So we should purposely allow games to only scale for current gpus and not account for faster gpus.I don't even understand what you are trying to say...
Some CPUs don't need multithreading,not for today's GPUs at least,let's wait for polaris and pascal and look again.
Other CPUs absolutely need it to drive today's GPUs not to even mention the new gen of GPUs.
You guys only look at benchmarks with the top intel CPUs and draw conclusions on scaling or hardware support,but you are forgetting that mantle/dx12 was made from the ground up for a combination of "way too weak CPU cores" + "way more powerful GPU cores"
If 2 or even 4 cores max out current GPUs then you can't see any scaling beyond those 2 or 4 cores,how is this so hard to understand?
So we should purposely allow games to only scale for current gpus and not account for faster gpus.
Thank God you aren't running game development or I'd be screened since I always use a gpu that wasn't out at the time to play a game. When I play the dx12 games out I won't be using Polaris or pascal but unreleased vega....
Funny how people have so many negative things to say about DX12, but can't find a single dev championing DX11 over DX12....
Didn't the Radeon 9800 also rule the roost for several gens afterward too, at least until the Geforce 8800 (the outlier in planned obsolescence).
That said, I also strongly detest the lack of AMDs driver optimizations during the early days of GCN as well. To think that you buy the card now, then have to wait for several years for drivers to finally make full use of the card strikes me as equally asanine as Nvidia's planned obsolescence.
It was either you buy a card that's good now, but performs poorly a couple years layer, or buy a "meh" card now, and wait for it to pick up steam (get it) later. There wasn't much in the way of "buy kickass card now that is still taking names years later."
Here's hoping we don't repeat the beginning of 28 nm.
Didn't the Radeon 9800 also rule the roost for several gens afterward too, at least until the Geforce 8800 (the outlier in planned obsolescence).
That said, I also strongly detest the lack of AMDs driver optimizations during the early days of GCN as well. To think that you buy the card now, then have to wait for several years for drivers to finally make full use of the card strikes me as equally asanine as Nvidia's planned obsolescence.
It was either you buy a card that's good now, but performs poorly a couple years layer, or buy a "meh" card now, and wait for it to pick up steam (get it) later. There wasn't much in the way of "buy kickass card now that is still taking names years later."
Here's hoping we don't repeat the beginning of 28 nm.
Sorry but Tahiti (HD7970 GHz) on release vs Kepler (GTX680) was on par or faster.
Same with Hawaii, R9 290X was faster than GTX 780 on release day one.
Both of GCN cards got way better than Kepler in the coming years. So you had top notch performance at day one AND even better performance 2-3 years later.
I do remember the 7970 GHZ Edition on par or faster than the GTX 680, though I don't think the original did nearly as well. If I remember correctly, the 680 was lauded on anandtech for landing the "technical trifecta" against the original (stock) 7970 while also being cheaper. Later, the 970 would also get the same praise from Anandtech vs the 290.
As for the 290X, I probably only looked at reference benchies (in fact, I tend to look at only reference benchmarks when comparing cards). Quite odd the reference part would be the one to exhibit throttling, but ehh.
Didn't the Radeon 9800 also rule the roost for several gens afterward too, at least until the Geforce 8800 (the outlier in planned obsolescence).
That said, I also strongly detest the lack of AMDs driver optimizations during the early days of GCN as well. To think that you buy the card now, then have to wait for several years for drivers to finally make full use of the card strikes me as equally asanine as Nvidia's planned obsolescence.
Sorry but Tahiti (HD7970 GHz) on release vs Kepler (GTX680) was on par or faster.
Same with Hawaii, R9 290X was faster than GTX 780 on release day one.
Both of GCN cards got way better than Kepler in the coming years. So you had top notch performance at day one AND even better performance 2-3 years later.
This is a fallacy created by Kepler owners to justify buying the inferior NV architecture of the time. HD7970Ghz was the fastest card as of June 2012. That means it only took 5 months for AMD to catch up and beat 680 over the next 4 years.
I don't even understand what you are trying to say...
Some CPUs don't need multithreading,not for today's GPUs at least,let's wait for polaris and pascal and look again.
Other CPUs absolutely need it to drive today's GPUs not to even mention the new gen of GPUs.
You guys only look at benchmarks with the top intel CPUs and draw conclusions on scaling or hardware support,but you are forgetting that mantle/dx12 was made from the ground up for a combination of "way too weak CPU cores" + "way more powerful GPU cores"
If 2 or even 4 cores max out current GPUs then you can't see any scaling beyond those 2 or 4 cores,how is this so hard to understand?
For my test I used the latest build of Windows 10 on a PC with an eight-core Core i7-5960X, 32GB of DDR4/2133 RAM, and an AMD Radeon Fury X GPU. To see the effect of losing cores, I manually switched off cores and Hyper-Threading while running the test.
Note that as you scale back the core count on the chip, Turbo Boost reacts by giving you a little more clock speed. Rather than turn off Turbo Boost, I’ll just note that up to two cores with Hyper-Threading, the chip runs at 3.5GHz. Beyond that it ran at 3.3GHz. In an ideal world, I'd use different CPUs, as each specific chip reacts a little differently, but this is a pretty reasonable approximation.
