I wanted to answer all of this yesterday but I didn't have time to watch the video.
Beware, huge wall of test incoming. This sort of turned into a huge rant about everything.
#1144
Yes. Mostly AMD Polaris though. Should be released around June/July. We're expecting an entry level chip and a mid range chip, both vastly more efficient than the chips they're replacing because it's the first process node change in 4 years and we're skipping one node altogether. But no one cares about efficiency right? Here's why you should be hyped anyway:
Mid range as in somewhere between 970/390 and 980/Fury. Now imagine a GPU that's more efficient than the 970/980 (the only downside of the 390), faster than the 970 and still cheaper than both 970 and 390.
Definitely something worth waiting for.
Now nVidia, that's complicated. This ties into
#1146
because a lot of this is explained in those videos.
Reasons why you should watch AdoredTV's videos:
- His accent is glorious and
- he's right.
Right now AMD's goals and decisions, both on the hardware, software, marketing and corporate side, happen to align with what's good for consumers. Most of it they aren't doing because they like us, rather because they have to. From our perspective they're doing the right things for the wrong reasons, which is not a bad thing for us.
On the hardware side of things nVidia is doing a lot of things that are technically (literally) the right thing to do, but aren't ideal for the consumers. It's not ideal, but if they had too large of an impact on performance you could always just vote with your wallet and force them to do something about it, right? We'll see about that later.
First some examples:
DX12. Kepler and Maxwell have basically no DX12 only features implemented in hardware. It does make sense since you can use all the saved space to improve DX11 performance but it also means that you won't benefit from DX12 on the GPU side (CPU side gains stay the same obviously). It's not exactly planned obsolescence, more like welcome obsolescence. They get better DX11 performance and don't have to spend money on engineering DX12 hardware before it's even released. Consumers being forced to upgrade earlier since their old cards don't won't get a boost and therefore a prolonged lifespan form DX12 is a pleasant side effect for nVidia but an unpleasant one for us. Whether or not they had that in mind doesn't really matter. Note that AMD didn't add asynchronous compute engines because they're so forthcoming and kind-hearted, but rather because they don't know how much longer they'll have to keep rebranding GCN architecture cards. They simply can't afford to develop two or three new chips per generation (it's been one or at best two for a while) so their chips have to last longer. Letting them become obsolete once DX12 was released just wasn't an option for AMD.
Double precision performance on the GM200. Traditionally nVidia has needed a bit more die space for the same performance as AMD. This isn't a problem when you can just make a bigger chip, although it increases cost. Both Fiji and the GM200 were balls to the wall ~600mm² chips. TSMC physically can not build a chip larger than 625mm², going bigger just isn't an option anymore. On the AMD side the DX12 stuff taking up some space that nVidia doesn't use is balanced by the GDDR5 memory interface being replaced with a far smaller and less power hungry HBM interface. nVidia didn't want to lose the performance crown to an architecture that even with the advantage of HBM is still less efficient simply because they couldn't fit enough units on a die, which meant that they had to cut something else. They choose douple precision performance. So while this made it possible for nVidia to produce a chip (GM200) that is more efficient and beats AMD (Fiji) although only barely, it puts the Titan X in a really awkward spot. The usual semi-professional segment that loved Titans before because of $5k Quadro DP performance for $1k doesn't have a reason to buy it when the 980 Ti offers almost the same at a far lower price and the Fury X offers more than double at an even lower price than that. More on that later under "corporate".
The 970 memory "issue" actually isn't a case of this. What they did allowed them to disable a part of the cache and ROPs which they otherwise couldn't have done. That meant higher yields and therefore a lower price. In fact the yields are so good that we still haven't seen a 960 Ti. With the GK104 we saw 3 GPUs: The full chip (680), the cut down chip with less shaders (670) and an even further cut down chip with less shaders, less ROPs, smaller L2 cache and smaller memory bus (660 Ti). With the GM204 there was never a need for the third one since they could cut down cache and ROPs without having to cut down the bus width. They didn't sell the 970 that "cheap" because it's broken, they spent a lot of money and effort on making it cheaper. I'm pretty sure they could sell the 970 even cheaper, worst case their cost is somewhere between a hypothetical "normal" 970 and a planned 960 Ti, best case it's exactly the cost of a 960 Ti which would sell at ~250$. Either way they're making a killing whereas the 390 is an improved 290 with more VRAM marked down 70$, that can't be good for your margin. I'm also pretty sure that this wasn't about GM204, rather about GM200. I'm sure nVidia is fine with making a couple of millions extra, but yields for a 398mm² chip were never going to be a problem. A 601mm² chip on the other hand, you want to use everything you can think of to improve your yields. I doubt that the yields are good enough for the 980 Ti that nVidia never even considered using it, it's more likely that they're afraid of the bad press if they "gimped" a high end card. My point is that the 980 Ti could be cheaper if people weren't so upset about something that isn't actually an issue. The driver did mask it pretty well, barely any performance impact in games and it took fairly long before anyone noticed something was up at all. I'm blaming bad journalism for this, clickbait titles and never bothering to explain that what they're getting sued over isn't the memory bandwidth but rather the other specs, more under "marketing".
On the corporate side they started doing stuff that was absolutely right from a business perspective and completely acceptable, for example the Titan, Titan Black and Titan Z. AMD couldn't beat the 780 so it was reasonable not to release the full GK110 chip since it is a very big (561mm²) and therefore very difficult and costly to produce chip. Releasing the Titan meant that you could get better performance (and double the VRAM) if you absolutely wanted/needed it at all costs, but you'd have to pay for it. The main point however was the you could get the double precision performance of a 5000$ professional grade GPU (Quadro K6000), although not the driver support, at a fraction (1/5) of the price. Once AMD released the 290X it got even better, the 780 Ti offered the full GK110 at a normal price while the Titan Black got you the double precision performance if you needed. Again at a premium over a gaming GPU, but still far cheaper than a workstation GPU.
And then came the Titan X. Again, holding back the full chip if there's no competition is fine. Not being able to fit higher double precision on it is fine too, it's already huge (601mm²). But they still wanted to make more money, so they forbade add-in board partners from manufacturing 12GB 980 Tis, because double the VRAM is now the only selling point of the Titan X. The 980 Ti and Titan X are so close that the added power draw of the doubled VRAM on the Titan can actually make it slower than the 980 Ti due to power budget restrictions. Yet they're still selling it at 1000$, asking 50% more just for double the VRAM and no other advantages.
continued in next post