Vega Rumors

[Araxie]

No, I think in some productivity benchmarks it beat the 1080 Ti handily. The 1080 Ti had a nice lead in CAD from what I remember. I don't feel like looking it up again. I did link it some pages back.

In a few games that HardwareUnBoxed ran yesterday in their 32 game shootout, the Vega 64 watercooled edition was nipping at its heels. I think Warhammer 40,000 Dawn of War III was one of them. Maybe a 5 frame difference between the Vega 64 and 1080 Ti.

But otherwise the 1080 Ti was kicking ass and chewing bubblegum.

and yet the Vega 64 WC cost the same as the 1080Ti.. an AIB 1080Ti destroy anything remotely of what VEGA could reach near 1080Ti FE.

Also of course there are some games that favor one vendor or another.. one single example of a popular game even old that favor Nvidia it's Fallout 4, there's a grotesque difference between Nvidia and AMD to the point that VEGA 64 Air perform at GTX 1070 FE Levels. however I try to not take those kind of games that favor one vendor or another in consideration unless someone it's looking to play specifically those kind of tittles.

Vega is not so bad at all, leaving Power consumption to a side I like Vega, in fact probably as I always do I may buy a Vega 64 Air cooled AIB later as I did with my 280X and 390X just to put my two cents in AMD account and play a bit with new toys, lol the 390X served me very well as an excellent back up card or testing card until some nasty driver fucked it up. I like to support some AMD cards others just nope (Fury Series).

 

[razor1]

Someone please post evidence that Vega is a good computer card compared to the 1080 and 1080Ti

Its all over the map sometimes it does better in compute, sometimes it does much worse, it all depends on the software. If the 1080 or 1080ti were forced to do Open Cl, they will have a tough time against most GCN cards even Polaris.

 

[Presbytier]

Its all over the map sometimes it does better in compute, sometimes it does much worse, it all depends on the software. If the 1080 or 1080ti were forced to do Open Cl, they will have a time time against most GCN cards even Polaris.

I think this is why NVIDIA decided to send a magical driver to the Titan Xp and unlock all of its compute functionality.

 

[razor1]

I think this is why NVIDIA decided to send a magical driver to the Titan Xp and unlock all of its compute functionality.

Well once Titan Xp got those drivers it takes a commanding lead in many of the tests. The Vega FE really competes with is the Quadro p5000, which is gp 104.

Purely drivers and software that are making such a difference. I'm sure if developers really program their way around GCN's through put issues, yeah it should beat Pascal hands down, but the amount of time to do that, probably not worth the dev's time nor is it worth the life cycle of the software.

 

[Presbytier]

Well once Titan Xp got those drivers it takes a commanding lead in many of the tests. The Vega FE really competes with is the Quadro p5000, which is gp 104.

Purely drivers and software that are making such a difference. I'm sure if developers really program their way around GCN's through put issues, yeah it should beat Pascal hands down, but the amount of time to do that, probably not worth the dev's time nor is it worth the life cycle of the software.

Not to mention the fact you would be alienating the vast majority of PC gamers who run NVIDIA.

 

[ChosenUno]

Yet here we are with a plethora of performance enhancing driver features now known to not be currently enabled and Vega64 is already ahead of 1080ti in some tests. Wishful thinking was assuming a superior architecture on paper wouldn't pass the inferior one when fully enabled. The real question is how far beyond 1080ti it goes. Already there are Nvidia pundits pushing the coming of Volta, with it's architecture that looks like GCN. Not surprising with SM6.0 based around GCN2. Nvidia are probably desperate to unload Pascal's while they still can, then have owners upgrade to Volta just to support everything within a year.

I've seen people being banned for writing stuff like this aka sine wave

 

[cybereality]

AMD also pioneered the new low-level graphics APIs with Mantle back in 2013. They later donated the API to Khronos and it became Vulkan.

DirectX 12 was probably in development during the same time, but I can't help but think that the early wins with Mantle and Battlefield paved the way for both DX12 and Vulkan becoming what they are.

AMD also has been more supportive of Linux, and the open-source AMD drivers are way ahead of the open-source Nvidia video drivers on Linux.

While I've traditionally been a big Windows buff, I'm starting to feel strained by all the proprietary strong-arm tactics some companies use, and it's nice to see a big player contributing to the open-source community.

 

[razor1]

AMD also pioneered the new low-level graphics APIs with Mantle back in 2013. They later donated the API to Khronos and it became Vulkan.

DirectX 12 was probably in development during the same time, but I can't help but think that the early wins with Mantle and Battlefield paved the way for both DX12 and Vulkan becoming what they are.

AMD also has been more supportive of Linux, and the open-source AMD drivers are way ahead of the open-source Nvidia video drivers on Linux.

While I've traditionally been a big Windows buff, I'm starting to feel strained by all the proprietary strong-arm tactics some companies use, and it's nice to see a big player contributing to the open-source community.

