NVIDIA Pascal Thread

[antihelten]

This road map. Perf/watt over time.

doh!, I don't know how I forgot about that one.

Based upon that graph it would look like Vega has ~50% better efficiency than Polaris (and thus 3.75 times the efficiency of 28nm GPUs).

Of course an increase in efficiency doesn't necessarily mean an increase in performance/IPC, but then again that was actually what we saw with the Kepler to Maxwell transition. Maxwell had roughly a 40-45% increase in efficiency (compared to Titan), along with roughly a 25% increase in IPC per core. If the Polaris to Vega transition is the same, then we might also be looking at roughly a 25% increase in IPC. This, along with the 60% increase in shader, would lead to a total improvement of more or less 100% (this all assumes that Vega and Polaris has the same clocks of course).

Combined with the numbers from my previous post, this would put Vega about 30-35% ahead of GP104. There's no way Nvidia would be able to compete with this without a new GPU.

Of course this assumes that the Polaris to Vega transition would be similar to the Kepler to Maxwell transition, which is very speculative to put it mildly.

The 2.5x already seems wrong since the CEO wont commit to it with investors. And there is no numbering on the left. So you dont even know if the scaling is supposed to be linear either.

Doing a bit of pixel counting reveals that the Polaris box is sitting almost exactly 2.5 times higher than the 28nm GPU box, relative to the x axis. So it would appear that the graph is actually linear, but of course it's still just a PR slide, so grain of salt and all that.

 

[3DVagabond]

doh!, I don't know how I forgot about that one.

Based upon that graph it would look like Vega has ~50% better efficiency than Polaris (and thus 3.75 times the efficiency of 28nm GPUs).

Of course an increase in efficiency doesn't necessarily mean an increase in performance/IPC, but then again that was actually what we saw with the Kepler to Maxwell transition. Maxwell had roughly a 40-45% increase in efficiency (compared to Titan), along with roughly a 25% increase in IPC per core. If the Polaris to Vega transition is the same, then we might also be looking at roughly a 25% increase in IPC. This, along with the 60% increase in shader, would lead to a total improvement of more or less 100% (this all assumes that Vega and Polaris has the same clocks of course).

Combined with the numbers from my previous post, this would put Vega about 30-35% ahead of GP104. There's no way Nvidia would be able to compete with this without a new GPU.

Of course this assumes that the Polaris to Vega transition would be similar to the Kepler to Maxwell transition, which is very speculative to put it mildly.

Yeah. I guess my point was that we don't really know. We at least need to see Polaris and Pascal 1st. Then we can guess about Vega. But from that road map, Vega is not simply a scaled up Polaris.

 

[Olecki]

tell me a company that actually had profits by selling fewer parts but with more cost per unit...name ONE(and survived to tell the story)
indeed it was that 1%....meanwhile any survey you can see on the net says a totally different story
lol just lol

I can give you example: Apple. They could sell iPhones at 30% of current pricing (with same sales volume) and they still would be profitable. Margin on iPhones is really ridiculous but peple are buying their products anyway.
Same is true for all companies positioning their products as a luxury.

 

[R0H1T]

I can give you example: Apple. They could sell iPhones at 30% of current pricing (with same sales volume) and they still would be profitable. Margin on iPhones is really ridiculous but peple are buying their products anyway.
Same is true for all companies positioning their products as a luxury.

How does your reply answer the poster you quoted

He said something along the lines of ~ Apple selling their current, or upcoming iphones, at 130% value & something like half the volumes (just some random numbers like yours) & how that will not end well for Apple. This is all true of course because in a declining market the last thing you do is to raise prices, that's like committing suicide!

 

[Olecki]

This is just a business calculation about margin to sells volume. Let's say that NV earning 80$ for each sold 1070 at 400$ . If they deicide to sell it for 500$ they will be earning 180$, even if they will sell 50% cards overal profit will be higher. You need only calculate where is best compromise between cost/sales volume.

 

[xpea]

raja said 2.5x and he is the head of rtg and also an engieneer...his words have more value

Lisa is also an engineer but more important, during the financial Q&A she was acting as the CEO of the company, replying to the investors. She can't lie or she can go to jail.
Raja can say any bullshit, it won't have any consequence. Some already have forgotten the Overclocker's dream told by the lead engineer...

 

[maddie]

The 2.5x already seems wrong since the CEO wont commit to it with investors. And there is no numbering on the left. So you dont even know if the scaling is supposed to be linear either.

Read the slide man. Your monomaniacal thinking blinds you at times.

1x perf/watt 28nm and 2.5x perf/watt Polaris.

 

[maddie]

Yeah. I guess my point was that we don't really know. We at least need to see Polaris and Pascal 1st. Then we can guess about Vega. But from that road map, Vega is not simply a scaled up Polaris.

This roadmap shows Polaris, Vega and Navi architectures, implying major changes. I think the patents recently being revealed here will be rolled out over several years and not all at once with Polaris.

