We've already had a discussion regarding the benefits of tiled rasterization above, so I won't go into details. But there are some myths that need dispelling:
No game is stupid enough to render things that are behind the camera. At the very least games check whether the bounding box of an object intersects with the camera's view ("frustum") and discard ("cull") any objects that don't. See "frustum culling" for details.
In fact games use a series of finer and finer (and thus more expensive) mechanisms to determine whether a triangle needs rendering. So even if an invisible/occluded object is rendered it will be discarded, either by game engine specific optimizations, clipping or the z-buffer. The earlier a primitive is discarded the less computationally expensive it is.
In conclusion, there is not nearly as much left for a tiled renderer to cull as one might think. Yes, it is useful, but don't get carried away by thinking that it can free up large amounts of resources. Tiled rasterization has been used in mobile devices for a while now, with the intent to save bandwidth. How useful bandwidth savings on a card with High-Bandwidth-Memory are is left up to you to contemplate.
On a side note: If a game would actually render objects behind the camera, you would see them. Been there, done that. Headache ensued.
I don't know enough or care enough to argue about this, just was sharing info I had from a trusted friend. Who also noted he had no idea what kind of perf impact it would have other than an improvement for some games. Also didnt read most of the thread tbh was sharing results I got with FE Vega cards. And this is also a thread for it. So yeah.
In other news Unigine is broken as all get out with Crossfired Vega FE. Got half the score of a single GPU with two installed because ???? Didnt do that many tests other than some 3dmark and GPUPI because i had 4 hours total testing time. Still.
I saw Nexus's video investigation on IPC Vega vs Fury and there was little difference. I'm really disappointed by this.
I think If RX Vega is going to be good they need to drop the HBM2. It adds a lot to a Baum cost of the card due to the poor yields, and not to mention the complexity of the memory controller going into it. GDDR5 has been around forever.
I don't think its actually doable though. AMD probably doesnt have the R&D budget to make two versions of a memory controller for this gpu.
Yeah so its slightly above 1080gtx, which is ok if AMD can undercut it decently price wise. If it was GDRR5 or X memory I'd be 100% certain of that, however because its HBM2, it really tosses a monkey wrench in the whole predicting retail price, its entirely possible it could be at 1080ti prices.
Also HBM2 can't be dropped at this stage, no doubt they have 100s thousand HBM Vega Chips made now, which is baked in on the GPU. I'd be surprised if there is ever a 590 card with Vega+GDDR5. If there was it would perform the same was the top end so I doubt they bother.
PS. SLI and Crossfire is a dead horse, why people even bother I dunno, good for select titles and testing only. Mostly its a headache to configure and get running flawlessly like a single GPU would run.
Well the Vega Chip is technically a near MCM technology already. And internally supports Infinity Fabric Interconnect.
The GPU die can technically be separated from the HBM2 stacks and interfaced to other memory or interfaced with other GPU's and memory aswell.
All of these are some interesting goings on with regards to engineering the Navi GPU's where it seems that AMD is looking to create something similar to the CCX architecture on Ryzen for the Navi GPU's to be used in the AMD Intrinsic GPU compute platform.
I wonder if XT and XTX will have any difference with the boards or if they're the same with different power settings and coolers. I'd imagine AMD could make the PCB a lot shorter than it is with the FE card and still use a cooler similar to the one found on the RX 480 where the fan overhangs the PCB for better intake airflow.
Personally I am not really interested in a watercooled card. So if I jump on RX Vega, it would probably be a Sapphire RX Vega XT Nitro or something like that. Still, my Fury is fine and the performance difference seems to be less than I had hoped for. So... I'll pass most likely.
We don't really know the true performance of vega yet, AMD rushed FE just so they didn't have to lie to investors about launching in first half due, they were delayed to due HBM2 production being too late.
I'm mainly interested if there is any difference between the watercooled version and the air cooled one other than coolers, since I'm going to put a custom block on it anyways I'm just going to get the aircooled 4096 one if there isn't any other difference.
I'd be keen on the XTX version because it normally is a smaller sized card and I can mount the cooler at the bottom of my ITX case.
The Air cooled cards might be pushing the max length for my case, however if the XTX is $1000aud then I probably won't be all that interested, which is the price of the 1080ti here which have 3gb extra memory and will absolutely perform better when overclocked guaranteed.
1100mhz on the HBM and 1700mhz on the core is VERY GOOD considering they don't have voltage control. Considering the stock voltage for Vega is 1.2v and there is tons of microthrottling along with the current drivers I'm very sure we're going to see 1800mhz vega sooner or later.
My only concern is the 375W power draw, the (Gigabyte GA-AB350N-Gaming WIFI MINI ITX) should be able to do the job but I just wonder about the long term survivability of such a card, I remember back in the 290X days when those cards draw lots of power, allot of them died eventually.
Power draw isn't an issue as long as you can cool it, lots of 290Xs ran 90c with the reference aircooler, the stock air cooler on Vega is actually very good considering it handles 300w at half the noise level of a titan xp and is cooler whilst doing so. Another thing about 290X is that it runs hot, and has a 512bit memory bus, aka 16 memory ICs, lots of failure points. You can find lots of defect Hawaii based cards on ebay and so on, whilst I have never seen a Fiji based card that wasn't broken by accident, seems like HBM is a lot more reliable by the looks of things.
You don't have to worry about your motherboard anyways, vega hardly pulls anything to the motherboard's pci-e slot as all the vcore phases are hooked up to the two 8 pin connectors.
Do you live in a shoebox? You can downclock Fiji very well if you find it pushes out too much heat when gaming. -28.5% in wattman along with 900mv for all p-states lets it sip 60w during gaming.
Makes me wonder what sort of power scaling vega has.
