Nvidia GTC16 Webcast Live Stream@5pm/9am PDT *Jen looks great in leather.*

[easyrider]

The number of times he repeated/started to repeat himself caught my attention - something clearly not right but might be personal not business.

I thought the whole thing was shoddy and amateurish.

 

[Rroff]

Someone on AT forums made this observation:



GCN and Pascal might be more a like than we think!!

Shoots down those people claiming vehemently that Pascal was just a shrunk Maxwell and would have "no" Async heh. Those changes definitely aren't by mistake.

 

[drunkenmaster]

FP32 isn't that bad - sure the hardware units hasn't gone up much but the clock speeds are a fair jump - the actual GFLOP increase isn't to be sniffed at.

The CUDA core per SM is an interesting change for those who've been keeping up

Keep in mind that 580gtx was 1.6TF or so and Titan was 4.5TF, Titan X was what 6.1TF and GP100 appears to be 10.5TF was it and with an increase in power to 300W. So 40 to 28nm was a nearly 3 fold increase in FP32 performance, from Titan X to GP100 is a significantly, dramatically smaller increase. If it had the same increase as last process node change it would have about 18TF of performance, instead it has 10.5TF, I think it's fair enough to say that is pretty tame.

The clock speed jump only makes that worse, it's 300W at very high clocks, at 250W with lower clocks it will be even less, and that listed performance is at full boost clocks not base clocks.

 

[JediFragger]

From the obvious changes they've made it certainly looks to be an interesting architecture (I was in the camp of thinking it would be v similar to Maxwell).

If it is indeed close to GCN in design then it would certainly make some sense...

 

[blasterbolt]

And the wait continues tor palace my 980s for 1440p!

 

[CAT-THE-FIFTH]

Shoots down those people claiming vehemently that Pascal was just a shrunk Maxwell and would have "no" Async heh. Those changes definitely aren't by mistake.

Yeah,I suppose it will be interesting to see how much functionality they will back into hardware like AMD has done more recently. Plus,it does kind of hint that Maxwell has some shortcomings too.

But I kind of sad yonks ago,that I see both companies kind of converging in some ways but from different angles.

 

[drunkenmaster]

Shoots down those people claiming vehemently that Pascal was just a shrunk Maxwell and would have "no" Async heh. Those changes definitely aren't by mistake.

The changes are very much about deep learning and HPC, it might still badly struggle to switch between graphics and compute functions easily. Less cores per SM make a lot of sense when you're trying to push throughout of FP16 deep learning stuff where you want to push through more smaller simpler blocks more quickly.

We'll also see if these changes carry over to the gaming/smaller cores. The FP64 is utterly worthless for gamers, the FP16 could be useful but it's in there as a versatile cuda core architecture so if they wanted to or found a way to get rid of FP64 it may remove FP16 as well.

For gaming a Maxwell shrunk for GP104 might be better for gaming performance and die size than just cutting down a GP100 into a GP104. I don't know if Nvidia will do that this generation, I see a huge increase in FP64 as bad in every way for gaming and eventually I can see AMD/Nvidia creating separate compute and gaming architectures, just not sure it will be this generation.

 

[D.P.]

Technically we should be comparing it to the Kepler based GK110 and GK210 cards and not the GM200 since Maxwell had most of DP compute stripped out.

So did the Fiji GPUs.

 

[D.P.]

I think it indicates that Nvidia are quite worried about the Intel compute cards,whch are slowly creeping into the supercomputer market.

This is definitely their biggest concern and is where they are putting most of their money in the actual discrete GPU development.

 

[bru]

Keep in mind that 580gtx was 1.6TF or so and Titan was 4.5TF, Titan X was what 6.1TF and GP100 appears to be 10.5TF was it and with an increase in power to 300W. So 40 to 28nm was a nearly 3 fold increase in FP32 performance, from Titan X to GP100 is a significantly, dramatically smaller increase. If it had the same increase as last process node change it would have about 18TF of performance, instead it has 10.5TF, I think it's fair enough to say that is pretty tame.

The clock speed jump only makes that worse, it's 300W at very high clocks, at 250W with lower clocks it will be even less, and that listed performance is at full boost clocks not base clocks.