They didn't pioneer low level API's, LLAPI's have been there for many years prior to Mantle in consoles. There was a reason HLAPI's were created to begin with, they are easier to work with. HAL has been a developers friend for many many years and still is. Without HAL the changes from DX's versions all to 9 would have been nightmarish. Even going from 9 to 11, is problematic with deprecated code. This is why LLAPI's weren't done in the past for PC's. With GPU's becoming programmable to a point they can be used as general purpose processors, that changed all that, but expectedly different hardware is going to need different tuning. Natural course when working close to metal. Always has been that way the closer you are to the silicon.

And this actually opened a can of worms, cause now there is a much bigger divide between code written between the two IHV's. LLAPI's like DX12 and Vulkan there are no more HAL's which means future generations of hardware and games, devs will have to take that into consideration, they won't be able to make a game that will work as effectively on lets say a 5 year old graphics card even if its the same DX version! They might not optimize for older hardware. So if AMD goes away from GCN or consoles switch to a different architecture (next generation now scorpio or the new ps4). What happens then? Those new Asic's are going to be different and they will need different optimizations for LLAPI games too.

Open source vs. closed shouldn't matter to the end user, what matters is if it functions the way it should and performs the way it should at the end. We can go back and forth with the pro's and cons of open sourced vs closed source but both have their strengths and weakness. At the end the only thing that matters is end results. Does it do what the end user needs it to do?

 

[ChosenUno]

They didn't pioneer low level API's, LLAPI's have been there for many years prior to Mantle in consoles. There was a reason HLAPI's were created to begin with, they are easier to work with. HAL has been a developers friend for many many years and still is. Without HAL the changes from DX's versions all to 9 would have been nightmarish. Even going from 9 to 11, is problematic with deprecated code. This is why LLAPI's weren't done in the past for PC's. With GPU's becoming programmable to a point they can be used as general purpose processors, that changed all that, but expectedly different hardware is going to need different tuning. Natural course when working close to metal. Always has been that way the closer you are to the silicon.

And this actually opened a can of worms, cause now there is a much bigger divide between code written between the two IHV's. LLAPI's like DX12 and Vulkan there are no more HAL's which means future generations of hardware and games, devs will have to take that into consideration, they won't be able to make a game that will work as effectively on lets say a 5 year old graphics card even if its the same DX version! They might not optimize for older hardware. So if AMD goes away from GCN or consoles switch to a different architecture (next generation now scorpio or the new ps4). What happens then? Those new Asic's are going to be different and they will need different optimizations for LLAPI games too.

Open source vs. closed shouldn't matter to the end user, what matters is if it functions the way it should and performs the way it should at the end. We can go back and forth with the pro's and cons of open sourced vs closed source but both have their strengths and weakness. At the end the only thing that maters is end results. Does it do what the end user needs it to do?

The one with the larger marketshare usually wins with LLAPIs. That's just the nature of businesses.

So really, the "win" for AMD is not really a win if you think about it long term. They have to have marketshare to stay relevant in devs' eyes, and to have marketshare they have to be relevant to devs.

If Navi doesn't turn things around for them, LLAPIs might be the cause for their marketshare to go into free fall, just because there's no incentive for devs to wait around until AMD extract all the performance via drivers. And with Nvidia being the dev environment for the majority of devs these days it's not gonna get better for AMD if they just maintain the status quo, which they haven't.

 

[_mockingbird]

Right, but that means the GPU architecture was a dumpster fire right from ground zero.

GCN was great back in the day, but now it's time to move on.

 

[juanrga]

I think that is the crux of this. If the performance was there I don't think anyone would be too concerned with power consumption. We might still be arguing that NVIDIA is more efficient per watt, but at the end of the day, the GPU king is still the GPU king.

Performance = Efficiency * Power

Increasing efficiency is one way to increase performance. In fact it is the main way to increase performance, because the history of computing is the history of packing more performance in smaller sizes to get more performance.

 

[juanrga]

AMD has innovated, and led the pack. These are some examples:
AMD was the first with x64 for CPU

Only because Intel used the transition to 64 bits to start a new ISA without the weirdness and limits of x86-32. Yes, they failed, but at least gave a try.

On the other hand, AMD did something trivial by adding 64bit extensions on top of 32bit. Thanks to AMD the current x64 is the more bloated and weird ISA in use and x64 chips are inefficient, power hungry, costly, and buggy.

There are persistent rumors that Intel is trying to clean the x86 ISA again, eliminating the weirdest and oldest legacy stuff from 64bits.

Threadripper is far better than anything Intel curently offers in the consumer space anywhere close to the same price

ThreadRipper is only 10--20% faster on non-AVX workloads that scale up to the 16 cores

On everything else TR is worse. And that is comparing 16C stock vs 10C stock. Things go worse for TR thanks to higher overclocks on the i9.

