NVIDIA Pascal Thread

[3DVagabond]

the first gen of those cards on the nodes will surely be more about power eff instead of perf but it would be stupid not to expect at least a 10-15% jump on perf from a flagship to another

also why this thread keep saying the cards will arrive on may? does OP have clairvoyance ?

Because a road map from a year ago resurfaced.

 

[mohit9206]

Any idea when entry level Pascal cards could arrive? I mean cards that cost $100-150 based on Pascal.

 

[airfathaaaaa]

You won't know until its in the hands of reviewers for testing. Since we obviously can't trust the specs released even if it's directly from NVIDIA, re: NV's 3.5 + 0.5GB and Async Compute claims.

But if they do have it (which I think Pascal will have hardware support), I would expect major Async Compute usage for new GameWorks features, it's gonna really reinforce the point for Kepler/Maxwell owners to upgrade to the latest & greatest.

I mean some folks here aren't happy with mid-range Pascal beating 980Ti/Titan X by ~20%, it could easily be 980Ti/Titan X +50% in DX12 if GameWorks pushes Async (which tanks on GPUs without support). Upgrade-worthy then? You bet.

mid 2015 they changed their roadmap to include volta and they put pascal almost on an identical plane with maxwell 2.0 that means their async capabilities will be close to none(since they wouldnt have the time to change their arch to include anything hardware wise)

 

[Silverforce11]

mid 2015 they changed their roadmap to include volta and they put pascal almost on an identical plane with maxwell 2.0 that means their async capabilities will be close to none(since they wouldnt have the time to change their arch to include anything hardware wise)

I'm hearing otherwise. Pascal is compute enhanced Maxwell 2 with limited multi-engine added. We'll see.

It would be a major blow to DX12/Vulkan adoption of Pascal is incapable of hardware async compute.

 

[Timmah!]

Because of transistor cost. The key advantage now is performance/watt.

But people dont buy cards for performance/watt. They buy for more performance. They care for perf/wattage ratio, only if they can have more performance at same wattage. Not equal performance at lower wattage. Why would someone waste another 400 USD/EUROs on 1070, if they already paid 700 for 980Ti. To save 100 Watts?

 

[ShintaiDK]

But people dont buy cards for performance/watt. They buy for more performance. They care for perf/wattage ratio, only if they can have more performance at same wattage. Not equal performance at lower wattage. Why would someone waste another 400 USD/EUROs on 1070, if they already paid 700 for 980Ti. To save 100 Watts?

They are not the target for such a product.

980TI/Fury/Fury X users isn't the target for an upgrade before 2017.

 

[Sweepr]

Because a road map from a year ago resurfaced.

Can you point us where Bench-life got this roadmap, or are you just trolling a NVIDIA-related thread as usual?

 

[3DVagabond]

But people dont buy cards for performance/watt. They buy for more performance. They care for perf/wattage ratio, only if they can have more performance at same wattage. Not equal performance at lower wattage. Why would someone waste another 400 USD/EUROs on 1070, if they already paid 700 for 980Ti. To save 100 Watts?

Not everyone owns a 980 ti. Some of us also have pretty random requirements. No PCIe connector, or maybe only one 6pin.

 

[dacostafilipe]

Can you point us where Bench-life got this roadmap, or are you just trolling a NVIDIA-related thread as usual?

Most articles around Pascal seem to be based on the the ascii.jp "roadmap" : https://translate.google.com/transl...n&u=http://ascii.jp/elem/000/001/109/1109096/

 

[3DVagabond]

Most articles around Pascal seem to be based on the the ascii.jp "roadmap" : https://translate.google.com/transl...n&u=http://ascii.jp/elem/000/001/109/1109096/

All of the Chinese writing on the lower right must be where it says it's by nVidia and gives the NDA date. /s

 

[ShintaiDK]

All of the Chinese writing on the lower right must be where it says it's by nVidia and gives the NDA date. /s

Its Japanese

 

[airfathaaaaa]

I'm hearing otherwise. Pascal is compute enhanced Maxwell 2 with limited multi-engine added. We'll see.

It would be a major blow to DX12/Vulkan adoption of Pascal is incapable of hardware async compute.

well i didnt say it was identical pascal basicly is a refresh towards computing(sp/dp) of maxwell and nothing more really

they dont have the time needed to actually remake the whole card to add anything more hardware wise and i hardly think it will be a blow since all of the big 5 companies are pushing for it if nvidia dont follow they will just get stuck behind unless ofc someone here thinks that nvidia has more influence on the matter any more than ms/amd/intel/samsung/apple/every arm company plus most of the big gaming houses that already said they will support it

 

[Timmah!]

They are not the target for such a product.

980TI/Fury/Fury X users isn't the target for an upgrade before 2017.

I doubt Nvidia would purposefuly ignore part of their customer base this year, because they plan them to be in for upgrade in 2017.

 

[NTMBK]

well i didnt say it was identical pascal basicly is a refresh towards computing(sp/dp) of maxwell and nothing more really

Be careful mixing those two roadmaps; they don't seem to refer to the same Maxwell. The first says that Maxwell would double DP/W over Kepler, which clearly didn't happen.

 

[airfathaaaaa]

the first one is from mid 2014 the second one is from 2015......clearly nothing nvidia say is to taken seriously untill further investigation

 

[ShintaiDK]

20nm is likely the answer. Or lack of same.

