Friday, June 26th 2015
AMD Didn't Get the R9 Fury X Wrong, but NVIDIA Got its GTX 980 Ti Right
This has been a roller-coaster month for high-end PC graphics. The timing of NVIDIA's GeForce GTX 980 Ti launch had us giving finishing touches to its review with our bags to Taipei still not packed. When it launched, the GTX 980 Ti set AMD a performance target and a price target. Then began a 3-week wait for AMD to launch its Radeon R9 Fury X graphics card. The dance is done, the dust has settled, and we know who has won - nobody. AMD didn't get the R9 Fury X wrong, but NVIDIA got its GTX 980 Ti right. At best, this stalemate yielded a 4K-capable single-GPU graphics option from each brand at $650. You already had those in the form of the $650-ish Radeon R9 295X2, or a pair GTX 970 cards. Those with no plans of a 4K display already had great options in the form of the GTX 970, and price-cut R9 290X.
The Radeon R9 290 series launch from Fall-2013 stirred up the high-end graphics market in a big way. The $399 R9 290 made NVIDIA look comically evil for asking $999 for the card it beat, the GTX TITAN; while the R9 290X remained the fastest single-GPU option, at $550, till NVIDIA launched the $699 GTX 780 Ti, to get people back to paying through their noses for the extra performance. Then there were two UFO sightings in the form of the GTX TITAN Black, and the GTX TITAN-Z, which made no tangible contributions to consumer choice. Sure, they gave you full double-precision floating point (DPFP) performance, but DPFP is of no use to gamers. So what could have been the calculation at AMD and NVIDIA as June 2015 approached? Here's a theory.
Image credit: Mahspoonis2big, Reddit
AMD's HBM Gamble
The "Fiji" silicon is formidable. It made performance/Watt gains over "Hawaii," despite a lack of significant shader architecture performance improvements between GCN 1.1 and GCN 1.2 (at least nowhere of the kind between NVIDIA's "Kepler" and "Maxwell.") AMD could do a 45% increase in stream processors for the Radeon R9 Fury X, at the same typical board power as its predecessor, the R9 290X. The company had to find other ways to bring down power consumption, and one way to do that, while not sacrificing performance, was implementing a more efficient memory standard, High Bandwidth Memory (HBM).
Implementing HBM, right now, is not as easy GDDR5 was, when it was new. HBM is more efficient than GDDR5, but it trades clock speed for bus-width, and a wider bus entails more pins (connections), which would have meant an insane amount of PCB wiring around the GPU, in AMD's case. The company had to co-develop the industry's first mass-producible interposer (silicon die that acts as substrate for other dies), relocate the memory to the GPU package, and still make do with the design limitation of first-generation HBM capping out at 8 Gb per stack, or 4 GB for AMD's silicon; after having laid a 4096-bit wide memory bus. This was a bold move.
Reviews show that 4 GB of HBM isn't Fiji's Achilles' heel. The card still competes in the same league as the 6 GB memory-laden GTX 980 Ti, at 4K Ultra HD (a resolution that's most taxing on the video memory). The card is just 2% slower than the GTX 980 Ti, at this resolution. Its performance/Watt is significantly higher than the R9 290X. We reckon that this outcome would have been impossible with GDDR5, if AMD never gambled with HBM, and stuck to the 512-bit wide GDDR5 interface of "Hawaii," just as it stuck to a front-end and render back-end configuration similar to it (the front-end is similar to that of "Tonga," while the ROP count is the same as "Hawaii.")
NVIDIA Accelerated GM200
NVIDIA's big "Maxwell" silicon, the GM200, wasn't expected to come out as soon as it did. The GTX 980 and the 5 billion-transistor GM204 silicon are just 9 months old in the market, NVIDIA has sold a lot of these; and given how the company milked its predecessor, the GK104, for a year in the high-end segment before bringing out the GK110 with the TITAN; something similar was expected of the GM200. Its March 2015 introduction - just six months following the GTX 980 - was unexpected. What was also unexpected, was NVIDIA launching the GTX 980 Ti, as early as it did. This card has effectively cannibalized the TITAN X, just 3 months post its launch. The GTX TITAN X is a halo product, overpriced at $999, and hence not a lot of GM200 chips were expected to be in production. We heard reports throughout Spring, that launch of a high-volume, money-making SKU based on the GM200 could be expected only after Summer. As it turns out, NVIDIA was preparing a welcoming party for the R9 Fury X, with the GTX 980 Ti.