Ryzen is far more power efficient than similarly configured Intel

What does "similarly configured" mean?

AMD 290x was better than Nvidia at the time

Sure?

AMD GPUs hit above their weight class in new dx12 and Vulcan now.

Nvidia beating AMD on Vulkan benchmarks, check page 2 and further

http://www.phoronix.com/scan.php?page=article&item=kblcpus-gl-vlk&num=2

AMD did with freesync in free software drivers what Nvidia did with gsync but without the $200-$300 hardware.

And each approach has its strengths and weaknesses.

Crossfire works/scales better than SLI - has for the last several generations.

Did you miss how AMD currently remains silent about Crossfire?

AMD GPU has been better at compute for several generations -- at most any given price point.

On raw compute numbers on selected OGL benches. Once we add efficiency to the equation, and CUDA then things change. Did you check the top500 list? Last time I checked it only a pair of supercomputers used AMD GPUs, everything else used Nvidia GPUs and Xeon Phi to crush numbers.

 

[juanrga]

AMD also has been more supportive of Linux, and the open-source AMD drivers are way ahead of the open-source Nvidia video drivers on Linux.

While I've traditionally been a big Windows buff, I'm starting to feel strained by all the proprietary strong-arm tactics some companies use, and it's nice to see a big player contributing to the open-source community.

Linux users claim otherwise. Developers claim otherwise...

 

[Ieldra]

Ah yes it's already faster than a 1080Ti, and the magic driver hasn't even landed yet. I'm a real boy!

 

[razor1]

Performance = Efficiency * Power

Increasing efficiency is one way to increase performance. In fact it is the main way to increase performance, because the history of computing is the history of packing more performance in smaller sizes to get more performance.

Exactly, performance efficiency and power are all tied together.

 

[razor1]

The one with the larger marketshare usually wins with LLAPIs. That's just the nature of businesses.

So really, the "win" for AMD is not really a win if you think about it long term. They have to have marketshare to stay relevant in devs' eyes, and to have marketshare they have to be relevant to devs.

If Navi doesn't turn things around for them, LLAPIs might be the cause for their marketshare to go into free fall, just because there's no incentive for devs to wait around until AMD extract all the performance via drivers. And with Nvidia being the dev environment for the majority of devs these days it's not gonna get better for AMD if they just maintain the status quo, which they haven't.

yeah market share dictates what Dev's will push their efforts on. And if AMD doesn't turn it around with competing hardware they will lose Xbox and Playstation deals, its just the normal course of events. MS already decided not to "upgrade" their console to a new generation of GCN, they decided to make changes on their own without AMD's involvement. Sony did go to Polaris like GCN. But its just a matter of time when power consumption, to keep those chips cool, and to get more performance out of them in a small form factor, the cost of cooling outweighs the low cost AMD is will to give.

Consoles have to stick with GDDR memory too, they aren't going to HBM, so really stuck there as well.

 

[Hameeeedo]

Yet here we are with a plethora of performance enhancing driver features now known to not be currently enabled and Vega64 is already ahead of 1080ti in some tests. Wishful thinking was assuming a superior architecture on paper wouldn't pass the inferior one when fully enabled. The real question is how far beyond 1080ti it goes. Already there are Nvidia pundits pushing the coming of Volta, with it's architecture that looks like GCN. Not surprising with SM6.0 based around GCN2. Nvidia are probably desperate to unload Pascal's while they still can, then have owners upgrade to Volta just to support everything within a year.

AMD is always behind NV innovation wise: Memory compression? Tiled Rasterization? Distributed Tessellation? High clocks? Power efficiency? These are the essence of any GPU, and in none of them is AMD even remotely competitive. Come back when they are. What does AMD have again? Yeah Async Compute and FP16! Console catering features.

Since the R600 days AMD NEVER took the lead from NVIDIA single GPU wise:
HD 2900X massively slower than 8800Ultra
HD 3870 massively slower than GTX 9800
HD HD4870/4890 slower than GTX 280/285
HD5870 slower than GTX 480
HD 6970 slower than GTX580
HD 7970 slower than GTX 680
HD 290X slower than GTX 780Ti/Titan Black
FuryX slower than 980Ti
Vega 64 massively slower than 1080Ti/TitanXp

See the pattern? with Vega AMD has now come full circle to being massively behind. And no amount of driver enhancements will do anything to dissolve this NVIDIA lead, They are just behind like always, It's just worse this time.

 

[Hameeeedo]

AMD also pioneered the new low-level graphics APIs with Mantle back in 2013. They later donated the API to Khronos and it became Vulkan.

DirectX 12 was probably in development during the same time, but I can't help but think that the early wins with Mantle and Battlefield paved the way for both DX12 and Vulkan becoming what they are.

