290x vs 390x

[Sakkura]

Doesn't change anything though. Overclock both and both will have a good amount of FPS increase. The strange thing is that 290x/390x edged out the 980 when it isn't supposed to on stock speeds. Might mean that drivers or the game itself needs a lil bit more work.

 

A 390X narrowly outperforming a GTX 980 at stock speeds is perfectly normal. The other relative performance levels also make sense, such as the GTX 980 Ti and Fury X being closely matched, with the 980 Ti a little ahead at 1080p and the Fury X a little ahead at 4K.

[Sakkura]

I just want to point out - AMD have roughly 50% more instructions per clock so 1MHz on an AMD card is about 1.5MHz on an Nvidia competitor - roughly.

 

Wut? No. Maxwell performs better than GCN per clock and "core." The AMD cards have significantly more shaders and texture units than Nvidia cards with comparable performance. 4096:256:64 on the Fury X vs. 2816:176:96 on the GTX 980 Ti.

[don_svetlio]

Wut? No. Maxwell performs better than GCN per clock and "core." The AMD cards have significantly more shaders and texture units than Nvidia cards with comparable performance. 4096:256:64 on the Fury X vs. 2816:176:96 on the GTX 980 Ti.

Really...? Set a 970 to 1GHz and a 390 to 1 GHz - guess what - 390 is ahead by 15-20 fucking per cent. It's been like so since early Kepler days.

[Kinda Bottlenecked]

A 390X narrowly outperforming a GTX 980 at stock speeds is perfectly normal. The other relative performance levels also make sense, such as the GTX 980 Ti and Fury X being closely matched, with the 980 Ti a little ahead at 1080p and the Fury X a little ahead at 4K.

 

For a 390x that might be the case, but very less likely for a 290x which has much lower clock speeds. Same goes for the 290 and the 970. 

[Curufinwe_wins]

In terms of clock speeds yea. Not so much for in game performance. Hawaii and the other GCN cards still overclock alright. Though not so much for GCN 1.3. Now those are not good overclockers. 

 

It is fair to say that the 980 Ti overclocks better than the Fury X in general(Fury/X can't overclock for nuts, unless u do LN2). Sadly the same is not said for the others like the 980/390x/290x or  970/390x/290 and so on. Because those when you have a nice overclock will have a good amount of FPS increase for all of those cards. Like buying a G1 Gaming or so for some nice overclocks. You gotta buy the right card on the other side as well. 

It depends on the situation quite a bit.

 

GM204 for example gets way more relative 'real-world' performance improvements (gm204 is a poster child for invalidating 3dmark as representative for gaming) from overclocking at 4k than it does at 1080p, due mainly in part to memory overclocking reducing a vram bottleneck, just like (to a much lesser extend) the fiji cards overcome their 4GB vram by having an insanely fast bus. Hawaii overclocking however is much more resolution independent (basically granting 3-5 fps no matter what) and since comparably speaking hawaii tends to be better even right now at 4k than gm204 (while being trading blows at 1080p) it is important to be able to see that effect in action.

 

Obviously fiji is awful at overclocking, but EDIT: Tonga doesn't seem to be which is the same architecture (obviously different memory solutions, so perhaps a bus interface is a limiting factor. I don't know.)

 

I agree the right card is needed, and there are fewer ultra-premium versions for AMD cards. (Indeed, the current consensus is that the MSI Gaming model while not being the coolest, tends to have the best overclocks, which is amusing.)

 

But back to the issue at hand, I don't use graphics cards at stock. I don't care what they do at stock. Overclocking is different for every architecture and core (but honestly fps differences between the same card in the modern age isn't very large. 50 Mhz on a nvidia card doesn't mean much and +/-35 Mhz on a Hawaii Gpu means only slightly more), so yes, I would prefer personally to only see overclocking results (and overclocked specs so I can compare to other benchmarks that also overclocked.)

[Curufinwe_wins]

Really...? Set a 970 to 1GHz and a 390 to 1 GHz - guess what - 390 is ahead by 15-20 fucking per cent. It's been like so since early Kepler days.

He did say per clock per core. Per clock ofc amd cards preform higher. They scale more linearly with clock speeds, and they start off with lower clocks. 

 

I honestly couldn't tell you if per clock per core which is higher because the core speed of nvidia gets attacked by the larger cores of AMD. It's a different philosophy which is why I mentioned 'performance' overclocking headroom, not 'frequency' overclocking headroom.

