Next NVIDIA Line will be named "Hopper"

[Nicnac]

^{This leak is a few days old by now but I didn't see a forum post yet so I hope this hasn't been posted already ^^}

 

NVIDIA has registered a new Trademark:

Quote

NVIDIA's architectures are always based on computer pioneers and this one appears to be no different. Nvidia's Hopper architecture is based on Grace Hopper, who was one of the pioneers of computer science and one of the first programmers of Harvard Mark 1, and also the inventor of the first linkers. She also popularized the idea of machine-independent programming languages, which led to the development of COBOL - an early high-level programming language still in use today. She enlisted in the Navy and helped the American War efforts during World War II.

 

They also registered the name "Aerial" for what we don't know.

 

Speculations say it is possibly going to utilize an MCM (multi-chip-module) philosophy. That would be NVIDIA Glue^TM then I guess...

(hey glueing stuff together wasn't such a bad idea after all huh?)

 

Be aware that this is still a lot of speculation and Hopper might be a line that might not even come to consumer hardware.

 

I think an MCM design is the logical next step and there are many analogies to how AMD is handling their CPU lines these days.

 

Sources: Videocardz and the Salt Mine

 

Edit:

There are a few more fun names on the list ^^

 

OmNiVeRsE might be my favorite

[Bombastinator]

There may already be a thread in this in news.  I recall seeing that name and picture previously.  There may be new information in this post though.  Not sure.

[Nicnac]

Just now, Bombastinator said:

There may already be a thread in this in news.  I recall seeing that name and picture previously.  There may be new information in this post though.  Not sure.

I checked until december 4th. if its earlier than that it might just be the rumour and not the registered trademark?

I guess mods will decide ^^

[Bombastinator]

3 minutes ago, Nicnac said:

I checked until december 4th. if its earlier than that it might just be the rumour and not the registered trademark?

I guess mods will decide ^^

It would have been before that. Late November possibly.

[Fasauceome]

35 minutes ago, Nicnac said:

hey glueing stuff together wasn't such a bad idea after all huh?

In terms of AMD CPU cores, hell yes

 

In terms of the FE cooler, hell no

[LeSheen]

She looks very happy. Would feel right at home with the modern it-crowd 

[Princess Luna]

nVidia OMNIVERSE.... okay that one made me giggle xD'

[rcmaehl]

1 hour ago, LeSheen said:

She looks very happy. Would feel right at home with the modern it-crowd 

Have you tried turning it off and back on again?

[Nicnac]

1 hour ago, Princess Luna said:

nVidia OMNIVERSE.... okay that one made me giggle xD'

I'd prefer a Multiverse that includes AMD

[Mira Yurizaki]

From https://www.tweaktown.com/news/69186/nvidia-next-gen-hopper-gpu-arrives-ampere-smash-intel-amd/index.html

 

 

This to me looks more like "SLI on a package" than say what AMD is doing with its recent processors.

[porina]

12 minutes ago, Mira Yurizaki said:

This to me looks more like "SLI on a package" than say what AMD is doing with its recent processors.

I've long wondered, what would it take to make multiple GPUs look and feel like a single one? Probably some combination of bandwidth and latency.

 

PCIe 3.0 is roughly 1 GB/s per lane, so up to 16GB/s bandwidth on HEDT platforms with sufficient lanes, half that on consumer more likely on 8x+8x. Double for PCIe 4.0. 

Nvidia SLI bridge depends on configuration but is ball park GB/s, so less than PCIe, but is dedicated to the cards. PCIe also has to deal with loading stuff from system ram for example.

Nvidia nvlink, for RTX cards, from memory is 50GB/s. A good step up in bandwidth, but put in context, still far below vram bandwdith. Higher end ones are >500GB/s.

 

If the GPUlets are on the same package, and have some kind of high bandwidth, low latency connection between them, maybe you can get it to look more like a single GPU. 

 

BTW the first version of Intel's new GPU is also a chiplet-like arrangement. https://www.anandtech.com/show/15119/intels-xe-for-hpc-ponte-vecchio-with-chiplets-emib-and-foveros-on-7nm-coming-2021

[Mira Yurizaki]

18 minutes ago, porina said:

I've long wondered, what would it take to make multiple GPUs look and feel like a single one? Probably some combination of bandwidth and latency.

