NVIDIA Pascal Thread

[kraatus77]

No it doesn't prove your point. That's a ROP limitation on FuryX causing the low memory bandwidth utilization.

GCNs 64 ROps are tapped out by Fiji's use of HBM memory in that test whereas Hawaii had its ROPs memory bandwidth limited with GDDR5.

Basically, Fiji has more than enough bandwidth to spare by using HBM but lacks ROPs to make any true use of the available memory bandwidth.

This allowed AMD to use the extra memory bandwidth for their 4GB framebuffer tunning. Transferring data from system memory to the framebuffer and back based on priority allowing Fiji the ability to run full speed even when a game went over the 4GB framebuffer. AMD have two engineers dedicated to that task. Rise of the Tomb Raider is an example of that fine tunning.

So no, AMD don't have weaker memory controllers, their memory controllers are superior to NVIDIAs. Heck, GCN's entire memory hierarchy, from cache to framebuffer is more robust than NVIDIAs.

What NVIDIA excelled at, especially with Maxwell, is color compression. This made up for NVIDIAs inferior memory controller's and cache hierarchy in Maxwell.

End result is what matters, 980/960 can deliver better/same performance using same/less memory bandwidth + color compression against 380x. that means they handling available bandwidth more efficiently.


you say amd has more efficient mc, than why the heck it's amd who needs more bandwidth to deliver same performance ? i mean nvidia are able too feed 64rops with 224gb/s meanwhile amd needs 320gb/s and you are telling me amd is more efficient ? lol. even 780ti is able to deliver same performance as 290x with less bandwidth and less rops. and without any color compression.

also look at 280x/380 for the matter. in 4k 280x is faster/same with it's more bandwidth and same rops. so that means 256bit is just enough to keep it's performance, on the other hand 960 is pretty close to 380 with just 128bit. and same 32rops.

you can eat 8lbs of meat and run 10km or eat 4lbs of meat and run 10km. it doesn't matter how much you really digest, what matters is how much you are able to run while eating less meat.

 

[Silverforce11]

You forget IPC improvements over Maxwell, and... GPC configurations. 1280 Pascal CUDA core GPU clocked at 1480 MHz with GDDR5X can achieve performance of GTX 970.

The question is: would it be able to maintain thermal envelope without power connector. That last bit is most questionable.

What makes you think GDDR5X is suitable for a high volume, low price SKU?

Come on man. Premium vram tech.

Even the GP104 rumors has been coming out that only the top uncut SKU is GDDR5X, the 970 replacement is GDDR5 only.

Now GP106 is already a 128 bit bus, we know this from the Drive leaks. It's also got GDDR5, ~80GB/s.

That isn't going to cut the mustard at getting 980 performance at 75W.

 

[IllogicalGlory]

http://videocardz.com/59154/nvidia-geforce-gtx-1080-pictured-again

That thing is seriously fugly. You're all crazy. Needs more polys.

 

[Silverforce11]

Once you pay for the interposer everything else on it is basically cheaper based on smaller die sizes. I'd guess an interposer won't come in at much over $20 but I haven't looked closely at it.

The interposer on Fiji was claimed to be $30 or so.

But they did not say the actual cost for assembly of the entire package with TSV and QC.

How much does Amkor or eSilicon charge for the process? We don't know. The wafer/interposer cost is low though.

 

[digitaldurandal]

End result is what matters, 980/960 can deliver better/same performance using same/less memory bandwidth + color compression against 380x. that means they handling available bandwidth more efficiently.


you say amd has more efficient mc, than why the heck it's amd who needs more bandwidth to deliver same performance ? i mean nvidia are able too feed 64rops with 224gb/s meanwhile amd needs 320gb/s and you are telling me amd is more efficient ? lol. even 780ti is able to deliver same performance as 290x with less bandwidth and less rops. and without any color compression.

also look at 280x/380 for the matter. in 4k 280x is faster/same with it's more bandwidth and same rops. so that means 256bit is just enough to keep it's performance, on the other hand 960 is pretty close to 380 with just 128bit. and same 32rops.

you can eat 8lbs of meat and run 10km or eat 4lbs of meat and run 10km. it doesn't matter how much you really digest, what matters is how much you are able to run while eating less meat.

What matters in the conversation is that you are saying one thing and meaning another. You need to express yourself correctly.

If people are discussing the efficiency of a single piece of hardware, in this case the memory controller, and you come in and make a comment on it based on the bigger picture, memory bandwidth utilization, then you are the one having a communication problem.

You can come back and say - "hey memory bandwidth utilization is what is important and fiji left a lot on the table." However, in the context of their conversation it was the controller efficiency that was key and other factors that influenced fiji's memory bandwidth utilization were not important in the context of the conversation.

 

[xorbe]

That thing is seriously fugly. You're all crazy. Needs more polys.

Yeah, the previous shroud had much better tessellation and smooth round corners.

 

[Arachnotronic]

That thing is seriously fugly. You're all crazy. Needs more polys.

Agree, it's pretty bad. But honestly don't care how it looks, the pair I buy will be shoved into a case that'll be placed next to my desk. I won't see them, but I hope to appreciate the FPS they pump out in my games

 

[Silverforce11]

I don't get why you guys thinking it's ugly. It's great, all those angular designs make it look like a sports car, very masculine, almost transformers-like. Aggressive and tech-cool.

On the design front, a lot of tech reviewers praised Fury for the industrial design, but I thought it was meh... very plain jane.

 

[Arachnotronic]

Yeah, the previous shroud had much better tessellation and smooth round corners.

Poor tesseleation performance in Pascal confirmed!!!1111

 

[antihelten]

I don't get why you guys thinking it's ugly. It's great, all those angular designs make it look like a sports car, very masculine, almost transformers-like. Aggressive and tech-cool.

