Nvidia announces Tesla P100 GPU with Pascal architecture

[Xlucine]

Nvlink 2 (nvilink ti?) will be released just in time for this all to flare up again

[lumireleon]

hehe so far the Hexus post with the longest arguments *comments*

[DanceswithUnix]

hehe so far the Hexus post with the longest arguments *comments*

No it isn't

[HalloweenJack]

micron have only just sampled GDDR5X to IHV`s, with an announced volume production `in the summer`. - unless GTX1080 is using GDDR5 we wont see it IMO before september

[scaryjim]

OMG is it impossible for you to agree to disagree?

OMG is it impossible for you to admit that what you originally said wasn't what you actually meant?

If as you say "It must have a connection" show me where on the P100 board the end-user would make that connection, show me where the end-user would join, link, or fasten together the P100 card to a PCIe connection.

See, now you're talking about connectors, but saying connection. This is the problem we're having.

As it is, the end user would use the mezzanine connector on the P100 to make the PCIe connection to the motherboard. Like I've already said in this thread.

They might do that directly, if they managed to source a motherboard with the appropriate mezzanine connector for PCIe. They might use a specially designed adapter board. They might buy a PCIe x16 flat-cable riser card, separate all the wires and manually connect them to the appropriate pins in the mezzanine connector. How they do it doesn't matter. What matters is that they're making a connection. The precise physical form of the connection is irrelevant. Whether it uses an edge connector, a mezzanine connector, or a soldered wire, it's still a connection.

And before you attempt your usual straw man argument as you've been doing repeatedly throughout this thread you may want to re-read what i originally posted, not what you have repeatedly been quoting out of context.

Sure, I'll do exactly that. Here's where we started, with two comments from you about PCIe/NVLink:

True but even if you could buy one you couldn't use it as you'd need a MoBo with a NVLink connection, apparently Nvidia says the PCI Express bus doesn't provide enough bandwidth to keep the card busy.

Note - "you'd need a motherboard with an NVLink connection"

I get the feeling though that the diagram you/they showed is like some of their other claims, that it will eventually come with a PCIe connection (that it's capable of using PCIe) but the first cards are only going to come with NVLink, i get the feeling that they're fabricating (and initially selling) large silicon and binning ones that don't meet make the grade for a lesser (lower power draw) HPC card that will use PCIe.

Note - "it will eventually come with a PCIe connection ... but the first cards are only going to come with NVlink"

So that's what you actually said. That's where all this started. Not a single mention of connectors. Two assertions that the P100 was NVLink only. And that's all the context any of us had when this started. You've since told us that what you meant was that they didn't have PCIe edge connectors that you could slot into a standard motherboard, and that statement is perfectly fine, but it's not what you said originally. Connection is a generic term. All connectors form connections, but not all connections use connectors. So saying something doesn't have a connector is a very different statement from saying it doesn't have a connection.

So, essentially, I can't "agree to disagree" because we don't actually disagree. You're simply using different language to describe it than pretty much any native English-speaker would. When you say "it doesn't have a connection", everyone else thinks that means something different from what you're trying to say. That's your perogative of course, but it's going to start - and prolong - a lot of arguments.

[Corky34]

See, now you're talking about connectors, but saying connection. This is the problem we're having.

What exactly do you think you do with a connector? You use it to make a connection.

As it is, the end user would use the mezzanine connector on the P100 to make the PCIe connection to the motherboard. Like I've already said in this thread.

Then show me where on a motherboard that the end user would connect the P100 to make use of that PCIe connection.

They might do that directly, if they managed to source a motherboard with the appropriate mezzanine connector for PCIe. They might use a specially designed adapter board. They might buy a PCIe x16 flat-cable riser card, separate all the wires and manually connect them to the appropriate pins in the mezzanine connector. How they do it doesn't matter. What matters is that they're making a connection. The precise physical form of the connection is irrelevant. Whether it uses an edge connector, a mezzanine connector, or a soldered wire, it's still a connection.

Then show me this motherboard with the appropriate mezzanine connector for PCIe, show me a specially designed adapter board that isn't Nvidia's hybrid cube mesh, show me the PCIe x16 flat-cable riser card that connects to the NVLink mezzanine connector, show me an example of a P100 card running where someone has separated all the wires and manually connect them to the appropriate pins, how the connections made is the only thing that's important because as I've been saying for all intents and purposes the end-user can't make a PCIe connection, if as you suggest the end-user can make a PCIe connection to the P100 card then show me how, show me how an end-user would connect a P100 card to a motherboard with a PCIe connection.

Note - "you'd need a motherboard with an NVLink connection"

Then show me the motherboard with an NVLink connection

So, essentially, I can't "agree to disagree" because we don't actually disagree. You're simply using different language to describe it than pretty much any native English-speaker would. When you say "it doesn't have a connection", everyone else thinks that means something different from what you're trying to say. That's your perogative of course, but it's going to start - and prolong - a lot of arguments.

And if I'm using different language to describe it than pretty much any native English-speaker would then perhaps you need to write a strongly worded letter to the OED as they say a connection is: The placing of parts of an electric circuit in contact so that a current may flow.

[scaryjim]

... The placing of parts of an electric circuit in contact so that a current may flow.

*sigh* yes, exactly. And the P100 has parts that you can place in contact so that a current may flow. So it has a connection. See. We're in agreement. Perhaps we can stop this now?