 

[Osjur]

Lisa is also an engineer but more important, during the financial Q&A she was acting as the CEO of the company, replying to the investors. She can't lie or she can go to jail.
Raja can say any bullshit, it won't have any consequence. Some already have forgotten the Overclocker's dream told by the lead engineer...

Yes but that 2x p/w could mean it was measured over Fiji and Raja's 2.5x p/w over Tonga for example. We don't really know.

Also who are you referring to with that lead engineer statement... because I'm pretty sure the guy who actually said that wasn't the lead engineer developing Fiji chips.

 

[antihelten]

This is all true of course because in a declining market the last thing you do is to raise prices, that's like committing suicide!

No this is not true at all. Point in case, Nvidia has been raising prices in a declining market, and has posted record quarter after record quarter.

 

[R0H1T]

No this is not true at all. Point in case, Nvidia has been raising prices in a declining market, and has posted record quarter after record quarter.

That's because the ASP is increasing isn't it, that's cause the IGP's have eaten up the low end. They've also taken away marketshare from AMD, so whilst strictly speaking IMO the statement is correct but it does come with certain qualifiers. For instance the smartphone market, the example that was given, is seeing ASP decline & Apple is also looking to lower their costs, not margins though, with the likes of SE.

The dGPU market is bifurcated between OEM & retail, direct to consumer, so it's not exactly the same as it'd be in case of an OEM or retail only market. Then you HPC which is the bread & butter for Nvidia, so now you have Apples & Oranges plus strawberries, figuratively speaking

The point is, at least what I'm saying, when you had lots of alternatives people could go to Matrox, 3dfx, ATI et al. However this isn't the case anymore, so it's now just a matter of which brand has more loyal followers i.e. Nvidia, you can see that with how the 970 fiasco didn't blow up in their face & in fact many buyers upgraded to 980. This isn't the case with smartphones, though there are still lots of Apple fans, & so you see a lot of churning happening over there.

The free market & all the "perfect" capitalism theories don't work these days, as they're supposed to, but that's something much deeper & maleficent I'm afraid.

 

[ShintaiDK]

He said something along the lines of ~ Apple selling their current, or upcoming iphones, at 130% value & something like half the volumes (just some random numbers like yours) & how that will not end well for Apple. This is all true of course because in a declining market the last thing you do is to raise prices, that's like committing suicide!

There are 2 things. Market share in the current market and market share total volume. Both mixed with the volume/profit ratio.

You see Apple today with a phone like the iPhone SE at 400 and 500$. Much cheaper than previous. However its done to capture more market share. Its not going to reverse the decline of phones.

With GPUs, they have IGPs to compete with. So you have to include IGPs when you think total share. Hence its not an issue as such with GPUs moving up in segment. Because they cant compete with IGPs else. Also why you only see 2 out of 20 or so SKUs in the M400 series as new GPUs.

And we dont exactly lack confirmation that people want to move up, for those that use it. Record K sales, record GTX sales. That some with entitlement issues may feel left out is another matter.

 

[Silverforce11]

So we're supposed to be 1 month from a Pascal release according to all these rumors.

Why is it we're getting board pics of QA GP104 samples, and the most recent, GDDR5X on full GP104 board, which is also a sampling chip from Micron?

Unless they are keeping leaks very tight..

~1 month before a retail release is when AIBs already have stocks with volume, moving to the docks and ports for shipping global. This means a lot of people would have their hands on it, and you can no longer stop detailed leaks from coming out.

 

[ShintaiDK]

So we're supposed to be 1 month from a Pascal release according to all these rumors.

Why is it we're getting board pics of QA GP104 samples, and the most recent, GDDR5X on full GP104 board, which is also a sample from Micron?

Unless they are keeping leaks very tight..

~1 month before a retail release is when AIBs have already started shipping in volume. This means a lot of people would have their hands on it, and you can no longer stop detailed leaks from coming out.

Why haven't we seen any Polaris boards? Its supposed to release in the same time frame. We know a lot more about Pascal and have seen a lot more than Polaris.

 

[Silverforce11]

Why haven't we seen any Polaris boards? Its supposed to release in the same time frame. We know a lot more about Pascal and have seen a lot more than Polaris.

Polaris is due in June.

The rumors are that Pascal will beat it to the release in May, no?

Basically, if we're a month from release and there's no detailed leaks, it's either a paper launch (press seeding/briefing, no retail til later), the rumors are wrong, or this is the most secretive mission impossible release ever.

 

[antihelten]

That's because the ASP is increasing isn't it, that's cause the IGP's have eaten up the low end. They've also taken away marketshare from AMD, so whilst strictly speaking IMO the statement is correct but it does come with certain qualifiers. For instance the smartphone market, the example that was given, is seeing ASP decline & Apple is also looking to lower their costs, not margins though, with the likes of SE.

