nvidia DGX A100 and Ampere Feature Rumors

[Results45]

So yeah.......it might just all exist 

Quote

.

• Ampere-based DGX A100
• GA100 to have between 7552-8192CUs @ ~1800Mhz
• DLSS 3.0
• 2x Tensor & RT cores or more compared to Turing (reduced "RTX ON" frame drop and up to 4x DXR performance)
• GeForce Experience NOW software
• Tensor-accelerated "Lossless VRAM Compression"
• up to 20%+ IPC gains
• 100% RTX lineup
• NVCache ~ leverages DDR/DRAM and SSD for extra memory access and faster load times (functions similar to XBSX & PS5 SoC)

.

 

SOURCES & IMAGES:

• https://trademarks.justia.com/888/54/dgx-88854783.html

• https://youtu.be/uqillUSZvMM*

 

 

*Tom from Moore's Law is Dead regularly gets bits and pieces of confirmed info from anonymous "insider sources" (employees of Nvidia, their partners, or 3rd parties with NDA info access from how he describes them) and one of them helped him piece together the lists above.

[Skiiwee29]

A) I dont trust WCCFTech.

 

B) I dont trust Tweak Town. Its bad when Raja Koduri had to rebuke there tweets/story on the intel GPUs yesterday on twitter because of the false information they are spreading and click baiting on there articles. 

 

Spoiler

 

[LAwLz]

I think "rumored" would be more accurate than "confirmed".

The source is a Youtube video from someone I've never heard of. Probably because it's a person with like 35K subs.

 

I wouldn't be surprised if his "source" is just someone messing with him, or maybe he just made it up and there isn't any source at all. I mean, he even says his source contracted him through email. He also spends several minutes going on and on about how cool it is to have an "insider source" basically.

[The_russian]

For anyone that wants a link directly to the trademark application:

https://trademarks.justia.com/888/54/dgx-88854783.html

Not sure why it wasn't included in the original post. 

[The_russian]

The title is really misleading, there should be a distinction made between what is confirmed and what is still rumored. As far as I can tell, the only thing that has been confirmed so far is the trademark of DGX A100, the rest is all rumors.

[Vulkan HeStan]

25 minutes ago, LAwLz said:

I think "rumored" would be more accurate than "confirmed".

The source is a Youtube video from someone I've never heard of. Probably because it's a person with like 35K subs.

 

I wouldn't be surprised if his "source" is just someone messing with him, or maybe he just made it up and there isn't any source at all. I mean, he even says his source contracted him through email. He also spends several minutes going on and on about how cool it is to have an "insider source" basically.

I've had his videos suggested to me before and in one of them he says "there's no chance of a zen 2 4 core", like a month before the 3100 and 3300x were announced. So his "insider sources" don't seem all that great. 

[LogicalDrm]

@Results45

Your thread is missing personal input and quote box for the speculated specs. As such thread has been moved to Graphics Cards. Please fix this and thread will be moved back to Tech News.

Edited May 4, 2020 by LogicalDrm

[leadeater]

Confirmed rumor, yep....

 

It's also confirmed I'm a billionaire, so my reliable sources say.

[Soppro]

Hmm ... probably related to a new Tesla A100, is it not?

[leadeater]

31 minutes ago, Soppro said:

Hmm ... probably related to a new Tesla A100, is it not?

In a way both. With the new Tesla A100 (go with that for now) GPUs they will update their DGX platforms as well. There's enough we know about how Nvidia does things we can make some pretty educated guesses and get a lot of things right, that means we can also make claims and rumors backed by 'claimed sources' which is really just spit balling and guessing hoping enough is actually true so people won't think you were completely full of it.

 

If I were to guess it would be that if Nvidia were looking at a more cost focused product, unlike Volta, then Ampere GA100 will not be larger than ~7500 CUDA cores (7200 my personal max) and will be around the 600mm2 to 700mm2 area.

[3vitor]

Honestly WCCFTECH should just be banned as a source, even for rumors. They will regurgitate any rumor for clicks without a reliable source.

[illegalwater]

No login required for geforce experience? Overhauled control panel? Hahahaha.. never happening.

 

Also is it implied that "DLSS 3.0" would be a driver level feature? That's impossible, DLSS requires engine level tweaks to work. Including changes to texture rendering as of 2.0, I doubt Nvidia wants to regress to the muddy textures of DLSS 1.0.

 

Edit:

 

Didn't Nvidia deny the Samsung 8nm rumors? Pretty sure they said they're using TSMC 7nm.

[Results45]

5 hours ago, leadeater said:

 

If I were to guess it would be that if Nvidia were looking at a more cost focused product, unlike Volta, then Ampere GA100 will not be larger than ~7500 CUDA cores (7200 my personal max) and will be around the 600mm2 to 700mm2 area.

 

Then what do you make of speculations based on a photo and Geekbench result?

• https://www.guru3d.com/news-story/rumor-nvidia-geforce-rtx-3070-and-3080-coming-q3-2020-specs.html
• https://browser.geekbench.com/v5/compute/206992

.

[BOOM BOX]

3 hours ago, Results45 said:

 

Then what do you make of speculations based on a photo and Geekbench result?

