Nvidia Post-Volta (Ampere?) Rumor and Speculation Thread

[Deleted member 13524]

But do we know what process was used on the Terminators chip? I mean, that could be billions of transistors more

Stupidest thing about deep learning inference is that there's no learning at all.
Since all Terminators can learn, that definitely isn't Terminator's microchip.

 

[BRiT]

But they didn't learn at all. If they did, their plan would have been to go undercover as female, seduced John Conner, get engaged, then married, buy a house and maybe even have kids. Thus ending his life.

 

[pharma]

NVIDIA Working on Tile-based Multi-GPU Rendering Technique Called CFR - Checkered Frame Rendering
November 21, 2019
"Forum user Blair at 3DCenter had a sharp eye noticed an added entry towards the drivers for Muli-GPU rendering, the technique is called CFR and basically slices up a frame in many small pieces, in order for the GPUs the render them in a parallel manner.

You could also refer to the technique as checkerboard rendering, where you split everything up into smaller tiles and have the GPUs render them based on an algorithm or simply, FIFO, first-in, fist-out, this could increase scaling performance but also helps with things like micro stuttering as frames and their output pacing are processed in way more stable manner. The basis is, of course, an existing technique applied in many solutions. NVIDIA, however, wants to use if for multi-GPU rendering.
...
Since CFR is currently activated with the help of extra tools and/or requires some manual work at Tweaking. The results and entries that NVIDIA is actively working on this methodology. The new technique would be DirectX compatible only, and as it seems for Turning and upcoming based GPUs as it will require NVLink."

https://www.guru3d.com/news-story/n...que-called-cfr-checkered-frame-rendering.html

 

[DavidGraham]

NVIDIA just announced Orin, the SoC that follows Xavier, coming with a next gen NVIDIA GPU.

Xavier had a 512 ALU Volta core, with 7 billion transistors and 30 TOPs of performance.
Orin has an unknown number of ALUs, but packs 17 billion transistors and 200 TOPs of performance, this is clearly made on 7nm, given it's insane density. Also it uses ARM Hercules CPUs which are made either on 7nm or 5nm.

For comparison, a full Turing GPU (Titan RTX ) is made of 18.5 billion transistors on 12nm.

https://wccftech.com/nvidia-next-gen-gpu-architecture-powered-orin-soc-announced/

 

[DegustatoR]

 

[no-X]

Orin has an unknown number of ALUs, but packs 17 billion transistors and 200 TOPs of performance, this is clearly made on 7nm, given it's insane density.

Where does this information come from? According to my source it is made on 8nm process (Samsung). Maybe my source is wrong.
/typo

 

[DavidGraham]

Where does this information come from? According to my source it is made one 8nm process (Samsung). Maybe my source is wrong.

It's mere speculation. Do you mean 8nm from Samsung?

 

[no-X]

Exactly.

 

[pharma]

Where does this information come from? According to my source it is made on 8nm process (Samsung). Maybe my source is wrong.

Anandtech says it pretty much a given that it will be on 7nm or 5nm, though Nvidia hasn't stated yet.

Overall, Orin is a 17 billion transistor chip, almost double the transistor count of Xavier and continuing the trend of very large, very powerful automotive SoCs. NVIDIA is not disclosing the manufacturing process being used at this time, but given their timeframe, some sort of 7nm or 5nm process (or derivative) is pretty much a given. And NVIDIA will definitely need a smaller manufacturing process – to put things in comparison, the company’s top-end Turing GPU, TU102, takes up 754mm2 for 18.6B transistors, so Orin will pack in almost as many transistors as one of NVIDIA’s best GPUs today.

https://www.anandtech.com/show/1524...-orin-a-herculean-arm-automotive-soc-for-2022

 

[DegustatoR]

It's a GV100 SoC in 25-50W power envelope. There's just no way it would be on 16nm.

 

[no-X]

pharma: I'd say it just means that they are guessing.

 

[Kaotik]

It's a GV100 SoC in 25-50W power envelope. There's just no way it would be on 16nm.

Assuming you meant transistorwise, GV100 is nearly 25% bigger.
Assuming you meant GV100+other stuff to make it SoC, no, it's not, not even close relative really.
But on "no way it would be on 16nm" you're definitely right. Samsung 8nm is quite likely candidate, unless they jumped to EUV with Samsung 7nm.
@pharma "pretty much given" means they don't know, but given the transistor and power budgets it's safe guess from their part. I wouldn't outrule 8nm Samsung, it's their closest match to 7nm when using DUV instead of EUV

 

[pharma]

pharma: I'd say it just means that they are guessing.

Is there a link for the Samsung 8nm speculation?

Edit: See some additional info may indicate 8 nm:

"Upon request, Nvidia has confirmed that Orin will be manufactured in the 8 nm process. Judging by the manufacturing processes currently available, the order should have gone to Samsung with 8LPP. TSMC does not offer a corresponding node."

https://www.computerbase.de/2019-12/nvidia-drive-agx-orin-gtc-china/

 

[no-X]

I don't think it's surprising. Or is it? There were some leaks (already in June) about using 8nm Samsung process for consumer Ampere GPUs.

 

[Deleted member 13524]

But on "no way it would be on 16nm" you're definitely right. Samsung 8nm is quite likely candidate, unless they jumped to EUV with Samsung 7nm.

Anandtech says the chip won't be ready for vehicles until 2022, so it's possible they don't even have it fabricated yet and 5nm is the target.

 

[pharma]

I don't think it's surprising. Or is it? There were some leaks (already in June) about using 8nm Samsung process for consumer Ampere GPUs.

Yes, it would be surprising since most leaks indicates Ampere using Samsung's 7nm process.(Digitimes).

 

[no-X]

According to the leaker, who was right about the existence of Hopper (mentioned it 6 months before other sources) and specifications of the Super-series, it should be 8nm Samsung for consumer GPUs and 7nm TSMC for HPC GPUs. He was semi-wrong abou AMD hardware, but his Nvidia-related leaks seems to be OK.

 

[Kaotik]

Anandtech says the chip won't be ready for vehicles until 2022, so it's possible they don't even have it fabricated yet and 5nm is the target.

As @pharma already edited to his post, NVIDIA has confirmed to ComputerBase that it's 8nm (which in turn confirms it's Samsung)

 

[Frenetic Pony]

According to the leaker, who was right about the existence of Hopper (mentioned it 6 months before other sources) and specifications of the Super-series, it should be 8nm Samsung for consumer GPUs and 7nm TSMC for HPC GPUs. He was semi-wrong abou AMD hardware, but his Nvidia-related leaks seems to be OK.

Huh, by all accounts Samsung's current processes can't hold up to TSMC's. Guess Nvidia got a good deal moneywise or something? A supported hypothesis, as Qualcomm has straight up said their new 865 Soc would do better on TSMC, while their mid range soc is on Samsung.

 

[CarstenS]

Huh, by all accounts Samsung's current processes can't hold up to TSMC's.

As you implied already, apart from the bleeding edge, processes suitability largely is a function of money.