NVIDIA Ampere: NVCache speeds up load times, optimize VRAM usage

erek

erek

8-D
2FA
Joined
Dec 19, 2005
Messages
18,774
Just a rumor. Hmm

"The same sources have said something similar to what I reported not too long ago: that the next-gen GeForce RTX 3060 as a "mid-range" card will have the performance of the now-flagship GeForce RTX 2080 Ti. Except, Tom reports that it is in ray tracing -- which seems to make sense. We don't know about the actual performance of this card obviously, which is where the salt is required.

We should hear all about Ampere (or at least as much as NVIDIA will share) during its GTC 2020 keynote in a few days. We should be introduced to the inner workings of the new Ampere architecture, with the GeForce RTX 3000 series cards to be unveiled somewhere around September."

Read more: https://www.tweaktown.com/news/7239...-up-load-times-optimize-vram-usage/index.html
 
I'm not sure that CUDA 11 "insight" is as useful as the OP cited in the video thinks. The drivers have been reporting back CUDA 11 support for at least a year when probed by things like Iray. It's been a little bit of a PITA actually, since tech support reviewing error logs uses this to fallback on "oh, your driver is too new" in order to close the ticket without doing anything.

I'm also not sure why there is buzz about 1900Mhz boost clocks. The 2080 Ti does that stock.
 
Last edited:
Thunderdolt said:
I'm not sure that CUDA 11 "insight" is as useful as the OP cited in the video thinks. The drivers have been reporting back CUDA 11 support for at least a year when probed by things like Iray. It's been a little bit of a PITA actually, since tech support reviewing error logs uses this to fallback on "oh, your driver is too new" in order to close the ticket without doing anything.

I'm also not sure why there is buzz about 1900Mhz boost clocks. The 2080 Ti does that stock.
Click to expand...

Think the 3080 Ti will have 16GB of VRAM?
 
Ray tracing has been so obnoxiously underwhelming across the industry that im not excited about anything nvidia yaps about concerning the tech. We were sold a bill.of goods that hardly anyone has taken advantage of. Maybe well see.
 
erek said:
Think the 3080 Ti will have 16GB of VRAM?
Click to expand...

We're now 3 years beyond the launch of the 11GB 1080 Ti and we don't see any games pushing the limits of that VRAM capacity. We're 4 years beyond the launch of the 8GB 1080, and now we're finally seeing some games come close to that. In terms of useful for gaming, sticking with 11GB would probably be just fine.

That said, the GTX & RTX cards pretty much own the professional market despite the Quadro line having the official label. In the pro market, there has has been an increasing need for more VRAM. I'm not sure what % of Geforce sales this segment accounts for, so it might not be large enough to move the needle.

With that said, I would guess that it is more likely to have 11 or 12, but I don't have much faith in the guess.
 
Thunderdolt said:
We're now 3 years beyond the launch of the 11GB 1080 Ti and we don't see any games pushing the limits of that VRAM capacity. We're 4 years beyond the launch of the 8GB 1080, and now we're finally seeing some games come close to that. In terms of useful for gaming, sticking with 11GB would probably be just fine.

That said, the GTX & RTX cards pretty much own the professional market despite the Quadro line having the official label. In the pro market, there has has been an increasing need for more VRAM. I'm not sure what % of Geforce sales this segment accounts for, so it might not be large enough to move the needle.

With that said, I would guess that it is more likely to have 11 or 12, but I don't have much faith in the guess.
Click to expand...


I need the 16GB on consumer for AI/ML purposes :( :(

"RuntimeError: CUDA out of memory. Tried to allocate 26.00 MiB (GPU 0; 11.00 GiB total capacity; 8.74 GiB already allocated; 2.59 MiB free; 8.75 GiB reserved in total by PyTorch)"


https://github.com/facebookresearch/ParlAI/issues/2619
 
Would be surprising if 3rd gen have 11gb ram.
It's time for more ram. Even though no game today makes use of it, I want to be somewhat future proof for a couple of generations.
 