 

But that wasn't the point of my comments or any intended criticism therein.

[don_svetlio]

It depends on the situation quite a bit.

 

GM204 for example gets way more relative 'real-world' performance improvements (gm204 is a poster child for invalidating 3dmark as representative for gaming) from overclocking at 4k than it does at 1080p, due mainly in part to memory overclocking reducing a vram bottleneck, just like (to a much lesser extend) the fiji cards overcome their 4GB vram by having an insanely fast bus. Hawaii overclocking however is much more resolution independent (basically granting 3-5 fps no matter what) and since comparably speaking hawaii tends to be better even right now at 4k than gm204 (while being trading blows at 1080p) it is important to be able to see that effect in action.

 

Obviously fiji is awful at overclocking, but Tahiti doesn't seem to be which is the same architecture (obviously different memory solutions, so perhaps a bus interface is a limiting factor. I don't know.)

 

I agree the right card is needed, and there are fewer ultra-premium versions for AMD cards. (Indeed, the current consensus is that the MSI Gaming model while not being the coolest, tends to have the best overclocks, which is amusing.)

 

But back to the issue at hand, I don't use graphics cards at stock. I don't care what they do at stock. Overclocking is different for every architecture and core (but honestly fps differences between the same card in the modern age isn't very large. 50 Mhz on a nvidia card doesn't mean much and +/-35 Mhz on a Hawaii Gpu means only slightly more), so yes, I would prefer personally to only see overclocking results (and overclocked specs so I can compare to other benchmarks that also overclocked.)

Fij =!= Tahiti

 

He did say per clock per core. Per clock ofc amd cards preform higher. They scale more linearly with clock speeds, and they start off with lower clocks. 

 

I honestly couldn't tell you if per clock per core which is higher because the core speed of nvidia gets attacked by the larger cores of AMD. It's a different philosophy which is why I mentioned 'performance' overclocking headroom, not 'frequency' overclocking headroom.

 

But that wasn't the point of my comments or any intended criticism therein.

At any rate - we as consumers only have access to the frequency thus we cannot really change anything else but that - hence why I am using it as a means of comparison - though it's not really fair to Nvidia in that case. MHz per MHz I mean - different architectures an all

[Kinda Bottlenecked]

It depends on the situation quite a bit.

 

GM204 for example gets way more relative 'real-world' performance improvements (gm204 is a poster child for invalidating 3dmark as representative for gaming) from overclocking at 4k than it does at 1080p, due mainly in part to memory overclocking reducing a vram bottleneck, just like (to a much lesser extend) the fiji cards overcome their 4GB vram by having an insanely fast bus. Hawaii overclocking however is much more resolution independent (basically granting 3-5 fps no matter what) and since comparably speaking hawaii tends to be better even right now at 4k than gm204 (while being trading blows at 1080p) it is important to be able to see that effect in action.

 

Obviously fiji is awful at overclocking, but Tahiti doesn't seem to be which is the same architecture (obviously different memory solutions, so perhaps a bus interface is a limiting factor. I don't know.)

 

I agree the right card is needed, and there are fewer ultra-premium versions for AMD cards. (Indeed, the current consensus is that the MSI Gaming model while not being the coolest, tends to have the best overclocks, which is amusing.)

 

But back to the issue at hand, I don't use graphics cards at stock. I don't care what they do at stock. Overclocking is different for every architecture and core (but honestly fps differences between the same card in the modern age isn't very large. 50 Mhz on a nvidia card doesn't mean much and +/-35 Mhz on a Hawaii Gpu means only slightly more), so yes, I would prefer personally to only see overclocking results (and overclocked specs so I can compare to other benchmarks that also overclocked.)

 

Ah, the right card however(if you want to overclock on a Hawaii core) doesn't come with in a 390x. It comes in a 290x. Though a Msi r9 390/x almost guarantees 1200mhz overclocks. It is the vapour-x/lightning 290x models that are better at pushing voltages and even higher clocks. 

 

Sooner or later someone is gonna but out a comparison between OC'ed cards for battlefront. Jay might do it if his users ask for it. 

[Sakkura]

Really...? Set a 970 to 1GHz and a 390 to 1 GHz - guess what - 390 is ahead by 15-20 fucking per cent. It's been like so since early Kepler days.

 

Because the 390 has WAY more shaders, not because it has higher IPC. To establish a higher IPC, you'd have to compare performance per shader/TMU/ROP running at the same clocks. So you have to add some 50% to the GTX 970's score before comparing it to a 390. Which means Maxwell clearly has higher IPC than GCN.