AMD didn't donate Mantle, they were forced to after it's failure, it had loads of bugs, didn't work well on new GCN hardware, needed constant driver involvement, and always came in late after months of game launch. The original plan was to have 15 Mantle games, it only got 7. The rest of the games didn't even bother and got released with DX11 (not even Vulkan or DX12).

DX12 is also on a similar trajectory, 2 years now and only 13 games, no games in sight other than Forza 7 for a long time. Developers actually hate it, they still see DX11 as the more reliable option, the only games that have DX12 are those that AMD pays for them, or those that Microsoft forces DX2 upon them. The rest of the industry just ignores DX12. And Vulkan, same thing. Very few games get released on PC with Vulkan (3?), and AMD pays for those too.

 

[Archaea]

Only because Intel used the transition to 64 bits to start a new ISA without the weirdness and limits of x86-32. Yes, they failed, but at least gave a try.

On the other hand, AMD did something trivial by adding 64bit extensions on top of 32bit. Thanks to AMD the current x64 is the more bloated and weird ISA in use and x64 chips are inefficient, power hungry, costly, and buggy.

There are persistent rumors that Intel is trying to clean the x86 ISA again, eliminating the weirdest and oldest legacy stuff from 64bits.



ThreadRipper is only 10--20% faster on non-AVX workloads that scale up to the 16 cores

On everything else TR is worse. And that is comparing 16C stock vs 10C stock. Things go worse for TR thanks to higher overclocks on the i9.



What does "similarly configured" mean?



Sure?



Nvidia beating AMD on Vulkan benchmarks, check page 2 and further

http://www.phoronix.com/scan.php?page=article&item=kblcpus-gl-vlk&num=2



And each approach has its strengths and weaknesses.



Did you miss how AMD currently remains silent about Crossfire?



On raw compute numbers on selected OGL benches. Once we add efficiency to the equation, and CUDA then things change. Did you check the top500 list? Last time I checked it only a pair of supercomputers used AMD GPUs, everything else used Nvidia GPUs and Xeon Phi to crush numbers.

Your post is disingenuous. I stand by everything I said.

 

[Archaea]

AMD is always behind NV innovation wise: Memory compression? Tiled Rasterization? Distributed Tessellation? High clocks? Power efficiency? These are the essence of any GPU, and in none of them is AMD even remotely competitive. Come back when they are. What does AMD have again? Yeah Async Compute and FP16! Console catering features.

Since the R600 days AMD NEVER took the lead from NVIDIA single GPU wise:
HD 2900X massively slower than 8800Ultra
HD 3870 massively slower than GTX 9800
HD HD4870/4890 slower than GTX 280/285
HD5870 slower than GTX 480
HD 6970 slower than GTX580
HD 7970 slower than GTX 680
HD 290X slower than GTX 780Ti/Titan Black
FuryX slower than 980Ti
Vega 64 massively slower than 1080Ti/TitanXp

See the pattern? with Vega AMD has now come full circle to being massively behind. And no amount of driver enhancements will do anything to dissolve this NVIDIA lead, They are just behind like always, It's just worse this time.

Here I'll rebutal your post with my own quoted post from another thread that fits in exactly to the conversation. Raw compute power still goes to AMD cards.

I don't agree with this.

as far as computational power AMD has had the edge for a very long time.

1080TI - tflops = 11.3
Vega - tflops = 12.6

980TI - tflops = 5.6
Fury X - tflops = 8.7

780TI - tflops = 5.0
290x - tflops = 5.6

680 - tflops = 3.1
HD7970 - tflops = 3.8

580 - tflops = 1.5
HD6970 - tflops = 2.7

480 - tflops = 1.3
HD5870 = 2.7

285 - tflops = .7
HD4870 = 1.2


sources:
https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units

It might go back further than this - I didn't search further back -- but to this point - for the modern GPU era -- you can see in every generation going back a decade that AMD had stronger hardware. As such, then they must always have had weaker driver support --- at least in regards to pulling out all the raw performance potential, this since nVidia cards typically had the edge on actual in-game performance.

That's why AMD cards have traditionally been better at password cracking, computational workload, or raw performance at mining - because the compute power is there. NVidia has a stronger driver team and has been able to make a higher performing gaming card with less computer power.

I think this hardware capability disparity also plays into the Fine Wine theme. AMD can make improvements over a longer period of time, because they left so much on the table to begin with.

 

[Presbytier]

Here I'll rebutal your post with my own quoted post from another thread that fits in exactly to the conversation. Raw compute power still goes to AMD cards.

Although raw compute rarely translates into raw performance when gaming is concerned and NVIDIA has so well leveraged CUDA with developers AMD compute capabilities are nill.

 

[razor1]

I don't agree with this.

as far as computational power AMD has had the edge for a very long time.