Nebell said:
Would be surprising if 3rd gen have 11gb ram.
It's time for more ram. Even though no game today makes use of it, I want to be somewhat future proof for a couple of generations.
Click to expand...

Needs more ram, bad !!! :( :(

"Chatting with the models
You may talk with our models. The 2.7B can be interacted with on a 16gb P100 GPU or better. The 9.4B parameter model requires at least two 32gb V100 GPUs to interact with."

https://parl.ai/projects/recipes/
 
erek said:
Needs more ram, bad !!! :( :(

"Chatting with the models
You may talk with our models. The 2.7B can be interacted with on a 16gb P100 GPU or better. The 9.4B parameter model requires at least two 32gb V100 GPUs to interact with."

https://parl.ai/projects/recipes/
Click to expand...

Can that be run without NVlinked cards? If so, maybe I should try it on my 4x Titan RTX setup (96GB).
 
Thunderdolt said:
Can that be run without NVlinked cards? If so, maybe I should try it on my 4x Titan RTX setup (96GB).
Click to expand...

Wow, 96GB, that's CRAZY !!!!!
I am not sure !!!! man what are you doing with 4x Titan RTX anyhow?
 
erek said:
I need the 16GB on consumer for AI/ML purposes :( :(

"RuntimeError: CUDA out of memory. Tried to allocate 26.00 MiB (GPU 0; 11.00 GiB total capacity; 8.74 GiB already allocated; 2.59 MiB free; 8.75 GiB reserved in total by PyTorch)"


https://github.com/facebookresearch/ParlAI/issues/2619
Click to expand...

NV will have to add even more bus from the gpu to the ram chips. I dont think they will do this. It segments the consumer from the pro. The pro has the highest bus widths and of course the premium prices

Ive already bitched about this here and the other people here flamed me to death. Shit like "if you need 24GiB of vram then buy titan etc..." no ability to empathize with the fact that a lot of people are super users at home and dont have the bank to afford a $9000 gpu when adding a little vram might cost another 200 markup for nvidia.
 
Last edited:
erek said:
is that some form of a joke?
Click to expand...

Yes.

I use the machine for simulations, rendering, and photogrammetry. I was hitting VRAM limits on 2080Tis, so I upgraded to the Titans. I'd still like to see a 48GB Titan Ampere though. I feel like with only 24GB, the effect of a Titan Ampere's compute capability would be hindered by its lack of increased VRAM, if that makes any sense. Like, if today I'm running a simulation for 30hrs, then a 24GB Titan Amp would maybe run the same sim in 24 hours. But if I'm already ok with letting it go for 30hrs, then I'm going to need more VRAM onboard in order hold the larger dataset required to keep the GPU loaded the whole time.
 
tangoseal said:
NV will have to add even more bus from the gpu to the ram chips. I dont think they will do this. It segments the consumer from the pro. The pro has the highest bus widths and of course the premium prices

Ive already bitched about this here and the other people here flamed me to death. Shit like "if you need 24GiB of vram then buy titan etc..." no ability to empathize with the fact that a lot of people are super users at home and dont have the bank to afford a $9000 gpu when adding a little vram might cost another 200 markup for nvidia.

So your barking up the wrong tree here at oft Forums
Click to expand...

It's more like we aren't [C]heap Forums. If you don't have the bank then learn to code.

:)
 
I think they might stretch it out to 12 or 16 GB for the top end 8 for the mid and 4 for the low. I could see the Titan coming in a 36gb configuration this time around. I am wondering if they are going to be using a HBM configuration for this generation.
 
I'm hoping it has 16GB VRAM, I want at least more than the 11GB the 1080TI has.

I already use RAMcache for games and it speeds up load times noticeably. Kinda shocked it isn't more often utilized, or at least, annoyed that the Windows default use of extra ram is so trash.
 
TheSlySyl said:
I'm hoping it has 16GB VRAM, I want at least more than the 11GB the 1080TI has.

I already use RAMcache for games and it speeds up load times noticeably. Kinda shocked it isn't more often utilized, or at least, annoyed that the Windows default use of extra ram is so trash.
Click to expand...

