Originally Posted by looncraz
Yeah, I'd like to think mGPU/SLI/CF as dead.
More so than nVidia, they are. All GCN GPUs are still completely relevant and can run any game using the latest APIs (with the right mix of settings). Contemporary nVidia cards can make no such claims. Prior to Pascal, DirectX 12 would experience no, or even negative, scaling. Now, it's either none or very little. An improvement, but still not great.
Tahiti lived on to be the R9 280, a mid-range card. The GTX 680 was the nVidia competition. They used the same power and had the same-ish performance when they were both on the market at the same time. Today, however, the R9 280 can be seen to be nearly TWICE as fast in modern games as the GTX 680 (Hitman comes to mind). That's what future proof means - the AMD cards still offer playable performance all these years later, nVidia cards do not - even if you buy their top cards.
Yup, who would have ever imagined that a company would gimp their own hardware so thoroughly in an effort to reduce power consumption? Sure, not everyone needed the scheduler capabilities or compute capabilities, but it wasn't a decision that was likely. nVidia didn't engineer better efficiency, they stripped capabilities to gain it. With Pascal all they did was optimize for frequency and rely on the process for power savings. They haven't been doing any real innovation on the power efficiency front. AMD, however, has been. They haven't stripped the GPU or its capabilities in order to save power, they've found areas where power usage was higher than needed.. then squandered their gains by running too much default voltage nearly across the board... still don't what happened to their binning process improvements.
Fiji easily bests the Titan, the 980Ti is just an overclocking beast. With its stripped hardware, it's also efficient. Of course, things change when DX12 or Vulkan are in play... and Fiji is seeing increasing performance month after month.
Not anymore, no. RX 480 supply is keeping up with demand and has been in constant supply for the last week.
Polaris is by no means power hungry. If you haven't noticed user tests, RX 480 can pull as little as ~130W at stock clocks with proper voltages (mine does!). Tweak the memory frequencies and you are knocking it with the GTX 1060 in both efficiency and performance. Not all cards are created equal, though. Some just won't go below 150W at stock clocks...
I will be doing a voltage scaling in regards to power consumption analysis very shortly - it's pretty insane how quickly power usage increases, even at idle, with increased voltage on these cards.
Yup, nVidia cranked up the clock speeds. Good for them. They probably aren't going to be able to pull off the same feat again. AMD has more engineers knowledgeable about high frequency computing - should they ever deploy them on the GPU front, nVidia will be unable to keep up.
We know a lot about how Polaris is bottlenecked. It could gain an easy 15~20% more performance with appropriate memory bandwidth... Vega will have that bandwidth, and more SPs, and even more updates. A Fiji-sized GCN4 GPU, running just two HB2 stacks, providing 512GB/s bandwidth, would match the 1080 in nearly every metric. Efficiency would be slightly worse due to inefficient CU performance scaling, but a 1.2GHz 64CU GCN4 GPU would match the GTX 1080.
Vega is NOT GCN4. It's updated tech. No idea what that really means, but it should mean that the smaller Vega chip will stomp all over the GTX 1080, quite possibly also in efficiency... and the Titan XP isn't all that far ahead in the grand scheme of things.
nVidia has been running their hardware much closer to their peak clocks and advertising lower frequencies so the overclocked frequencies look more impressive... but the reality is that the GTX 1060 and RX 480 overclock nearly the same, with a slight edge to the RX 480 (an overclock of 2.05Ghz on GTX 1060 is the same 1340MHz on RX 480).
RX 480 scales nearly perfectly linearly with memory frequency. It is ABSURDLY bound by memory performance:
Anyone who says otherwise is just experiencing driver problems and may not even know it.
I've put hours of work into examining Polaris 10 in both RX 470 and RX 480 clothes. I have even created full curves of memory sensitivity in 200MHz increments from 600Mhz:
You can see that scaling only really stops with the 750MHz GPU and 1600MHz RAM. That would be 2.667GHz RAM clocks for 1250MHz... or 340GB/s. Meaning Polaris 10 needs 33% more bandwidth than it has for optimal performance. I haven't done the math, but I think it would be about 25% faster with that much bandwidth, with no increase in GPU clocks... and almost no change in power usage.