1080TI - tflops = 11.3
Vega - tflops = 12.6

980TI - tflops = 5.6
Fury X - tflops = 8.7

780TI - tflops = 5.0
290x - tflops = 5.6

680 - tflops = 3.1
HD7970 - tflops = 3.8

580 - tflops = 1.5
HD6970 - tflops = 2.7

480 - tflops = 1.3
HD5870 = 2.7

285 - tflops = .7
HD4870 = 1.2


sources:
https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units

It might go back further than this - I didn't search further back -- but to this point - for the modern GPU era -- you can see in every generation going back a decade that AMD had stronger hardware. As such, then they must always have had weaker driver support --- at least in regards to pulling out all the raw performance potential, this since nVidia cards typically had the edge on actual in-game performance.

That's why AMD cards have traditionally been better at password cracking, computational workload, or raw performance at mining - because the compute power is there. NVidia has a stronger driver team and has been able to make a higher performing gaming card with less computer power.

I think this hardware capability disparity also plays into the Fine Wine theme. AMD can make improvements over a longer period of time, because they left so much on the table to begin with.

I agree they do but has it been advantageous in real world applications for them? Yeah but very few though, And as you can see nV has almost caught up with AMD on the Tflop side of things, if we consider boost clocks nV is ahead of of AMD albeit a small margin.

nV has been able to create a smaller chip, with similar computational power with the gp102 as Vega, which is much more power efficient and better at utilizing all of its units in many more applications.

This is what RTG must focus on with Navi. nV did with Maxwell, since then they made their architecture scale-able. They were able to remove all the unnecessary components for gaming cards since then. DP units, all the plumbing for those units. They will do the same for Volta, I don't think any of those tensor units are needed for gaming, it would be nice to have some of them, more from a programming perspective to see what can be done in games, but for now not needed. AMD doesn't have dp units in current Vega. But why the hell is its die so big? Gotta be something in there that GP102 doesn't have right? And that something is consuming lots of power and not giving performance that is acceptble, and this problem has been there since the first GCN product and now has gotten to a point that can't be ignored.

Prior generation GPU's AMD's ALU count was considerably higher than nV's, even that gap has lessened. Its under 25% now? Why is AMD's die getting relatively larger than nV's while nV is able to keep higher throughput and increase Flops figures and ALU amounts? Wierd right, and AMD is supposed to be on a smaller node 14nm.......

 

[sabrewolf732]

AMD is always behind NV innovation wise: Memory compression? Tiled Rasterization? Distributed Tessellation? High clocks? Power efficiency? These are the essence of any GPU, and in none of them is AMD even remotely competitive. Come back when they are. What does AMD have again? Yeah Async Compute and FP16! Console catering features.

Since the R600 days AMD NEVER took the lead from NVIDIA single GPU wise:
HD 2900X massively slower than 8800Ultra
HD 3870 massively slower than GTX 9800
HD HD4870/4890 slower than GTX 280/285
HD5870 slower than GTX 480
HD 6970 slower than GTX580
HD 7970 slower than GTX 680
HD 290X slower than GTX 780Ti/Titan Black
FuryX slower than 980Ti
Vega 64 massively slower than 1080Ti/TitanXp

See the pattern? with Vega AMD has now come full circle to being massively behind. And no amount of driver enhancements will do anything to dissolve this NVIDIA lead, They are just behind like always, It's just worse this time.

the 4870 was nearly as fast as the gtx280 and priced to compete with the 260.

the 480 soft launched like 6-7 months after the 5870, cost more money, used more power, and traded blows (480 is faster in more recent games)

the 580 was faster depending on the game, but also cost a substantial amount more.

the 7970 beat the 580 and then the 680 was launched a bit later, and was faster than the 7970/traded blows with the 7970. This is also where nvidia started to emphasize efficiency.

The 290x beat the 780, ti launched a bit after and was priced substantially higher.

Your eyes are bleeding green dude. AMD has remained rather competitive with nvidia, even the fury was competitive performance wise.

Vega is a bit of a let down and is reminiscent of the 2900xt.

 

[Hameeeedo]

Here I'll rebutal your post with my own quoted post from another thread that fits in exactly to the conversation. Raw compute power still goes to AMD cards.

Raw compute alone is worthless for graphics, you need Texturing, fill rate, polygon rate, tessellation rate, memory bandwidth, and scheduling (lowest compute throughput case). So compute is only good on marketing slides.

 

[OutOfPhase]

Only because Intel used the transition to 64 bits to start a new ISA without the weirdness and limits of x86-32. Yes, they failed, but at least gave a try.

On the other hand, AMD did something trivial by adding 64bit extensions on top of 32bit. Thanks to AMD the current x64 is the more bloated and weird ISA in use and x64 chips are inefficient, power hungry, costly, and buggy.

There are persistent rumors that Intel is trying to clean the x86 ISA again, eliminating the weirdest and oldest legacy stuff from 64bits.

