 (higher than the price of the 2070 I paid at that time.

[psaam0001]

Right now, I put that Zotac GTX 1650 back into my Windows 10 machine (now that I massaged a few UEFI settings to get the AMD Ryzen 3's iGPU running smoother), without the noisy fan. I'm still waiting for an e-mail from the folks at gpufanreplacement.com to give me an estimate on the replacement pair of fans. If it's >$50.00 USD, I'm just going to put the working fan I'm salvaging on the GT 1030 from Gigabyte (that also needed a single replacement fan).

Paul

I know Asus and Gigabyte won't replace the fan.
They will charge for 'repairs' which for Gigabyte was $150 (triple fan), and Asus was $700 (higher than the price of the 2070 I paid at that time.

That Zotac GTX 1650 Super is only a 2 fan card, in which one of them started whining. So, the one that was not raising a fuss will go on that GT 1030 (it uses the same size fan).

I'm only buying EVGA or MSI cards when I need them.

Paul