Getting up to Speed with Fiji and Maxwell
If you have no idea about what’s happening with AMD and NVIDIA, then you either live under a rock, or you just don’t care about computer components that much. Assuming that you’re here because you don’t fit into either category, then let me get you up to speed with what’s been happening in the GPU world, especially since the anticipated release of stacked memory is right around the corner. The stage is set with NVIDIA dominating the GPU market. Maxwell was impressive when it released, but has managed to become one of the most notable GPUs to date. AMD on the other hand, is telling us they aren’t out yet. With a slew of refreshes, as well as two high end GPUs that are the first ever to feature stacked memory, the Green team might be facing some stiff competition in the upcoming weeks. DirectX 12 could also have some new implications on the gaming front so let’s throw all of this together and take a stab and what the future holds for us.
We now have a fully unlocked GM200-400 die in the form of the Titan-X and a slightly cut down GM200-310 die in the GTX 980 Ti. The Titan-X is the graphics card most of us loftily dream about having some day, but never seriously imagine owning. The GTX 980 Ti is the card that we might actually sell a kidney for. Many people were shocked to see a $649 starting price tag for what would be considered NVIDIAs go-to enthusiast card. When you factor in that the gaming performance is right up there with the Titan-X, the card becomes a no-brainer. There is hardly any reason to pay a price premium for the fully unlocked, 12 GB Maxwell card when the slightly cut down, 6 GB version will give you almost the same results. Even though the $649 tag is still a pretty hefty, it’s much more reasonable considering the level of performance the card gives. It’d be pretty hard to understand why NVIDIA brought there MSRP so low, unless AMD is making something that actually competes.
AMD started bringing up the idea of stacked memory a long time ago. The idea is that HBM (High-Bandwidth Memory) has a much higher bandwidth thresh hold, is much more power efficient, and can occupy a much smaller space than GDDR5. While we’ll get to see the full benefits soon, it’s already beginning to look like HBM is going to be a huge thing for the future of graphics cards. Currently, the code-named Fiji XT die is going to be in the flagship Radeon Fury X card, while a slightly cut down Fiji Pro die is placed in the standard Fury card. The Fury X will have both liquid-cooled and air-cooled designs, while the Fury will initially release with just an air-cooled design. You may have noticed that AMD seems to have taken a page from NVIDIA in that it is using a new naming scheme for its ultra-enthusiast cards, using the normal naming scheme currently adopted for the refreshes. Here’s where Fiji gets interesting, though. While full review testing is still a week or two out, some initial benchmarks released show that Fury X seems to be trading blows with NVIDIA’s Titan-X. In fact, the upcoming GPU pulls slightly ahead of the Titan in the Firestrike Ultra benchmark. Considering the Fury X only has 4 GBs of HBM, that’s an impressive feat to say the least. In regards to 4k resolution, the HBM has a memory bandwidth of 512GB/s, a significant jump compared to the R9 290X only having 320GB/s memory bandwidth, which seems to play a huge factor in large resolutions. Fury X could be bringing the tight competition we enthusiasts have been hoping for rather soon.
While we’re on the competition note, the R9 300 series refreshes seems poised to offer some good competition as well. The GTX 970 is an extremely hard card to compete with. It has more of a mid-ranged price tag, but the heart of a top-tier card. It handily beat out the R9 290X in most everything, while still being able to hang in there at 4k resolutions. The good news is, the R9 390X is looking to offer a better rivalry than it’s predecessor. AMD, while still using the Hawaii core, looks like they’ve optimized it a bit more, increased the memory clock rate, and buffered the capacity to include a full 8 GBs of ram right off the hop. This makes the card more competitive at lower resolutions, but gives it a fairly solid advantage at the 4k range. Most likely, users will tend to stick with the 970 if they know they will be stuck with 1080p for a while, but if you want that safety net for upgrading to an ultra high-res monitor, the R9 390X might be the ticket. With a confirmed MSRP of $389, the revamped Hawaii card can be a great option for a mid to high ranged budget. We’ll also get to finally see a fully unlocked Tonga GPU in the R9 380X with a slightly smaller version in the R9 380. Considering Tonga showed some promise, but had a smaller memory interface, it will be interesting to see what it can do with a 384-bit interface and 4-6 GBs of memory. Pricing looks to be in the $200-$300 range, but final specs and exact costs haven’t been confirmed for the 380X yet.
With the GPU landscape laid out, we can finally address some possibilities of what DirectX 12 and low level APIs can mean for gaming. I’ve heard it said before that AMD would have great performance if they just had drivers to take advantage of their hardware. It would almost seem to make sense when you compare the 4096 stream processors in Fury X to the 3072 CUDA cores in Titan-X, but the reality is, there is a ton of untapped potential in all of our GPUs do to the way previous APIs worked. One of the things that works in favor for AMD is the use of async shaders. In short, these are mechanisms that use parts of the GPU that would otherwise be forced to sit idle. While this can lead to some performance benefits, it may not be anything that can visually be seen short of bringing framerates to a level that’s noticeably better. Since DirectX 12 already improves the performance of GPUs, the other feature to note recently is the conservative rasterization support. This is something that Maxwell 2.0 can take full advantage of and actually leads to visual improvements. NVIDIA demoed this and basically, when we currently see a shadow in a game, it’s usually mapped out and pre-defined to work with an object. In a sense, it’s almost two objects working together which can become very graphically demanding. Rasterization will enable ray tracing which artificially simulates how real shadows work by shading the pixels that are being obscured by an object between the light source. Unfortunately, this feature will have to be specifically supported, but NVIDIA typically doesn’t have too much trouble garnering support. Overall, DirectX 12 looks like it will be a big deal, has helped inspire other low-level APIs, and very well could change how we experience games in some of the best ways possible.
So there you have it. I know that’s a lot of content to go through, but this is what the GPU landscape is looking like pretty much through the end of the year. The good news is, NVIDIA plans on integrating HBM next year with Pascal, which shows that AMD seems to be on the right track. While I was hoping for more new releases from team Red, the price points on the refreshes seem to be very competitive and Fury X looks like it’s the first card to really give the Titan series a run for it’s money. Check out some of the links below for further details and expect an exciting day as Fiji is officially launched during E3 next week!