Nvidia Ampere Discussion [2020-05-14]

[MDolenc]

CarstenS: tensor cores lower register file bandwidth per math operation because they work on tensors rather than scalars (N^3/N^2).
Nvidia actually shows this phenomenon in their animated tensor core cartoons in their keynotes.

So it’s likely they are close to peak RF bandwidth both at 78 scalar TFlops as well as at 312 tensor TFlops.

That's said it much better yes.
The implication from my post might be that there's a ton of RF bandwidth available if not using tensor cores which is not the case. Tensor cores could also raise number of active registers of a kernel thus hurting occupancy a bit.

 

[swaaye]

Just noticed the RTX3080 requires a 750W power supply. I have a 650W one and my CPU is Ryzen 3900X. I wonder if my power supply would be constrained if I got a 3080. What do you think?

I think it would be questionable but I would put an AC wattmeter on your current setup and see what you get with various games.

I have a little Seasonic G 550RM powering a 4.8 GHz 8600K with 1080 Ti. No problems. The PSU fan does spin up with some game loads though which is annoying. I would definitely swap it out if I went to an even beefier GPU.

 

[pharma]

The GeForce RTX 30 Cards Are Insane. What’s Coming Up for Quadro?

Here’s where we have to take a step back and recognize that these three new cards, like all GeForce cards, are consumer tier. They’re made for gamers, and gamers are pretty psyched. But NVIDIA also makes a professional lineup of graphics cards, the Quadro series. If you don’t know the differences between the two, read What’s the Difference Between GeForce and Quadro Graphics Cards?

So if these consumer cards are such a generational leap forward, what will the next Quadro cards look like? My hopes are pretty high. That top-of-the-line BFGPU up there boasts 24GB of VRAM, more than double the last-gen 2080 Ti. The current best Quadro card, the Quadro RTX 8000, is already sitting at 48GB. Will we see a new Quadro card with 96GB of VRAM?

Here’s another point of reference. The Quadro RTX 8000 has 4608 CUDA Cores, the main computational element of NVIDIA GPUs. The entry-level GeForce RTX 3070 has 5888, and the BFGPU has a staggering 10,496. How many more cores are we going to see in the next Quadro series?

https://www.engineering.com/Hardware/ArticleID/20679/The-GeForce-RTX-30-Cards-Are-Insane-Whats-Coming-Up-for-Quadro.aspx

 

[Scott_Arm]

I think it would be questionable but I would put an AC wattmeter on your current setup and see what you get with various games.

I have a little Seasonic G 550RM powering a 4.8 GHz 8600K with 1080 Ti. No problems. The PSU fan does spin up with some game loads though which is annoying. I would definitely swap it out if I went to an even beefier GPU.

I'm thinking of doing the same. If I get one I'm going to undervolt it, so it's very possible I'll be able to run on my 650Watt supply. The problem is, you don't know the results you're going to get, and you have to work your way down from stock voltage, so you need to be able to support stock voltage anyway. It's like buying the supply to use it for a couple hours until you can lower the power so you don't need it anymore. I might measure my system and see if it looks possible that I could run the card.

Actually, I wonder if it's possible if I found a 2nd supply that I could borrow, maybe even from work, that I could rig it up to power the gpu directly with the side panel open. Then I could potentially undervolt the gpu then remove and return the second supply. Frig, that won't work because of the pcie power. Probably super complicated. Voltage differences would do bad things between the two supplies.

 

[pharma]

https://twitter.com/i/web/status/1302216936563126272

 

[Scott_Arm]

https://twitter.com/i/web/status/1302216936563126272

Guess we'll have very accurate numbers.

 

[pharma]

The beauty is it also works on competitor cards.

 

[pjbliverpool]

The beauty is it also works on competitor cards.

I'm sure that's the point. Nvidia want to highlight their latency superiority over the competition and traditional tools don't allow that.