I can't speak to rumors, but I can speak to the impact of x86 decode - it's really not a significant issue anymore. The cost for us in the decode stage is actually a really small portion of the overall chip, and it's a smaller percentage every generation. Everything else expands and gets more complex, but the decode to uOps is really pretty simple. And honestly - you're going to do this anyway. The instruction stream will never be identical to your internal processing after any length of time. You need to keep innovating and software will remain fixed, thus the need for some translation.

As for the design itself - calling it trivial is disingenuous. I'd argue it's brilliant. It solved the real problems we really had (word size, (unsegmented) addressing space, register pressure) without blowing everything up in the process. That's actually a very difficult job. Too many people focus on the complexity of the x86 instruction set when that is only a problem for the top slice of the pipe.

Edit:
And ironically perhaps, the reason Itanium struggles is the same reason GCN struggles now. The instructions are incredibly cumbersome to deal with in reality.

 

[razor1]

the 4870 was nearly as fast as the gtx280 and priced to compete with the 260.

the 480 soft launched like 6-7 months after the 5870, cost more money, used more power, and traded blows (480 is faster in more recent games)

the 580 was faster depending on the game, but also cost a substantial amount more.

the 7970 beat the 580 and then the 680 was launched a bit later, and was faster than the 7970/traded blows with the 7970. This is also where nvidia started to emphasize efficiency.

The 290x beat the 780, ti launched a bit after and was priced substantially higher.

Your eyes are bleeding green dude. AMD has remained rather competitive with nvidia, even the fury was competitive performance wise.

Vega is a bit of a let down and is reminiscent of the 2900xt.

ATI has gotten close in graphics many times but only once has it really taken a lead, and that was with the 9700, and 9800 pro, this happened because nV screwed up with the FX.

The other times it got close was when nV was lazy and reused their architecture for 3 generations, 1 generation more than usual and that was with the tesla architecture like you mentioned with the x4xx vs then 2xx, and with Fermi nV had power issues and again AMD was able to keep it close. One more time ATi was right around nV was with the x1x00 which for its time was right around the 7x00 series and after was better than the 7xxx series but when the x1900x (which is defiantly better than the updated 7x00 series on a smaller node, was released the g80 was only 6 months away.

When ATi had better tech than nV they weren't able to capitalize because at the time the x1800 was shader limited and it was rectified by the x1900 which was ROP limited with the current applications at the time.

We have seen this many times, nV just makes tech that are better for the current games out for the longevity of those products. By the time games shifted over to more ALU needs, the g80 just clobbered the x1900, and the r600 was no where to be found. When it finally arrived it had issues.

 

[Hameeeedo]

Your eyes are bleeding green dude. AMD has remained rather competitive with nvidia, even the fury was competitive performance wise.

This is a comparison by generation. I didn't arrange them chronologically, or with price/perf ratio. This is about absolute performance, of which AMD never had a win in any generation.

Vega is a bit of a let down and is reminiscent of the 2900xt.

Exactly my point.

the 4870 was nearly as fast as the gtx280 and priced to compete with the 260.

Nope, GTX 280 remained the faster option, that extended to GTX285 vs HD4890. The HD4890 was trading blows with the GTX 275.

the 580 was faster depending on the game, but also cost a substantial amount more.

Nope always. GTX 580 was always the faster option.

and traded blows (480 is faster in more recent games)

Nope again it was the faster option:

To wrap things up, let’s start with the obvious: NVIDIA has reclaimed their crown – they have the fastest single-GPU card. The GTX 480 is between 10 and 15% faster than the Radeon 5870 depending on the resolution, giving it a comfortable lead over AMD’s best single-GPU card.

http://www.anandtech.com/show/2977/...x-470-6-months-late-was-it-worth-the-wait-/20

the 7970 beat the 580 and then the 680 was launched a bit later, and was faster than the 7970/traded blows with the 7970. This is also where nvidia started to emphasize efficiency.

Wrong for the 5th? time ..

the GTX 680 is faster, less power hungry, and quieter than the Radeon HD 7970. NVIDIA has landed the technical trifecta, and to top it off they’ve priced it comfortably below the competition.

http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/20

 

[Archaea]

Raw compute alone is worthless for graphics, you need Texturing, fill rate, polygon rate, tessellation rate, memory bandwidth, and scheduling (lowest compute throughput case). So compute is only good on marketing slides.

Not every use for video cards is gaming.

compute apps such as password cracking software, or design applications can benefit from the extra calculation prowess.

 

[Archaea]

Raw compute alone is worthless for graphics, you need Texturing, fill rate, polygon rate, tessellation rate, memory bandwidth, and scheduling (lowest compute throughput case). So compute is only good on marketing slides.

And developers in your pocket. Remember that little story about how AMD couldn't run hairworks in Witcher 3 without a severe performance dedregation and all the Nvidia guys rang the bell. Then later someone found a way to enable hairworks on AMD cards and it actually ran better on the AMD cards than the Nvidia cards?
Proprietary, vendor marketed/locked feature nonsense FTL.