Something that looks like what AMD is doing with Zen 2's Threadripper and EPYC, then do something like checkboard rendering such that each GPU chiplet renders say every nth pixel or block of the screen to avoid the problem of uneven workloads.

 

The lingering concern I had was memory access can be indeterminate, but I think that's not really the case with GPUs, considering their workload can be determined ahead of time. So as long as the drivers can schedule memory access such that by the time the GPU chiplets are ready to work on the next thing, they'll have what they need, then the system should work.

 

The issue to overcome here is that current multi-GPU contain the same dataset in each of their VRAM pools. If you eliminate that, then I see an MCM-based GPU more plausible. Like if we take an RTX 2080 Ti with its 6 GPCs and make each GPC a chiplet, it won't necessarily need more bandwidth than what it normally uses if you can find a way to make them all use the same VRAM pool

[porina]

13 minutes ago, Mira Yurizaki said:

Something that looks like what AMD is doing with Zen 2's Threadripper and EPYC, then do something like checkboard rendering such that each GPU chiplet renders say every nth pixel or block of the screen to avoid the problem of uneven workloads.

Chiplet + IO arrangement would be particularly challenging for a GPU though. IF bandwidth is very low in comparison to GPU bandwidths. There will be major challenges in moving that much data around. Coherent NUMA-like approach is probably technically a lot easier to manage, providing inter-note bandwidth is high enough.

 

 

13 minutes ago, Mira Yurizaki said:

The issue to overcome here is that current multi-GPU contain the same dataset in each of their VRAM pools.

I'm not sure how it works, but when RTX was launched, the claims made by Jensen at the time in many presentations implied you can connect two cards by nvlink and get effective use of the combined vram. As said, the consumer nvlink I think is around 50MB/s bandwidth (comparable to dual channel 3200 ram). I think the professional nvlink is even higher.

[Mira Yurizaki]

4 minutes ago, porina said:

Chiplet + IO arrangement would be particularly challenging for a GPU though. IF bandwidth is very low in comparison to GPU bandwidths. There will be major challenges in moving that much data around. Coherent NUMA-like approach is probably technically a lot easier to manage, providing inter-note bandwidth is high enough.

My assumption is a smaller GPU means less bandwidth is needed to feed it. Something like an RTX 2080 Ti only needs 600+ GB/s of bandwidth because of all the execution units it has to feed. If you break them up and have them separated, presumably you don't need anywhere near that much per chiplet, but combined yes.

 

EPYC has to feed what is essentially 8 Ryzen 7 dies with its I/O chip. You can do the same technique with GPUs to get a combined VRAM pool rather than try to make separate GPU modules that work together in an internal SLI fashion. AMD solved the NUMA issue by making memory go through one source.

[porina]

7 minutes ago, Mira Yurizaki said:

EPYC has to feed what is essentially 8 Ryzen 7 dies with its I/O chip. You can do the same technique with GPUs to get a combined VRAM pool rather than try to make separate GPU modules that work together in an internal SLI fashion. AMD solved the NUMA issue by making memory go through one source.

Don't know what ram speeds Epyc support, but let's assume it is running 8ch ram at 3200. I make that ball park 200 GB/s ram bandwidth going through IOD... so I guess in that sense, the 600GB+ isn't much further of a stretch.

 

Still, from what's been shown so far, it doesn't look like either nvidia or Intel are going this route with their modular GPUs.

[Dylanc1500]

@porina

NVlink 2.0 has a transmission rate of 25 GT/s per lane. Each sub link has 8 lanes, with single and dual sub-links on the RTX line 800GT/s or 100GB/s for dual sub-links. The V100 Mezzanine version having 6 sub-links totaling 2500 GT/s or 300GB/s.

 

Don't get me started on the NVswitch.

[Bombastinator]

So much terminology.  There’s “gigabyte” gigabit” and “gt” whatever that is.  They’re all different and don’t have a whole lot to do with each other.  I personally find “gigabyte” to be the least useful since people haven’t really used 8 bit bytes since cp/m

[Tamesh16]

9 hours ago, Nicnac said:

OmNiVeRsE

This is what comes to mind immediately  Omnibot !

[The1Dickens]

9 minutes ago, Bombastinator said:

and “gt” whatever that is.

Gigatexels. From Wiki:

Quote

in case of texture fillrate the number of texture map elements (texels) GPU can map to pixels in a second.

