Sapphire Radeon RX 460 Nitro 4GB

[DanceswithUnix]

He was comparing a Maxwell card to an old Terascale based HD6670 though.

GCN images seem to show rendering spread across the screen: http://www.realworldtech.com/forum/?...rpostid=159935

It just looks like the Nvidia card is rasterising the image in a way more cache friendly manner, which will have both performance and power advantages. If that is the case, then the 460 would be at a definite disadvantage in a competition between bus powered cards.

And before anyone asks, yes I had a Dreamcast and a softspot for tile based rendering

[CAT-THE-FIFTH]

GCN images seem to show rendering spread across the screen: http://www.realworldtech.com/forum/?...rpostid=159935

It just looks like the Nvidia card is rasterising the image in a way more cache friendly manner, which will have both performance and power advantages. If that is the case, then the 460 would be at a definite disadvantage in a competition between bus powered cards.

And before anyone asks, yes I had a Dreamcast and a softspot for tile based rendering

It depends if you are comparing the bus powered GTX950 and a RX460,because the TR showed a bus powered GTX950 based system was consuming the same amount of power as a pre-overclocked RX460 one!

It might be somewhat in favour of Nvidia with the GTX1050,but interestingly enough the relative positions of Polaris and Pascal is not as worse as GCN1.2 and Maxwell,and interestingly Polaris is basically still considered GCN1.2/1.3 for dev purposes.

Remember,the RX460 sold in shops is only a 14 CU part with boosted clockspeeds past the optimal frequencies- the WX4100 is a full P10 chip in a single slot and low profile card,most likely consuming under 75W.

The heatsink looks more akin to the single slot HD7750 cards which were well under 75W.

Its why I think it is a shame we will not see a fully enabled P11 based consumer card.

What is more surprising is that the desktop and laptop RX460 based cards seem to be very close in speed,and some of that might be due to the better CPU in the desktop system(not sure what AMD were thinking about there!!).

It might because the mobile cards use fully enabled chips though,but if it isn't that is quite a bit of progress AMD has made in that regard IMHO.

It also makes me wonder,if the tile based system was the main reason Nvidia improved efficiency,what happens with Volta if they are going towards a more compute focussed GPU again.

Remember,Nvidia pushed a reasonable amount of its scheduling towards using software methods too which helped.

[DanceswithUnix]

It also makes me wonder,if the tile based system was the main reason Nvidia improved efficiency,what happens with Volta if they are going towards a more compute focussed GPU again.

Remember,Nvidia pushed a reasonable amount of its scheduling towards using software methods too which helped.

Whether a device is compute focused would be down to how the shaders are configured, so quite compatible with making the rasterizer work in tiles.

I can see how scheduling in software would allow more transistors to be put into shaders so helping performance, but I don't see how it would help power consumption.

[CAT-THE-FIFTH]

Whether a device is compute focused would be down to how the shaders are configured, so quite compatible with making the rasterizer work in tiles.

I can see how scheduling in software would allow more transistors to be put into shaders so helping performance, but I don't see how it would help power consumption.

From Nvidia:

However based on their own internal research and simulations, in their search for efficiency NVIDIA found that hardware scheduling was consuming a fair bit of power and area for few benefits.

Look at things like async compute:

https://msdn.microsoft.com/en-us/lib...aphics_example

http://international.download.nvidia...pi-support.png

MS were aware of it and Nvidia said Maxwell supported it. It just tells me Nvidia were more concerned about efficiency and DX11 performance,so optimised the uarch towards that,and removed any functionality which was not required in their opinion(async compute being one of them) and AMD were expecting something like a compute focus in games and APIs like DX12 much earlier,so pushed that. It was a miscalculation for AMD no doubt at the time,but we are starting to see the issues with Nvidia's design decisions with Kepler and Maxwell also starting to flare up.

Even at the lower end of the spectrum you might want to look at the consumer cards,GCN seems to be a more general compute orientated uarch(it seems to be replicated right down the range even in older cards).

Remember,the compute focussed Pascal card is a separate and SPECIALIST die with added functionality units,and Nvidia bifurcated the lines on 28NM to Kepler and Maxwell and Volta is meant to be a more general purpose GPU which probably unifies them again.

Have you not noticed Volta is technically delayed as Pascal appeared relatively recently as an intermediate product line?? It is probably quite a big redesign over Pascal IMHO....if it were a reconfiguration of shaders then I expect we would have not seen Pascal released.

Maybe the will have some new memory standard with Volta - it will be interesting to see how big the effiency jump will be. Kepler had its software scheduling,Maxwell had the use of a tile rasterizer,and the lines were bifurcated and Pascal had a node shrink and had dedicated GPUs for each line.

AMD does not use dedicated GPUs for each line,and they have to span different segments,ie,are more general purpose.

There are other things like this:

http://vrworld.com/2016/08/05/amd-ra...rt-4tb-memory/

The way how Radeon Pro SSG works is quite ingenious. When AMD decided to reshape its mainstream GPU codenamed as “Ellesmere” into a “Polaris 10”, part of that effort was to bring more than 16 PCIe lanes inside the GPU. Polaris 10 has as much PCIe lanes as Intel mainstream processors – from Haswell, Broadwell of yesterday to today’s Skylake and tomorrow’s Kabylake. In our conversation with John Swinimer and his colleague, we learned that the initial prototype card is just the beginning. The GPU attaches to the on-board PCIe bridge just like Radeon Pro Duo does. However, the data does not go down to the motherboard but rather stays on the discrete board, ‘talking’ to the NVMe controller which features two M.2 slots.

However,even look at AMD Hawaii,and for such a relatively small chip it has a load of functionality crammed into it.I think people underestimate how much extra functionality AMD has plonked into their GPUs,which is probably a consequence of them not being able to afford to make as many dedicated lines as Nvidia,so need more general purpose GPUs to bridge the different lines they do.

But at the same time I expect it comes with extra costs in power.

[CAT-THE-FIFTH]

The two most compact RX460 cards I have seen are made by HIS and MSI.



The HIS card is £100 delivered from OcUK.

https://www.msi.com/Graphics-card/Ra...#hero-overview



Edit!!

There is an XFX one too:

http://xfxforce.com/en-us/products/a...f-rx-460p2sfg5

Looks like the HIS RX460 2GB is now actually in stock for just under £100 delivered:

https://www.overclockers.co.uk/his-r...gx-10d-hs.html