• https://www.guru3d.com/news-story/rumor-nvidia-geforce-rtx-3070-and-3080-coming-q3-2020-specs.html
• https://browser.geekbench.com/v5/compute/206992

.

Firstly: a fine gentleman in the guru3d comments made a valid point saying and I quote, "The 3080Ti having almost twice the cores than the 3080 seems a little fishy..." I have a hard time trying to justify any comments denying such a statement. The only comment that could be made in opposition would be perhaps the name SKU is misleading and Nvidia is considering reverting back to the original release format that they did with the 10xx series and prior. But that would be dumb as it's been proven that releasing the Titan before the Ti only cannibalizes Titan sales, not that they really sell a lot of them in the first place.

 

Secondly: The geekbench results throw a massive red flag when you look at the CUDA information. Now, someone correct me if I'm wrong but 1.11 GHz max frequency on a 30xx compared to a 1.545 GHz boost clock on a 2080 Ti with base being 1.35 GHz seem to be a little bit of a step backwards. I would be willing to argue that just because you theoretically have near double the CUDAs, doesn't mean you should back off on clocks, if anything keep the clocks the same or give a marginal bump in clock speeds. If I were to spit ball here, which I am, I would spit ball that the info in this spread sheet could possibly be pointing to a workstation card or a server card.

 

Like I said before, I'm spit balling here based on what my gut is telling me, I'm sure someone with more knowledge or experience around this stuff would be willing to fill in the gaps or correct me on any of this. Not trying to be rude, I'm just laying things out that I find that could be red flags.

[leadeater]

4 hours ago, Results45 said:

 

Then what do you make of speculations based on a photo and Geekbench result?

• https://www.guru3d.com/news-story/rumor-nvidia-geforce-rtx-3070-and-3080-coming-q3-2020-specs.html
• https://browser.geekbench.com/v5/compute/206992

.

Already posted what i think about those in the news thread here about it. The first image is just wrong, the die code names don't match where Nvidia actually uses them nor do they really do Gx103 dies. The difference between the 3080 Ti and 3080 is so so far out of wack, not that the 3080 Ti is going to use GA100 that's a Tesla die with HBM memory controller.

 

As for the Geekbench here is one for a Tesla V100 (GV100 with 80 SMs and 64 CUDA cores per SM)

 

So this rumor one is most certainly a Tesla part using HBM2E and has, if reported correctly, 124 SMs but we don't know if the CUDA core count per SM has changed.  Things don't always scale nicely so going from the current 5120 CUDA cores to near 8000 may not work and it also many not leave room for improvements in Tensor cores and, if Nvidia wants these in a Tesla, RT cores.

 

I don't know how easy it is to fake GeekBench data but the score is too close to a Tesla V100 so it could just be that with faked CUDA information. It doesn't look trustworthy enough for me to believe any of it.

 

If there is one thing you can take away from this is that anything you see about a Tesla is not going to be close to a 3080 Ti.

[Results45]

For all you doubters:

 

As many shipfulls of salt you may need to gorge on, some of the info and analysis above has to be correct, right?

[LAwLz]

15 hours ago, Results45 said:

As many shipfulls of salt you may need to gorge on, some of the info and analysis above has to be correct, right?

Well yeah, some info might be correct. I can make a ton of guesses and get some stuff correct too.

Even a broken clock is correct twice a day.

 

Guess we'll have to wait and see how much stuff he got right.

If he's wrong about a bunch of stuff, then he is an unreliable source and can't be trusted in the future. If he's right about non-obvious stuff then he might be reliable and worth believing in the future.

[Results45]

STH's NF5488M5 Review: Inspur's half-size DGX-2

 

Patrick does a pretty good job explaining in general the overall networking structure, what the DGX/HGX-2 layout is, and how NV-Switch works to connect all 8 cards more effectively than NV-Link (8:05 to 13:35):

 

https://www.servethehome.com/inspur-nf5488m5-review-a-unique-8x-nvidia-tesla-v100-server/

https://de.inspur.com/de/2464830/2464866/2507544/index.html

 

I kinda hope that GA100 or GA102 based Titan cards will support the aforementioned "NV-SWITCH".

[leadeater]

@Results45 @Soppro

Think I was close 

 

[Soppro]

1 hour ago, leadeater said:

@Results45 @Soppro

Think I was close 

The fact that they managed an 826mm^2 die on 7nm is impressive!

[Results45]

.

 

Nvidia: has built and run a private multi-exaflop supercomputer for the past 1-2 years

Also Nvidia: uses Epyc Rome CPUs in DGX A100 servers

Intel: double roasted (yes, like the salsa from Trader Joe's)

 

11 hours ago, leadeater said:

@Results45 @Soppro

Think I was close 

 

 

Here's additional specs from TechPowerUp: https://www.techpowerup.com/gpu-specs/tesla-a100-sxm4.c3506

 

So the A100 spits out 19.49 raw TFLOFS at 1410Mhz, but since Tech Jesus comfortably overclocked the Titan V to more than 1600Mhz and the Titan RTX to more than 1900Mhz (before switching to water or LN2) I think it's safe to say that an OC'd "Titan A" card will easily hit 22.12 TFLOPS if based on GA100.