 

[Malo]

I'm sure that's the point. Nvidia want to highlight their latency superiority over the competition and traditional tools don't allow that.

Isn't this for power usage analysis?

 

[CarstenS]

Isn't this for power usage analysis?

Yes.

 

[Cyan]

So that 310W is already taking into account boost! That's not so bad hmm..

not too shabby. Still for my PC's psu in particular, the 3070 -still waiting for what AMD is going to show in november- seems the best value.

https://videocardz.com/newz/nvidia-details-geforce-rtx-30-ampere-architecture (screengrab from this article)

 

[CarstenS]

The GeForce RTX 30 Cards Are Insane. What’s Coming Up for Quadro?

https://www.engineering.com/Hardware/ArticleID/20679/The-GeForce-RTX-30-Cards-Are-Insane-Whats-Coming-Up-for-Quadro.aspx

That article is misleading about a couple of things at best (that ISVs tune their applications for Quadro and not Nvidia disabling certain things in the driver for ginormous performance gains for example).
Quadro RTX could use the full complements of GA102 (10752 ALUs) and 104 (6144 ALUs), additionally, they can use twice as much memory (in clamshell mode) as RTX 3080 and 3070. For doubling of VRAM over RTX 3090, they would need to wait until Micron uses 16 GBit dies.

 

[pjbliverpool]

Isn't this for power usage analysis?

My bad, I didn't pay much attention to that part of the presentation tbh!

 

[trinibwoy]

The beauty is it also works on competitor cards.

What does nvidia gain from providing reviewers with more accurate power testing equipment.

 

[BRiT]

How do we know it's accurate? Has anyone done a calibration verification? </conspiracy>

 

[Picao84]

not too shabby. Still for my PC's psu in particular, the 3070 -still waiting for what AMD is going to show in november- seems the best value.

https://videocardz.com/newz/nvidia-details-geforce-rtx-30-ampere-architecture (screengrab from this article)

Yeah, I'll probably will end up with a 3070 for the same reason. Just annoyed that it doesn't have GDDR6X. That 760GB/s on memory is impressive.

 

[pharma]

What does nvidia gain from providing reviewers with more accurate power testing equipment.

I assume reviewers will see what the variance is between the traditional method and the nvidia provided tool. It would be interesting to see if there is a significant difference.

 

[Scott_Arm]

Imagine console gaming at 30fps ... lol

 

[swaaye]

I'm thinking of doing the same. If I get one I'm going to undervolt it, so it's very possible I'll be able to run on my 650Watt supply. The problem is, you don't know the results you're going to get, and you have to work your way down from stock voltage, so you need to be able to support stock voltage anyway. It's like buying the supply to use it for a couple hours until you can lower the power so you don't need it anymore. I might measure my system and see if it looks possible that I could run the card.

Actually, I wonder if it's possible if I found a 2nd supply that I could borrow, maybe even from work, that I could rig it up to power the gpu directly with the side panel open. Then I could potentially undervolt the gpu then remove and return the second supply. Frig, that won't work because of the pcie power. Probably super complicated. Voltage differences would do bad things between the two supplies.

It wouldn't hit peak power use just sitting in the BIOS or Windows. Problems could come once you are running AVX2 thru multiple CPU cores and have the GPU doing something that pushes it to high utilization. So there shouldn't be any problem with stock voltages unless you start playing games.

 

[Scott_Arm]

It wouldn't hit peak power use just sitting in the BIOS or Windows. Problems could come once you are running AVX2 thru multiple CPU cores and have the GPU doing something that pushes it to high utilization. So there shouldn't be any problem with stock voltages unless you start playing games.

The problem is to undervolt you'll basically run a benchmark at 100% load and then slowly adjust your voltage vs frequency curve so it flattens out peak frequency to lower and lower voltages until you start getting errors. You'll never really know how well it's going to work, or how low you'll be able to go, especially on a new architecture. So you're probably going to be running 100% gpu for some time near stock voltage as you lower it a bit at a time.