 

[Silverforce11]

well i didnt say it was identical pascal basicly is a refresh towards computing(sp/dp) of maxwell and nothing more really

That's the obvious conclusion, but I think NV has a ton of money to afford redesigns late into the cycle. It's only a question of money.

I am expecting Pascal with extra engines, ACEs (though obviously they won't call it such) if you will, perhaps limited in number and queue depth, but able to properly execute graphics + compute in parallel.

 

[airfathaaaaa]

That's the obvious conclusion, but I think NV has a ton of money to afford redesigns late into the cycle. It's only a question of money.

I am expecting Pascal with extra engines, ACEs (though obviously they won't call it such) if you will, perhaps limited in number and queue depth, but able to properly execute graphics + compute in parallel.

its not about money we arent talking about a minor redesign here it will need a major overhaul to actually move from a software sc to a hardware sc with ace like engines also having flipping cores and such this thing takes time to be a proof of concept let alone tape out test and such

i dont think that we have seen such a move in any industry so far
(obviously the money will be a factor but the time to adjust + the complete rewrite of drivers and find the correct design of the core to actually follow dx12 is a no go on this matter)

 

[poofyhairguy]

I doubt Nvidia would purposefuly ignore part of their customer base this year, because they plan them to be in for upgrade in 2017.

Why do you doubt that? In 2014 Nvidia released the 980, not the 980 TI, aka they didn't target 2013 780 ti customers then either. The 780 ti customer base got ignored.

And guess what- some 780 ti owners bought a 980. I bet someone out there then bought a 980 ti after that.

Nvidia knows what they are doing. Everyone is just upset that what Nvidia is doing (and has done) isn't what they want.

 

[tential]

Why do you doubt that? In 2014 Nvidia released the 980, not the 980 TI, aka they didn't target 2013 780 ti customers then either. The 780 ti customer base got ignored.

And guess what- some 780 ti owners bought a 980. I bet someone out there then bought a 980 ti after that.

Nvidia knows what they are doing. Everyone is just upset that what Nvidia is doing (and has done) isn't what they want.

I want to dispel the notion that Nvidia doesnt know what theyre doing when they release a midrange chip first.
They know exactly what they're doing.

 

[Silverforce11]

I want to dispel the notion that Nvidia doesnt know what theyre doing when they release a midrange chip first.
They know exactly what they're doing.

And what they are doing is very logical.

Especially on a new unproven node, going big first for a consumer SKU release is extremely risky.

What they did with Maxwell, they did not out of necessity or risk management, but purely because they saw that it worked and it's simply more profitable to milk small chips for max prices for as long as the competition allows it.

And with AMD being so public with their roadmap about Vega + HMB2 in 2017, NV knows full well they won't have any competition at the high-end, and so expect nothing but low/mid Pascal for this year.

 

[Timmah!]

Why do you doubt that? In 2014 Nvidia released the 980, not the 980 TI, aka they didn't target 2013 780 ti customers then either. The 780 ti customer base got ignored.

And guess what- some 780 ti owners bought a 980. I bet someone out there then bought a 980 ti after that.

Nvidia knows what they are doing. Everyone is just upset that what Nvidia is doing (and has done) isn't what they want.

They released Titan Black and Titan Z that year. Titan Z was obviously fail at that exorbitant price, but regardless of what you think of Titan Black, it was superior to 980 in many tasks thanks to bigger vram or double precision support.

Anyway, they offered top end product every year, whether it was gtx690, original Titan, etc...

I dont question they know what they are doing. Thats exactly why i doubt they wont offer upgrade path to 980Ti/Titan X owners this year.

 

[Silverforce11]

Thats exactly why i doubt they wont offer upgrade path to 980Ti/Titan X owners this year.

Have you considered at all, the possibility of what I mentioned here?

http://forums.anandtech.com/showpost.php?p=38112718&postcount=568

If Pascal has hardware Async Compute, you can bet NV is going to push it in GameWorks. That 20% lead for GP104 may well grow to 50% in DX12 titles. Is that a viable upgrade path, considering price & perf/w? Heck yes.

 

[Kris194]

But people dont buy cards for performance/watt. They buy for more performance. They care for perf/wattage ratio, only if they can have more performance at same wattage. Not equal performance at lower wattage.

Exactly, we don't need to lower power consumption, it's already very low and gamers (GP104 target) do not pay for lower power consumption, they pay for performance.

 

[Genx87]

There's also this very plausible theory that Maxwell only existed because of the cancellation of the 20nm node.

NV's prior road maps did not have Maxwell. It was a Kepler -> Pascal leap.

This theory propose that Maxwell is Pascal, with most of the FP64 stripped, mix-mode ops stripped, a purely gaming or FP32 focus design. This results in the nice gains versus Kepler, at the same 28nm, despite only small die size increase (GK104 -> GM204 etc), NV has extracted a lot of performance from it.

Now, if that theory is true, the real comparison for Pascal should be made versus Kepler.

We're talking GK104 vs GP104, and GK110 vs GP100. What kind of performance leap is that? It's a massive perf leap.

Maxwell was always on the road map. But I think you are right that the cancellation of 20nm changed Nvidias plans for what Maxwell was supposed to become. They stripped out parts of the original design as a stop gap to get through 28nm.