 

[sabrewolf732]

Hameeeedo
"The GTX 480 is quite simply not a "Radeon HD 5870 Killer

We are seeing some trades back and forth; in some games the GTX 480 is slightly faster, in other games the HD 5870 is slightly faster. There wasn’t anything that really stood out, other than Metro 2033. Once again, we have to factor in cost, power, and heat. The GeForce GTX 480 consumes more power than the HD 5870, produces more heat, and costs more money (which might also include the cost of a new PSU). It doesn’t seem like the best value with all this factored in. The only thing that "blew us away" was the heat coming out of the video card and the sound of the fan. If you have not taken a look and a listen to the videos on the previous page, you should."

"
n F1 2010 the Radeon HD 6970 ended up being around 10% faster than the GeForce GTX 580. In Civ 5there is a 17% performance gap favoring the GeForce GTX 580. In Metro 2033 there is only a 5% difference between the two video cards. In BC2 there is a 13% performance gap favoring the GeForce GTX 580. In Mafia II the video cards are nearly identical. It seems that performance is dependent on the game and can range from an overall tie to about a 20% difference.

"Let’s be clear here: the 5870 is the single fastest single-GPU card we have tested, by a wide margin. Looking at its performance in today’s games, as a $379 card it makes the GTX 285 at its current prices ($300+) completely irrelevant. The price difference isn’t enough to make up for the performance difference, and NVIDIA also has to contend with the 5850, which should perform near the GTX 285 but at a price of $259. As is often the case with a new generation of cards, we’re going to see a shakeup here in the market as NVIDIA in particular needs to adjust to these new cards."

When you consider the Radeon HD 6970 is $140 less than the GeForce GTX 580, for at most a 20% difference in performance in some games, the impression of the Radeon HD 6970 turns more positive."

"NVIDIA has also surprised us by providing an efficient GPU. Efficiency is out of character for NVIDIA, but surely we welcome this cool running and quiet GTX 680. NVIDIA has delivered better performance than the Radeon HD 7970 with a TDP and power envelope well below that of the Radeon HD 7970. NVIDIA has made a huge leap in efficiency, and a very large step up from what we saw with the GeForce GTX 580." (as I stated, the 680 is where nvidia started to take efficiency seriously).


Yawn.

 

[juanrga]

I can't speak to rumors, but I can speak to the impact of x86 decode - it's really not a significant issue anymore. The cost for us in the decode stage is actually a really small portion of the overall chip, and it's a smaller percentage every generation. Everything else expands and gets more complex, but the decode to uOps is really pretty simple. And honestly - you're going to do this anyway. The instruction stream will never be identical to your internal processing after any length of time. You need to keep innovating and software will remain fixed, thus the need for some translation.

As for the design itself - calling it trivial is disingenuous. I'd argue it's brilliant. It solved the real problems we really had (word size, (unsegmented) addressing space, register pressure) without blowing everything up in the process. That's actually a very difficult job. Too many people focus on the complexity of the x86 instruction set when that is only a problem for the top slice of the pipe.

Edit:
And ironically perhaps, the reason Itanium struggles is the same reason GCN struggles now. The instructions are incredibly cumbersome to deal with in reality.

The weirdness of x86 don't affect only to the decode stage, but to the whole pipeline. Also the decode penalty is so significant that both AMD and Intel use now uop caches to avoid the penalty of decoding again x86 instructions that were decoded recently.

I call "brilliant" to what ARM did with 64bits, not only it is a separate ISA, but ARM used the transition to 64bits to clean up the ISA, improving implementation and efficiency. What AMD did, to add a 64bit extension to existent ISA was both trivial and cheap.

 

[OutOfPhase]

The weirdness of x86 don't affect only to the decode stage, but to the whole pipeline. Also the decode penalty is so significant that both AMD and Intel use now uop caches to avoid the penalty of decoding again x86 instructions that were decoded recently.

I call "brilliant" to what ARM did with 64bits, not only it is a separate ISA, but ARM used the transition to 64bits to clean up the ISA, improving implementation and efficiency. What AMD did, to add a 64bit extension to existent ISA was both trivial and cheap.

Yes, doing a blatantly stupid decode would have huge impacts. Good thing we don't do that then, right? Yes, uOp caches are one aspect of making it a non issue.
Short version - of course x86/x64 decode could be a problem. But it isn't, and the amount of chip complexity dedicated to making sure it isn't is really not very big.

I'm glad the x64 extension looks trivial. That means we all did our job in making huge improvements pretty painless.

 

[Hameeeedo]

And developers in your pocket. Remember that little story about how AMD couldn't run hairworks in Witcher 3 without a severe performance dedregation and all the Nvidia guys rang the bell. Then later someone found a way to enable hairworks on AMD cards and it actually ran better on the AMD cards than the Nvidia cards?
Proprietary, vendor marketed/locked feature nonsense FTL.