The dGPU market is bifurcated between OEM & retail, direct to consumer, so it's not exactly the same as it'd be in case of an OEM or retail only market. Then you HPC which is the bread & butter for Nvidia, so now you have Apples & Oranges plus strawberries, figuratively speaking

The point is, at least what I'm saying, when you had lots of alternatives people could go to Matrox, 3dfx, ATI et al. However this isn't the case anymore, so it's now just a matter of which brand has more loyal followers i.e. Nvidia, you can see that with how the 970 fiasco didn't blow up in their face & in fact many buyers upgraded to 980. This isn't the case with smartphones, though there are still lots of Apple fans, & so you see a lot of churning happening over there.

The free market & all the "perfect" capitalism theories don't work these days, as they're supposed to, but that's something much deeper & maleficent I'm afraid.

The ASP has gone up because Nvidia has started selling mainstream GPUs for enthusiast prices, you know the thing that half this thread has been complaining about. And so what if they've taken away market share from AMD, doesn't change the fact that their strategy of increasing prices in a declining market has lead to record earnings.

Point is that airfathaaaaa couldn't comprehend how a company could go for higher margin and low volume vs lower margin and high volume, and still remain profitable, even though there are tons of examples of companies doing so and being very successful (Nvidia themselves being one of them).

At the end of the day it's simply about finding the optimal inflection point between price and volume (volume being driven by demand at the given price), all companies try to do this and Nvidia's strategy of raising prices has clearly succeeded and lead to higher total profits.

 

[ShintaiDK]

Polaris is due in June.

The rumors are that Pascal will beat it to the release in May, no?

Basically, if we're a month from release and there's no detailed leaks, it's either a paper launch (press seeding/briefing, no retail til later), the rumors are wrong, or this is the most secretive mission impossible release ever.

By your logic both Polaris and Pascal will be paper launched.

 

[antihelten]

So we're supposed to be 1 month from a Pascal release according to all these rumors.

Why is it we're getting board pics of QA GP104 samples, and the most recent, GDDR5X on full GP104 board, which is also a sampling chip from Micron?

Unless they are keeping leaks very tight..

~1 month before a retail release is when AIBs already have stocks with volume, moving to the docks and ports for shipping global. This means a lot of people would have their hands on it, and you can no longer stop detailed leaks from coming out.

For what it's worth the leaked pics of the two GP104 samples both had the date stamp as 1614, or in other words the 14th week of 2016, which would be just 2 weeks ago.

So the samples might only have gotten into the hands of OEMs very recently, which might help explain the lack of leaks so far.

 

[Timmah!]

If you wanted to do some performance guesses, could you for simplicity´sake use the P100 FP32 peak performance number in GFlops and pretend this number is a result of its CC number (3584) multiplied by their boost clock frequency? Then do the same for Titan X and then do some comparisons? Or this is not how it works at all, so any attempt to do it this way would prove futile?

 

[Silverforce11]

By your logic both Polaris and Pascal will be paper launched.

How? May and June aren't the same time. Seriously.

 

[Silverforce11]

For what it's worth the leaked pics of the two GP104 samples both had the date stamp as 1614, or in other words the 14th week of 2016, which would be just 2 weeks ago.

So the samples might only have gotten into the hands of OEMs very recently, which might help explain the lack of leaks so far.

Yes, which leads to the other point, these are sampling chips for QA. Not production chips, for both Pascal & GDDR5X (which isn't due for mass production until Q3).

If they intend a retail launch in May, and they only just got QA samples now?

Think back to the Bitsandchip info leak from earlier in the year. They nailed GP100 timing/leaks months in advanced. What did they say about GP104?

Release around Computex with more retail availability in Q3.

 

[tviceman]

So we're supposed to be 1 month from a Pascal release according to all these rumors.

Why is it we're getting board pics of QA GP104 samples, and the most recent, GDDR5X on full GP104 board, which is also a sampling chip from Micron?

Unless they are keeping leaks very tight..

~1 month before a retail release is when AIBs already have stocks with volume, moving to the docks and ports for shipping global. This means a lot of people would have their hands on it, and you can no longer stop detailed leaks from coming out.

GK104 and GM204 performance didn't leak until days before release. But hey, Nvidia hasn't demonstrated working silicon so they're obviously months behind AMD, right?!

 

[ShintaiDK]

If you wanted to do some performance guesses, could you for simplicity´sake use the P100 FP32 peak performance number in GFlops and pretend this number is a result of its CC number (3584) multiplied by their boost clock frequency? Then do the same for Titan X and then do some comparisons? Or this is not how it works at all, so any attempt to do it this way would prove futile?

Nope. Its theoretical numbers only.

 

[ShintaiDK]

How? May and June aren't the same time. Seriously.

Tell me when Pascal launches and tell me when Polaris launches. Exact dates. And then tell me when we should see all kinds of Polaris leaks.

 

[Timmah!]

Nope. Its theoretical numbers only.

OK, but they still dont pull that theoretical number out of their ..... They gotta have some way to calculate it. But you say its not cc number x frequency?