And that is exactly why I think it's ugly (and your remaining descriptors are also pretty damn iffy in my book).

 

[AnandThenMan]

And that is exactly why I think it's ugly

Same I think it looks hideous. I normally don't care much at all about how a vid card actually looks but this thing hurts my eyes.

 

[Maverick177]

That thing is seriously fugly. You're all crazy. Needs more polys.

The shroud looks like one of those "Gamerz" Case or mice, tells a lot about nVidia's customer as well.

 

[SteveGrabowski]

The shroud can be pink and teal and covered in swastikas for all I care if it's a decent upgrade from the 980 Ti.

 

[Arachnotronic]

The shroud can be pink and teal and covered in swastikas for all I care if it's a decent upgrade from the 980 Ti.

I would never buy a product with swastikas on it. Would rather use an iGPU.

 

[Adored]

The swastika is thousands of years old, not 100.

 

[SteveGrabowski]

I would never buy a product with swastikas on it. Would rather use an iGPU.

It could have an ISIS flag for the backplate for all I care.

 

[JDG1980]

What makes you think GDDR5X is suitable for a high volume, low price SKU?

What makes you think it's not?

 

[Silverforce11]

What makes you think it's not?

What makes you think NV prefers lower margins on low-end SKUs?

GDDR5X for GP106 or GDDR5 would make more sense to you? ie, do you think GDDR5X is cheaper than GDDR5 & has more volume?

 

[JDG1980]

What makes you think NV prefers lower margins on low-end SKUs?

GDDR5X for GP106 or GDDR5 would make more sense to you?

A 128-bit GDDR5X bus may well be cheaper than a 256-bit GDDR5 bus with similar throughput. And GP106 can't use standard GDDR5 on a 128-bit bus or it will be badly bottlenecked. Even GM206 (GTX 960) is pushing it. They would either have to go up to at least a 192-bit bus, or go with GDDR5X. Compared to AMD, Nvidia prefers narrower buses and high-speed RAM.

Remember that Polaris 10, which is going to be in the same size class as GP106, will have a 256-bit bus.

 

[Silverforce11]

A 128-bit GDDR5X bus may well be cheaper than a 256-bit GDDR5 bus with similar throughput. And GP106 can't use standard GDDR5 on a 128-bit bus or it will be badly bottlenecked. Even GM206 (GTX 960) is pushing it. They would either have to go up to at least a 192-bit bus, or go with GDDR5X. Compared to AMD, Nvidia prefers narrower buses and high-speed RAM.

Remember that Polaris 10, which is going to be in the same size class as GP106, will have a 256-bit bus.

GM206 is 128bus with 7Gbps GDDR5. It's cheap and bandwidth starved, but it does the job fine! The 960 is very popular afterall.

Why would GP106 go with a bigger bus or GDDR5X? It adds to cost. Them margins must remain high or get better.

GP106 with 128bus and 8Gbps GDDR5 and better memory compression and improved register/cache on SM/CC could give you 30-40% more performance while still being cheap to produce.

The only scenario where you need GDDR5X on it, would be if the performance target is much higher.

All they have to deliver, is 970 performance class in a 960 replacement, at $200 and gamers would love it. The bar isn't that high.

 

[xorbe]

It could have an ISIS flag for the backplate for all I care.

Allahu Asyncbar!!!1! ~oh snap~

 

[RussianSensation]

The shroud looks like one of those "Gamerz" Case or mice, tells a lot about nVidia's customer as well.

The fact that so many people still pay $$$ for hot, loud, thermal-throttling, poor overclocking blower reference cards tells us even more, unless most of the market is using 3-4 cards in SLI and/or is water-cooling them. :sneaky:

Ironically, for 99% of PC builders with proper modern cases, blowers are a pretty poor choice.

"We found that with the default settings on GeForce GTX 980 SLI the lowest clock rate it hit while in-game was 1126MHz. That clock speed is actually below the boost clock of 1216MHz for GTX 980. This is the first time we've seen the real-time in-game clock speed clock throttle below the boost clock in SLI in games. It seems GTX 980 SLI is clock throttling in SLI on reference video cards. This is something we did NOT expect, but it is happening with reference cards."
http://www.hardocp.com/article/2014...980_sli_overclocked_gpu_review/2#.VxcWG9UrKHs

Open-air cooled cards. 0 dBA idle, quiet at load, cool, cheaper NV PCB/VRM goes out the window, upgraded VRMs/power circuitry, almost always lower noise levels stock and esp. overclocked, etc. :thumbsup:
https://www.techpowerup.com/reviews/Gigabyte/GTX_980_Ti_XtremeGaming/22.html

Anyway, the card is ugly, far worse than the current version imo, but not that it matters since when it's upside down, all you'll see is the back-plate (or not).

There is a reason all the top overclocking 980Ti NV cards (Galaxy HOF, Asus Matrix, EVGA Classified, MSI Lightning, etc.) are using non-reference PCBs/no blowers.

 

[Cookie Monster]

What were they thinking?? Looks hideous imo.

Sometimes I feel like Im the only person who favors say motherboards with normal heatsinks vs those massive chunky colorful things and just no frills mechanical/heatsink designs tailored for practicality rather than aesthetics..

 

[Silverforce11]

What were they thinking?? Looks hideous imo.

Sometimes I feel like Im the only person who favors say motherboards with normal heatsinks vs those massive chunky colorful things and just no frills mechanical/heatsink designs tailored for practicality rather than aesthetics..

You mean fancy heatsinks on RAM that serves no function besides looking cool? lol

It's a pet peeve of mine too. But it sells.

 

[nvgpu]

http://videocardz.com/59189/nvidia-geforce-gtx-10801070-cooler-shroud-taken-apart