"But that's 4xTitan X!" True, but unfortunately this video doesn't show where the guy says rendered 4k, but down sampled to 1080p.6 to 12x more than we could do with DX11... also textures are 8k by 8k, significant more than we could do
Ah, ok, think I remember now. Hawaii was the response to GK110. Makes sense it would beat out the 780. That last benchmark shows the 680 pulling ahead of the 7970 GHZ/280X though (the one Atena posted.)
What is interesting is even though Titan has more raw compute, the 280X is really giving it a run for it's money now.
I certainly wouldn't argue driver support on Nvidia's part, especially if the gpu killing drivers are to be believed.
Ironically, some of those people spent weeks trying to show how GCN lacks certain DX12 features (12.1, etc.) and how Maxwell was the way to go for DX12 gaming. Now that 7970/280X demolishes 680/770/780 and 290-390X demolish 970/980 in DX12 games, DX12 games is downplayed and supposedly a waste of time for developers. Yet, Deus Ex Mankind Divided is a DX12 game, and one of the most highly anticipated titles of 2016.
NV actually has a remarkable track record of delivering GPU architecture that last for about 2 years and then fall apart in modern titles thereafter. Looking back, GeForce 7 underwent something very similar in modern titles of the time like how now Kepler and Maxwell are showing signs of struggle under DX12.
No worries, that's part of the NV's planned obsolescence marketing plan, just like the good old days of GeForce 5 and 7, since then long forgotten. If Pascal has massive gains under DX12 vs. Maxwell in 2016-2017, well, there is our answer.
Won't that play to AMD's strengths? It will gimp previous Nvidia generations however.Great post man, as usual, backed with solid evidence.
I have no doubts that Pascal will do much better in modern games DX11 or DX12 than Maxwell.
Pascal has functional fine-grained preemption, something Maxwell claims but actually cannot do (lie), and Kepler lacks. The older GPUs not only lack this feature, so that compute can't interrupt a graphics task in the pipe already to proceed... the basic graphics & compute rendering requires a slow context switch. This means as games use more compute, these GPUs will tank in performance.
Pascal will not tank in compute heavy games, because it's engine has become GCN-like for handling graphics & compute. Even if it lacks multi-engine (ACEs) and cannot do Async Compute, it won't be gimped by it.
Credit goes to @zlatan as he said this about Pascal a long time ago before anyone had any idea. NV's detailed info release about P100 confirms the above features.
Due to this large ability gap that Pascal brings, what is the best way to optimize it so that it performs great in modern games? More compute usage for effects & physics. Expect to see some Pascal-optimized compute heavy GameWork features.
Not sure if you're sarcastic or not, but I don't see the dillema. I expect nvidia domination at least until Vega and until then nvidia needs to hit hard and do anything to widen the gap between maxwell and pascal, to entice upgrades. We might even see it with battlefield.Won't that play to AMD's strengths? It will gimp previous Nvidia generations however.
That's a dilemma.
Won't that play to AMD's strengths? It will gimp previous Nvidia generations however.
That's a dilemma.
I doubt Maxwell will be gimped so much as compute limited. Even in AMD biased games, Maxwell seems to fall in line with GCN according to raw compute. (280x closing in on the 970 due to having similar compute capabilities)Won't that play to AMD's strengths? It will gimp previous Nvidia generations however.
That's a dilemma.
Didn't the Radeon 9800 also rule the roost for several gens afterward too, at least until the Geforce 8800 (the outlier in planned obsolescence).
That said, I also strongly detest the lack of AMDs driver optimizations during the early days of GCN as well. To think that you buy the card now, then have to wait for several years for drivers to finally make full use of the card strikes me as equally asanine as Nvidia's planned obsolescence.
It was either you buy a card that's good now, but performs poorly a couple years layer, or buy a "meh" card now, and wait for it to pick up steam (get it) later. There wasn't much in the way of "buy kickass card now that is still taking names years later."
Here's hoping we don't repeat the beginning of 28 nm.
Well yes, the 290/290x would have been monsters if they released with decent coolers and optimized drivers.I doubt Maxwell will be gimped so much as compute limited. Even in AMD biased games, Maxwell seems to fall in line with GCN according to raw compute. (280x closing in on the 970 due to having similar compute capabilities)
I believe games right now are utilizing Maxwell fairly well. Right now, getting more brute in there is probably the main focus behind Pascal. We'll see how that works out. Without ACEs, it will probably be difficult to maintain peak efficiency in all situations, though if they are able to effectively utilize the chip without, they could very well simply out brute Polaris now, then work on ACEs and other such additions later.
One other thing is the bad PR Nvidia has been gaining as of late. Enthusiasts are already fully aware of Nvidia GPUs losing steam against their competitors, not to mention the memory issue in the 970, and now poor drivers, it would be dumb to recommend an Nvidia card at all unless CUDA was a necessity. I'm not sure if Nvidia can hold back again this time and come away unscathed.
The reason the 290X did so well for longevity is it's a flat out better chip than the 780 and it's built for a future that AMD made.
NV's architectures have been consistently short-sighted because they can't read the future as well as AMD, because AMD knows what they're making.