Do you have mVME? Or just sata? Ive noticed no diff with a ram cache on my nvme system. I think mostly because there is decompression that is limited by CPU throughput.
 
Thunderdolt said:
I'm not sure that CUDA 11 "insight" is as useful as the OP cited in the video thinks. The drivers have been reporting back CUDA 11 support for at least a year when probed by things like Iray. It's been a little bit of a PITA actually, since tech support reviewing error logs uses this to fallback on "oh, your driver is too new" in order to close the ticket without doing anything.

I'm also not sure why there is buzz about 1900Mhz boost clocks. The 2080 Ti does that stock.
Click to expand...

Unless they are talking about the actual on the box boost where the 2080ti reference is 1635Mhz. Assuming all things equal, it’d be a 300Mhz increase and real world boost would be ~2250Mhz.

If they are already talking about real world boost - yeah, who cares.

Optimistically from rumors we’re looking at 15% IPC *10% more Mhz = 27% faster for same core count. Lets assume 50% more cores that’s a 90% increase in rasterized.

Optimistically...
 
Dayaks said:
Unless they are talking about the actual on the box boost where the 2080ti reference is 1635Mhz. Assuming all things equal, it’d be a 300Mhz increase and real world boost would be ~2250Mhz.
Click to expand...

This is exactly what makes it so difficult to read into.
  • If it's just 1900Mhz actual, then I really don't care.
  • If it's 1900Mhz advertised, but the change is that Nvidia is going to start advertising the actual boost freq, then I still don't care.
  • If this is the advertised and Nvidia is going to keep the same safety margin between advertised and actual, then sweet baby Jesus this is going to be a beefy GPU.
Let's hope it's the last one.
 
Thunderdolt said:
Yes.

I use the machine for simulations, rendering, and photogrammetry. I was hitting VRAM limits on 2080Tis, so I upgraded to the Titans. I'd still like to see a 48GB Titan Ampere though. I feel like with only 24GB, the effect of a Titan Ampere's compute capability would be hindered by its lack of increased VRAM, if that makes any sense. Like, if today I'm running a simulation for 30hrs, then a 24GB Titan Amp would maybe run the same sim in 24 hours. But if I'm already ok with letting it go for 30hrs, then I'm going to need more VRAM onboard in order hold the larger dataset required to keep the GPU loaded the whole time.
Click to expand...

Out of curiosity- what kinds of simulations are you running?
 
tangoseal said:
Do you have mVME? Or just sata? Ive noticed no diff with a ram cache on my nvme system. I think mostly because there is decompression that is limited by CPU throughput.
Click to expand...
NVME.
Two tiered cache, games are on HDD, L2 cache is 1TB NVME (3500mbps) and then 8 gigs L1 RAMcache.

I use Primocache.

It definitely depends on the game, but the ones that are using huge high res texture packs it makes a difference.

I also use 8GB RAMcache on my OS drive, where it's utilized over 50% of the time (68.7% as of right now), which at the very least saves a ton of wear on my OS drive. Nearly impossible to benchmark or guesstimate performance increase there, aside from a general "faster" feeling.
 
Last edited:
I think we're due for an increase in ram capacity on 256-bit Nvidia cards - it's been 4 tears since he GTX 1080 launched at 8GB, and even the new consoles have managed to double their VRAM.

They could pull a GTX 660 Ti 2GB on 192-bit trick, but I'm not sure Nvidia wants to do that again? But if they did it, they could give you 8GB RTX 3060, 12GB RTX 3070 , and a full 16GB for 3080!

This thing must be available in quantity by-now, and it will be a whole lot cheaper than going GDDR5SX on Pasca (and you're going to need the speed bump to feed the new 3080 card!)

So, use lower-density chips on 3060 and 3070, and only pay the premium for the high-end parts!

For me, the biggest question will be if 3060 launches at 192-bit (likely 9GB split ram) or jumps-up to 256?

I think the major reaason here fir more VRAM is the 4x RT performance increase: from what In can gather, turning on RT takes significantly more VRAM! This would give the new high-performance cards more flexibility in newer games.
 
Last edited:
You must log in or register to reply here.
Back
Top