EDIT: sorry, I should probably leave this, but I've only just been struck by the hypocrisy of someone accusing me of setting up straw men then saying this:

... Then show me this motherboard with the appropriate mezzanine connector for PCIe, show me a specially designed adapter board that isn't Nvidia's hybrid cube mesh, show me the PCIe x16 flat-cable riser card that connects to the NVLink mezzanine connector, show me an example of a P100 card running where someone has separated all the wires and manually connect them to the appropriate pins, how the connections made is the only thing that's important because as I've been saying for all intents and purposes the end-user can't make a PCIe connection, if as you suggest the end-user can make a PCIe connection to the P100 card then show me how, show me how an end-user would connect a P100 card to a motherboard with a PCIe connection. ...

[DanceswithUnix]

Supposedly P100 has 60 shader modules, of which 4 are disabled.

http://www.fudzilla.com/news/graphic...56-sms-enabled

[scaryjim]

Supposedly P100 has 60 shader modules, of which 4 are disabled.

http://www.fudzilla.com/news/graphic...56-sms-enabled

Fermi mk 2? Got to wonder why nvidia keep trying to do these great big dies first up on a new process, when they never get them to yield well fully-enabled. Wonder when GP110 will turn up....

[shaithis]

Fermi mk 2? Got to wonder why nvidia keep trying to do these great big dies first up on a new process, when they never get them to yield well fully-enabled. Wonder when GP110 will turn up....

Probably so they don't need to redesign once the process has matured and they are ready to release the full-fat version.

Plus, I guess that ANY 4 units can be disabled, meaning that any card coming off the line with 60 (of 64) working units can be sold, reducing waste.

[DanceswithUnix]

Probably so they don't need to redesign once the process has matured and they are ready to release the full-fat version.

Plus, I guess that ANY 4 units can be disabled, meaning that any card coming off the line with 60 (of 64) working units can be sold, reducing waste.

I expect so. On the GTX480 the worst performing SMs were disabled, dead ones followed by most power hungry allowing the remaining ones to clock faster with the freed power budget so it isn't so bad.

For such a massive chip on a still fairly young process, I think only 4 disabled sounds rather good.

[scaryjim]

Probably so they don't need to redesign once the process has matured and they are ready to release the full-fat version.

*shrug* perhaps; or they felt the need to have the biggest and best compute card off the lines as quickly as possible. When they went 28nm they did the GK104 first, got excellent yields, and then scaled up. I just wonder what's got them desperate to have the big die out first this time. It makes a lot of sense to run a smaller die first on a new process - you get to find out what the kinks in the process are on something that costs you a lot less money if the yields aren't where you want them to be, and you can apply those lessons to the larger die later (it's what AMD did at 40nm very successfully).

I mean, you don't design a die specifically to not use 7% of the functional components, so presumably the idea was to release the 60 SM version but yields weren't good enough. leaves you wondering if they could've refined their designs for 16nm by doing a smaller die first, so when they released the P100 it could've had the full complement of shaders enabled...

EDIT for crosspost:

... For such a massive chip on a still fairly young process, I think only 4 disabled sounds rather good.

GF100 had 1 of 16 clusters disabled to make the GTX480. That's actually a slightly lower percentage, and why I drew the Fermi comparison in the first place. FI suppose the architecture and DP performance make it clear that GP100 is a compute and HPC chip first and foremost, so perhaps part of the reason it's not yielding well enough for a full implementation is that the silicon is more complex? *shrug*

Just feels a lot like history repeating...

[DanceswithUnix]

I just wonder what's got them desperate to have the big die out first this time.

Well they are contracted to build those big supercomputers, so if the sales are guaranteed then you might as well crack on with building them!

[scaryjim]

Well they are contracted to build those big supercomputers, so if the sales are guaranteed then you might as well crack on with building them!

From memory they're not due until late 2017, are they? Perhaps that's the thing then - get GP100 out early so they've got time to do GP110 for the real money-spinner...!

[Corky34]

*sigh* yes, exactly. And the P100 has parts that you can place in contact so that a current may flow. So it has a connection. See. We're in agreement. Perhaps we can stop this now?

So basically what you're saying is that after almost 70 posts you agree with my original statement "that it will eventually come with a PCIe connection (that it's capable of using PCIe)"? That eventually the P100 or more likely the P50, will have the parts needed so the end-user can make an electric circuit in contact so that a current may flow using only PCIe.

EDIT: sorry, I should probably leave this, but I've only just been struck by the hypocrisy of someone accusing me of setting up straw men then saying this:

There's nothing straw man about that, you claimed that someone could connect to the PCIe signal on a P100 card if only they could source a motherboard with an appropriate mezzanine connector for PCIe, or adapter board, or a PCIe x16 flat-cable riser card, separate all the wires and manually connect them, but unless you can show me a motherboard or adapter board that converts the NVLink connector to a PCIe connector they can't as the P100 card only come with NVLink connectors.

It's like saying you could plug a TV that you bought in France into a UK socket when hypothetically there's no French sockets in your house, you hypothetically don't have an adapter, and hypothetically you're not capable or rewiring the plug.

[DanceswithUnix]

, and hypothetically you're not capable or rewiring the plug.

That bit doesn't work, not with the crowd around here