Don't see how this is relevant to GPU specs, you say compute is used for password cracking, yeah but we don't buy GPUs for that, majority of AMD and NV's customers don't buy GPUs for that, they buy them to game on them. And for that specific use compute alone doesn't mean jack.

Hameeeedo
"The GTX 480 is quite simply not a "Radeon HD 5870 Killer

Yeah, when you only test 4 games that will be your conclusion!

You are still ignoring the generation distinction, insisting on a useless "few months" of chronological order. By your flawed method, Volta should be compared to Vega. Not Vega to Pascal. The rest of your post is just confirmation for my line of thoughts. GTX 480 claimed the crown despite being a cut-down underclocked chip, and it's brother the GTX 580 also claimed the crown for the consecutive generation just by being the full chip with original clocks. So you could say Fermi held the gates for NV for 2 generations. Despite AMD transitioning from VILW5 to VLIW4 arc. That's how ahead NVIDIA was back then, (and still is).

 

[Dayaks]

And developers in your pocket. Remember that little story about how AMD couldn't run hairworks in Witcher 3 without a severe performance dedregation and all the Nvidia guys rang the bell. Then later someone found a way to enable hairworks on AMD cards and it actually ran better on the AMD cards than the Nvidia cards?
Proprietary, vendor marketed/locked feature nonsense FTL.

They murdered the AA on AMD to make hairworks work better.... not like it was free.

The vast majority of gameworks has the same effect on both nVidia and AMD. If there is a difference it's usually fixed after a few patches. nVidia could have made gameworks not work at all on AMD if they really wanted.

Gameworks is basically plug and play for AMD. AMD's tech is almost never plug and play for nVidia.

 

[JustReason]

The weirdness of x86 don't affect only to the decode stage, but to the whole pipeline. Also the decode penalty is so significant that both AMD and Intel use now uop caches to avoid the penalty of decoding again x86 instructions that were decoded recently.

I call "brilliant" to what ARM did with 64bits, not only it is a separate ISA, but ARM used the transition to 64bits to clean up the ISA, improving implementation and efficiency. What AMD did, to add a 64bit extension to existent ISA was both trivial and cheap.

Man it would be wonderful if we could see that parallel universe where it was Intel that created x86-64, and see how your statement would differ. Only you have ever downplayed every AMD advancement as "trivial". That has got to be a lot of work ensuring that every positive AMD post gets quashed with irreverent disposition across the web. I wonder if anyone takes you seriously? Seeing posts like this just seems obvious to me you have an agenda, and it isn't the truth or reason.

 

[Magic Hate Ball]

x86-64 was the best way to move the market forward towards 64-bit OSes and applications.

It let people exist on their current platform and eventually switch up to 64-bit, saving enormous costs for everyone.

Period.

Integrating the memory controller into the CPU was a significant performance boost for AMD, which is again something they did before Intel.

 

[JustReason]

They murdered the AA on AMD to make hairworks work better.... not like it was free.

The vast majority of gameworks has the same effect on both nVidia and AMD. If there is a difference it's usually fixed after a few patches. nVidia could have made gameworks not work at all on AMD if they really wanted.

Gameworks is basically plug and play for AMD. AMD's tech is almost never plug and play for nVidia.

he was a tad mistaken as well as you. AMD has in their drivers an overide for tessellation which was the fix for the heavy tess used and no slider to adjust for varying degree of cards. Unfortunately pre-Maxwell cards did not have this option.

 

[Dayaks]

he was a tad mistaken as well as you. AMD has in their drivers an overide for tessellation which was the fix for the heavy tess used and no slider to adjust for varying degree of cards. Unfortunately pre-Maxwell cards did not have this option.

Oh that was a thing too. I was remembering this:
Either way you're sacraficing something.

 

[razor1]

Yeah the whole tessellation thing is just AMD's pipelines stalling when a certain amount of triangles and over has to be rendered that has been their weakness since DX11 games started using tessellation. This is purely the amount of Geometry units they have. The only reason I can think of this, is that they allocate so much die space to the shader array they don't have space to put more geometry units in. I'm pretty sure the geometry units are decoupled with GCN. If they are not then that would be the issue but with GCN 1.3 they were able to increase the GU's from 2 to 4 so I'm thinking they are decoupled.

 

[Presbytier]

And developers in your pocket. Remember that little story about how AMD couldn't run hairworks in Witcher 3 without a severe performance dedregation and all the Nvidia guys rang the bell. Then later someone found a way to enable hairworks on AMD cards and it actually ran better on the AMD cards than the Nvidia cards?
Proprietary, vendor marketed/locked feature nonsense FTL.

So some random guy did a better job then AMD driver team. That is not proprietary lock out that is AMD shitty driver team. Same with Gameworks. NVIDIA does not cripple games AMD shitty driver team does.