NVIDIA ‘Ampere’ 8nm Graphics Cards

KillBoY_UK · 16 Jul 2020 at 15:24

five8five said:
It's a strange comparison though, eight racks versus one. The space required, power and cooling difference must be huge.

To be fare we are talking about a chip whos soul focus is AI while Ampere/A100 silicon can so a lot more

deuse · 16 Jul 2020 at 15:25

Gerard said:
How the hell do they go about making these and making them look real, just goes to show what kind of fakes could be making news.

Near sure this was already posted but it's impressive how well made it is.

Very

Wonder if they will go on ebay?

scubes · 16 Jul 2020 at 16:04

deuse said:
Very

Wonder if they will go on ebay?

M

Lmfao ******* over it.

Gerard · 16 Jul 2020 at 16:20

Seems a lot more sites are reporting on this supposed new power connector that ampere needs, looks as if there could be something to it after all. Can't even recall the last time a gpu came out and needed a new power connector.

K4Y84 · 16 Jul 2020 at 16:21

Gerard said:
Seems a lot more sites are reporting on this supposed new power connector that ampere needs, looks as if there could be something to it after all. Can't even recall the last time a gpu came out and needed a new power connector.

It's confirmed fake.

Quartz · 16 Jul 2020 at 16:23

K4Y84 said:
It's confirmed fake.

Where?

K4Y84 · 16 Jul 2020 at 16:23

Quartz said:
Where?

https://twitter.com/Eegras/status/1283502323893964800

melmac · 16 Jul 2020 at 16:41

Quartz said:
Looks like Ampere is getting a Colossal spanking in the AI space.

TNA said:
Might be worth looking into if it is possible to invest into the company. Could end up being very lucrative.

five8five said:
It's a strange comparison though, eight racks versus one. The space required, power and cooling difference must be huge.

It's a bizarre!! It's 8 racks vs 1 rack? AS fivevfive said, the power, space and cooling requirements would be huge compared to the DGX-A100. It would probably be more cost effective just to buy a second DGX-A100 (and better performance!!)

Quartz · 16 Jul 2020 at 17:15

K4Y84 said:
https://twitter.com/Eegras/status/1283502323893964800

Hahahahaha!

Quartz · 16 Jul 2020 at 17:16

melmac said:
It's a bizarre!! It's 8 racks vs 1 rack?

Huh? It's a 1U rack-mounted server.

melmac · 16 Jul 2020 at 17:20

Quartz said:
Huh? It's a 1U rack-mounted server.

In terms of performance metrics, Graphcore has compared eight M2000 IPU-Machine racks to a single DGX-A100

Seems to be 8 racks against 1.

TNA · 16 Jul 2020 at 17:29

melmac said:
It's a bizarre!! It's 8 racks vs 1 rack? AS fivevfive said, the power, space and cooling requirements would be huge compared to the DGX-A100. It would probably be more cost effective just to buy a second DGX-A100 (and better performance!!)

I just glanced over it so did not see that. I made the assumption that they are better as they are only dealing with AI like ASIC's great at only doing the one thing. Will look into it later.

melmac · 16 Jul 2020 at 17:39

TNA said:
I just glanced over it so did not see that. I made the assumption that they are better as they are only dealing with AI like ASIC's great at only doing the one thing. Will look into it later.

I thought that was the case, it was an easy detail to miss if you just did a quick glance at it.

Quartz · 16 Jul 2020 at 18:11

melmac said:
Seems to be 8 racks against 1.

No, it's 8 vs 8: 8x M2000 vs 8x A100. The DGX is a 4U part IIRC and merely the host box (granted it has an Epyc CPU inside). If you wanted to compare U vs U you would have 4x M2000 vs 8x A100. And the A100 would still get spanked.

Stu999 · 16 Jul 2020 at 18:16

K4Y84 said:
https://twitter.com/Eegras/status/1283502323893964800

First reply says other sources translate as 'I edited the above'. Either way a whole new connector isn't a way to go but IF it does happen hopefully third parties will use standard connectors.

melmac · 16 Jul 2020 at 18:26

Quartz said:
No, it's 8 vs 8: 8x M2000 vs 8x A100. The DGX is a 4U part IIRC and merely the host box (granted it has an Epyc CPU inside). If you wanted to compare U vs U you would have 4x M2000 vs 8x A100. And the A100 would still get spanked.

Here the relevant quotes, not sure how are making it out to be 8 vs 8.

In addition to the Colossus MK200 IPU, Graphcore is also unveiling its competitor to the NVIDIA DGX A100 rack codenamed the IPU-M2000. This rack is composed of four MK200 IPUs

In terms of performance metrics, Graphcore has compared eight M2000 IPU-Machine racks to a single DGX-A100. The reason is the performance per dollar metric for these comparisons. The DGX A100 costs $199,000 (MSRP) while eight M2000 racks would cost $259,600, MSRP)

Quartz · 16 Jul 2020 at 18:32

melmac said:
Here the relevant quotes, not sure how are making it out to be 8 vs 8.

The DGX contains 8 A100 chips.

EastCoastHandle · 16 Jul 2020 at 18:47

Rumor out there is that by the time the new cards hit the market it will make Turing doing ray tracing look like kepler doing dx12/vulkan.
The question is, since there aren't many RT games out there right now is will it matter? Could one live with just turning RT off in these upcoming games in that scenerio?

melmac · 16 Jul 2020 at 19:41

Quartz said:
The DGX contains 8 A100 chips.

And each IPU-M2000 contains 4 MK200 Chips.

Satchfanuk · 16 Jul 2020 at 20:14

EastCoastHandle said:
Rumor out there is that by the time the new cards hit the market it will make Turing doing ray tracing look like kepler doing dx12/vulkan.
The question is, since there aren't many RT games out there right now is will it matter? Could one live with just turning RT off in these upcoming games in that scenerio?

The problem as I see it is that, by bringing RT functionality so early and with such a heavy penalty to gamers, in both cost and the negative performance impact, nvidia could actually turn consumers off the idea of RT altogether.

It was so bad that nvidia also had to introduce dlss to mitigate the impact that RT had on framerates.

Hopefully the second coming will be much improved by both nvidia and amd.