The GTX 980 Ti was more likely designed with R9 Fury X performance, rather than a target price, as the pivot. The $650 price tag is likely something NVIDIA came up with later, after having achieved a performance lead over the R9 Fury X, by stripping down the GM200 as much as it could to get there. How NVIDIA figured out R9 Fury X performance is anybody's guess. It's more likely that the price of R9 Fury X would have been different, if the GTX 980 Ti wasn't around; than the other way around.
Who Won?
Short answer - nobody. The high-end graphics card market isn't as shaken up as it was, right after the R9 290 series launch. The "Hawaii" twins held onto their own, and continued to offer great bang for the buck, until NVIDIA stepped in with the GTX 970 and GTX 980 last September. $300 gets you not much more from what it did a month ago. At least now you have a choice between the GTX 970 and the R9 390 (which appears to have caught up), at $430, the R9 390X offers competition to the $499 GTX 980; and then there are leftovers from the previous-gen, such as the R9 290 series and the GTX 780 Ti, but these aren't really the high-end we were looking for. It was gleeful to watch the $399 R9 290 dethrone the $999 GTX TITAN in September 2013, as people upgraded their rigs for Holiday 2013. We didn't see that kind of a spectacle this month. There is a silver lining, though. There is a rather big gap between the GTX 980 and GTX 980 Ti just waiting to be filled.
Hopefully July will churn out something exciting (and bonafide high-end) around the $500 mark.
The Radeon R9 290 series launch from Fall-2013 stirred up the high-end graphics market in a big way. The $399 R9 290 made NVIDIA look comically evil for asking $999 for the card it beat, the GTX TITAN; while the R9 290X remained the fastest single-GPU option, at $550, till NVIDIA launched the $699 GTX 780 Ti, to get people back to paying through their noses for the extra performance. Then there were two UFO sightings in the form of the GTX TITAN Black, and the GTX TITAN-Z, which made no tangible contributions to consumer choice. Sure, they gave you full double-precision floating point (DPFP) performance, but DPFP is of no use to gamers. So what could have been the calculation at AMD and NVIDIA as June 2015 approached? Here's a theory.
AMD's HBM Gamble
The "Fiji" silicon is formidable. It made performance/Watt gains over "Hawaii," despite a lack of significant shader architecture performance improvements between GCN 1.1 and GCN 1.2 (at least nowhere of the kind between NVIDIA's "Kepler" and "Maxwell.") AMD could do a 45% increase in stream processors for the Radeon R9 Fury X, at the same typical board power as its predecessor, the R9 290X. The company had to find other ways to bring down power consumption, and one way to do that, while not sacrificing performance, was implementing a more efficient memory standard, High Bandwidth Memory (HBM).
Implementing HBM, right now, is not as easy GDDR5 was, when it was new. HBM is more efficient than GDDR5, but it trades clock speed for bus-width, and a wider bus entails more pins (connections), which would have meant an insane amount of PCB wiring around the GPU, in AMD's case. The company had to co-develop the industry's first mass-producible interposer (silicon die that acts as substrate for other dies), relocate the memory to the GPU package, and still make do with the design limitation of first-generation HBM capping out at 8 Gb per stack, or 4 GB for AMD's silicon; after having laid a 4096-bit wide memory bus. This was a bold move.
Reviews show that 4 GB of HBM isn't Fiji's Achilles' heel. The card still competes in the same league as the 6 GB memory-laden GTX 980 Ti, at 4K Ultra HD (a resolution that's most taxing on the video memory). The card is just 2% slower than the GTX 980 Ti, at this resolution. Its performance/Watt is significantly higher than the R9 290X. We reckon that this outcome would have been impossible with GDDR5, if AMD never gambled with HBM, and stuck to the 512-bit wide GDDR5 interface of "Hawaii," just as it stuck to a front-end and render back-end configuration similar to it (the front-end is similar to that of "Tonga," while the ROP count is the same as "Hawaii.")
NVIDIA Accelerated GM200
NVIDIA's big "Maxwell" silicon, the GM200, wasn't expected to come out as soon as it did. The GTX 980 and the 5 billion-transistor GM204 silicon are just 9 months old in the market, NVIDIA has sold a lot of these; and given how the company milked its predecessor, the GK104, for a year in the high-end segment before bringing out the GK110 with the TITAN; something similar was expected of the GM200. Its March 2015 introduction - just six months following the GTX 980 - was unexpected. What was also unexpected, was NVIDIA launching the GTX 980 Ti, as early as it did. This card has effectively cannibalized the TITAN X, just 3 months post its launch. The GTX TITAN X is a halo product, overpriced at $999, and hence not a lot of GM200 chips were expected to be in production. We heard reports throughout Spring, that launch of a high-volume, money-making SKU based on the GM200 could be expected only after Summer. As it turns out, NVIDIA was preparing a welcoming party for the R9 Fury X, with the GTX 980 Ti.