 

 

5 hours ago, porina said:

I've long wondered, what would it take to make multiple GPUs look and feel like a single one? Probably some combination of bandwidth and latency.

 

Didn't AMD have dual GPUs on a single board some time ago? Some information that seems to pop out in my head is R9 or Fury dual cards... That might be me remembering incorrectly, though.

 

And I'm guessing Ampere will still run on PCIe 3.0? So then Hopper will theoretically run on 4.0?

4 hours ago, Mira Yurizaki said:

You can do the same technique with GPUs to get a combined VRAM pool rather than try to make separate GPU modules that work together in an internal SLI fashion.

Following this thought train, does this mean that it would then be easier to use multiple cards in a system (in a system that would benefit) for even greater VRAM pool?

[Bombastinator]

6 minutes ago, The1Dickens said:

Gigatexels. From Wiki:

 

Didn't AMD have dual GPUs on a single board some time ago? Some information that seems to pop out in my head is R9 or Fury dual cards... That might be me remembering incorrectly, though.

 

And I'm guessing Ampere will still run on PCIe 3.0? So then Hopper will theoretically run on 4.0?

Following this thought train, does this mean that it would then be easier to use multiple cards in a system (in a system that would benefit) for even greater VRAM pool?

https://en.m.wikipedia.org/wiki/Texel_(graphics)

So it’s a semi BS term when describing data rates because the number of 1s and 0s can vary by how the texels are made.  It goes in the pile along with “gigabyte” then.

[Nowak]

"NVIDIA's architectures are always based on computer pioneers and this one appears to be no different."

Didn't they used to name them after scientists? Like Johannes Kepler, Nikola Tesla, Blaise Pascal and Enrico Fermi?

[Drak3]

13 minutes ago, The1Dickens said:

Didn't AMD have dual GPUs on a single board some time ago?

Yup, and they’re making custom ones for Apple now too.

Nvidia also made dual GPU cards years ago.

[porina]

29 minutes ago, Bombastinator said:

I personally find “gigabyte” to be the least useful since people haven’t really used 8 bit bytes since cp/m

8 bits make a byte for as long as I've been into computing. You thinking of words? https://en.wikipedia.org/wiki/Word_(computer_architecture)

 

5 minutes ago, The1Dickens said:

Gigatexels.

GigaTransfers sprang to mind without looking it up. Maybe I should click that link...

 

5 minutes ago, The1Dickens said:

Didn't AMD have dual GPUs on a single board some time ago? Some information that seems to pop out in my head is R9 or Fury dual cards... That might be me remembering incorrectly, though.

They did, but I believe they were effectively crossfire on a card.

 

5 minutes ago, The1Dickens said:

And I'm guessing Ampere will still run on PCIe 3.0? So then Hopper will theoretically run on 4.0?

Following this thought train, does this mean that it would then be easier to use multiple cards in a system (in a system that would benefit) for even greater VRAM pool?

IMO even with both cards on 16x 4.0 the bandwidth is on the low side for two cards to share. 32GB/s is less than nvlink, which is in addition to PCIe.

[Bombastinator]

14 minutes ago, porina said:

8 bits make a byte for as long as I've been into computing. You thinking of words? https://en.wikipedia.org/wiki/Word_(computer_architecture)

 

GigaTransfers sprang to mind without looking it up. Maybe I should click that link...

 

They did, but I believe they were effectively crossfire on a card.

 

IMO even with both cards on 16x 4.0 the bandwidth is on the low side for two cards to share. 32GB/s is less than nvlink, which is in addition to PCIe.

Exactly.  A quasi arbitrary number only used in computers for a brief time.  It’s more accurate than gigatexel which appears to be only very specifically useful, just as bytes were.

 

i have the suspicion that gigtexel came out of marketing from one company or another and was used when they started to have less bandwidth than their competitors and they wanted a way to point out “yeah, maybe, but our texels are smaller than yours!”  If they actually wanted to convey information it would have been defined as a larger than one modifier for each instance in each brand.

[leadeater]

5 hours ago, Mira Yurizaki said:

then do something like checkboard rendering such that each GPU chiplet renders say every nth pixel or block of the screen to avoid the problem of uneven workloads

The problem with that which also effects SLI/Corssfire is it breaks post processing effects, certain lighting techniques and a few other things. That's why for a long time now SLI/Crossfire used AFR as it is the most compatible and also the simplest to implement.