Only fair to compare tesla versus Tesla, so that makes the X2090 (580) 1.3TF and the Tesla K20X (original Titan) 3.9TF and the Tesla M40 (TitanX) 5.8TF.
But even with those numbers I do agree with you the numbers for the P100 do look pretty tame, especially considering the clock speeds it is running at.

 

[meldarthx]

DM I doubt we'll see AMD split their cards - I mean cut down on FP 64 - but totally gut? doubt it.

I've been saying this for months and people wouldn't listen on Pascal. Everyone in the industry has come out and stated Nvidia is 6+ months behind AMD; part of that is due to the process they are on 16ff+

We'll hear about small Pascal in June - possibly paper launch in July; but won't see cards until at earliest Sept; if not Oct. Big Pascal won't be seen until next year. Nvidia has never; ever.............EVER been good at node shrinks - suddenly people think they will be ahead of AMD? Originally - Nvidia announced their would only be updates to their mobile chips - aka maxwell again. Then suddenly that's changed to they will have small Pascal chips out for back to school.

Honestly I don't know what to believe - it might be possible they'll get some small pascal mobile out - because what I'm hearing is; Nvidia is worried about Polaris full stack coming out for back to school; and getting a lot of contracts over them; which I can easily see happening.

Nvidia's also terrified of AMD and Intel continue to gain in pro market; Intel is going for full throttle - and AMD isn't backing down - Cuda still has a grip; but its not as massive as it was people think - pro market actually wants more players there.......this is going to be interesting. Nvidia's bread and butter - *though that think its gaming are very wrong; its pro market that pays the bills*

 

[Razor Time]

because what I'm hearing is; Nvidia is worried about Polaris full stack coming out for back to school; and getting a lot of contracts over them; which I can easily see happening.

I hope so, competition is always good in a two horse race and we all know nvidia like their prices high and advancements slow if they could have their own way.

 

[ICDP]

So did the Fiji GPUs.

What the hell has that got to do with the topic being discussed? It is not even remotely a green vs red thing going on here so don't turn it into one.

 

[mahius]

For those who missed it live (like myself), you can watch it here. Jen seems to like Deep-learning. It's like NV's equivalent of Asynchronous compute.

Edit: Turns out that was last years. Damn... I missed it. Thankfully the Titan X was very early so I knew something was up when they started announcing a card that already exists. Doesn't seem to be any videos of this year's currently, I've read an article about it though and it's disappointing. No mention of a Pascal card. Whatever happened to announcing a great GPU every year?

Meh, I called it anyway though. I think they're saving it for last. That and there's the possibility that it won't release this year.

 

[r7slayer]

I cba watching all that babbiling on he keeps doing and where he keeps keeps repeating the same, repeating the same thing over and over again.

Is there a news website that has taking all the important things from the video and condensed it down yet?

 

[bru]

For those who missed it live (like myself), you can watch it here. Jen seems to like Deep-learning. It's like NV's equivalent of Asynchronous compute.

ER that was last years event.

 

[humbug]

I cba watching all that babbiling on he keeps doing and where he keeps keeps repeating the same, repeating the same thing over and over again.

Is there a news website that has taking all the important things from the video and condensed it down yet?

Sure, i can do that;
Driver-less Cars, powerful Work Station Cards, Cars that drive themselves, Autonomous Cars, Compute Cards, Multi Lightprobe VR, learning, the Cars of Jeremy Clarkson nightmares, GPU's that go to School and Cars that don't need You.

 

[Lokken86]

Wow this thread moved along quick what did I miss.

 

[JBuk]

Sure, i can do that;
Driver-less Cars, powerful Work Station Cards, Cars that drive themselves, Autonomous Cars, Compute Cards, Multi Lightprobe VR, learning, the Cars of Jeremy Clarkson nightmares, GPU's that go to School and Cars that don't need You.

So...was there any mention of anything car related ?

 

[D.P.]

What the hell has that got to do with the topic being discussed? It is not even remotely a green vs red thing going on here so don't turn it into one.

It's very relevant. Both companies cut compute and both companies have to add it back in.The kind of improvements we see in pascal will be mirrored in Vega