The GTX 980 Ti was more likely designed with R9 Fury X performance, rather than a target price, as the pivot. The $650 price tag is likely something NVIDIA came up with later, after having achieved a performance lead over the R9 Fury X, by stripping down the GM200 as much as it could to get there. How NVIDIA figured out R9 Fury X performance is anybody's guess. It's more likely that the price of R9 Fury X would have been different, if the GTX 980 Ti wasn't around; than the other way around.
Who Won?
Short answer - nobody. The high-end graphics card market isn't as shaken up as it was, right after the R9 290 series launch. The "Hawaii" twins held onto their own, and continued to offer great bang for the buck, until NVIDIA stepped in with the GTX 970 and GTX 980 last September. $300 gets you not much more from what it did a month ago. At least now you have a choice between the GTX 970 and the R9 390 (which appears to have caught up), at $430, the R9 390X offers competition to the $499 GTX 980; and then there are leftovers from the previous-gen, such as the R9 290 series and the GTX 780 Ti, but these aren't really the high-end we were looking for. It was gleeful to watch the $399 R9 290 dethrone the $999 GTX TITAN in September 2013, as people upgraded their rigs for Holiday 2013. We didn't see that kind of a spectacle this month. There is a silver lining, though. There is a rather big gap between the GTX 980 and GTX 980 Ti just waiting to be filled.
Hopefully July will churn out something exciting (and bonafide high-end) around the $500 mark.
223 Comments on AMD Didn't Get the R9 Fury X Wrong, but NVIDIA Got its GTX 980 Ti Right
While we might be seeing some Enthusiast from just a year or so ago starting to feel the "drum-beat" of 4K, I see it’s more reality in the next node shrink. I perhaps think there’s gamers that might try 3x 1440p for Eyefinity/Surround if the price was right, that’s perhaps a better the hook to sell these top-end cards.
What I was postulating, is mainstream gamers need to see the path in transitioning from 1080p. It seems right today panel manufactures aren't or can't move higher resolution and their subsequent tech down in price, so the whole thing goes stagnate. Sure GPU side can offer the hardware (what they can within 28nm); however folk aren’t going to buy into that if they can’t see a monitor(s) upgrade as obtainable in the near term. I really thought panel manufactures would've ran with the whole Adaptive Sync to aid in some new market push, but it (or they) don’t seem to be generating that momentum. A mainstream gamer (@1080p) should be able right now to realize a card that's in a $280-400 range, and then see in the relative near future decent though generic 1440p Adaptive Sync panel for $350-400. From where I’m watching, I’m not sensing a move in that direction.
It's still probably fair to assume the game was better tested/optimized on NVIDIA hardware though if they include that at all.
At the same time, games devs that do work with one team or the other still make their games completely playable on the other's cards. They'd be stupid not to. The other team's cards might get a few FPS lower, but it has never been to a point that I said "damn this game is crippled because my card is made by team XYZ". It just doesn't happen.
Finally, don't even try to say nVidia is the only one outright lying to our faces. AMD has lied a lot more, and said much worse lies, than nVidia. But for some reason they just seem to get a pass...
Creating physics effects with the use of PhysX, usually means that there is NOT going to be available a second physics engine that can offer the same quality of effects using the CPU or/and the GPU. That means that a person that does not use an Nvidia GPU sees something inferior on his screen.
Even if that person has a secondary Nvidia card, he can't use it because Nvidia forbids it. Patches that where removing the lock Nvidia implements in it's drivers, showed that using an Nvidia secondary card for physics was 100% working.
On the other hand Mantle does NOT affect the quality for someone running the game with DX11. What the person running the DX11 version sees on his screen, is the same as to what the person running Mantle sees on his screen. No one is treated as a second class customer.
And of course Nvidia would have never agreed in supporting an AMD's tech with AMD's logo on it. Instead it does support DX12 and I believe Vulkan. So in a way it does have an equivalent to Mantle. Nvidia on the other hand never promoted or supported the idea of a physics engine that will take advantage of what PhysX offers today and let someone else make it run on all gpus. Instead they create more proprietary techs thought GameWorks that can guarantee that whoever is not paying for Nvidia hardware will be treated as a second class customer with inferior quality and/or performance in the game.Runs perfectly on Nvidia GPUs.
I merely object when people like you spout a PR line about things being available for all equally, when they really aren't. However ALL those effects from both sides are gimmicks, and not necessary for good, fun gameplay, so not having them doesn' bother me. I just object to the statement you made about Tress-FX being just as usable on Nvidia cards.