 

You can also compare the GTX 980 to the R9 280X. Both have 2048 shaders and 128 TMUs, but the GTX 980 performs much better. That again demonstrates the higher IPC of the Maxwell architecture.

[Sakkura]

For a 390x that might be the case, but very less likely for a 290x which has much lower clock speeds. Same goes for the 290 and the 970. 

 

The clock speed difference between the 290X and 390X is quite small.

[don_svetlio]

Because the 390 has WAY more shaders, not because it has higher IPC. To establish a higher IPC, you'd have to compare performance per shader/TMU/ROP running at the same clocks. So you have to add some 50% to the GTX 970's score before comparing it to a 390. Which means Maxwell clearly has higher IPC than GCN.

 

You can also compare the GTX 980 to the R9 280X. Both have 2048 shaders and 128 TMUs, but the GTX 980 performs much better. That again demonstrates the higher IPC of the Maxwell architecture.

You forget one major thing - a sizable portion of AMD's hardware is not being utilized thanks to the CPU overhead - the 390 is overbuilt compared to a 970. 970 is a VERY cut-down GPU - Nvidia has been removing anything they can in order to produce less heat and are carried solely by drivers as we speak. Hence the DX12 benches.

280X is GCN 1.0

390 is GCN 1.1

Big difference.

Also - shall we go back to Tera Scale and Fermi? Things were mighty different back then

[Kinda Bottlenecked]

The clock speed difference between the 290X and 390X is quite small.

 

50mhz is no small amount for a Hawaii core 0_0

[Curufinwe_wins]

Fij =!= Tahiti

 

At any rate - we as consumers only have access to the frequency thus we cannot really change anything else but that - hence why I am using it as a means of comparison - though it's not really fair to Nvidia in that case. MHz per MHz I mean - different architectures an all

Sorry Tonga. MB.

[Sakkura]

You forget one major thing - a sizable portion of AMD's hardware is not being utilized thanks to the CPU overhead - the 390 is overbuilt compared to a 970. 970 is a VERY cut-down GPU - Nvidia has been removing anything they can in order to produce less heat and are carried solely by drivers as we speak. Hence the DX12 benches.

280X is GCN 1.0

390 is GCN 1.1

Big difference.

Also - shall we go back to Tera Scale and Fermi? Things were mighty different back then

 

Not really. The GM204 GPU is 398 mm², the Hawaii GPU is 438 mm². That's a much smaller difference. The real difference is that AMD has come up with an architecture that uses a lot of small shaders etc. while Nvidia uses an architecture with a smaller number of shaders etc. that each take up more space and perform better. Those differences largely cancel out.

 

The only thing Nvidia really cut out was double precision. That's why you get seemingly odd results like a freaking GTX 580 beating a Titan X.

[don_svetlio]

Not really. The GM204 GPU is 398 mm², the Hawaii GPU is 438 mm². That's a much smaller difference. The real difference is that AMD has come up with an architecture that uses a lot of small shaders etc. while Nvidia uses an architecture with a smaller number of shaders etc. that each take up more space and perform better. Those differences largely cancel out.

 

The only thing Nvidia really cut out was double precision. That's why you get seemingly odd results like a freaking GTX 580 beating a Titan X.

Not only double precision - anything not directly related to gaming is cut. CUDA core count went down from Kepler, everything went down - transistors. Maxwell is a "fake" efficiency gimmick carried by Nvidia's wizardry software. Simple.

[Curufinwe_wins]

Not only double precision - anything not directly related to gaming is cut. CUDA core count went down from Kepler, everything went down - transistors. Maxwell is a "fake" efficiency gimmick carried by Nvidia's wizardry software. Simple.

IDK about that. GM204 certainly has issues, but GM200 still outclasses GK110 in literally everything but DP (and is basically the same core count as GK110, more transistors etc), and still does impressively well in DX12 (well compared to how shitty all the rest of Maxwell does in DX12)

[WaitForMeImOnDialUp]

A 290x/390x matching a 980? Looks like Nvidia hasn't finalize the drivers for battlefront. 

no dude nvidea released drivers for this game which was used in this benchmark

 

http://www.anandtech.com/show/9698/nvidia-releases-35850-game-ready-drivers-for-star-wars-battlefront

[WaitForMeImOnDialUp]

I just want to point out - AMD have roughly 50% more instructions per clock so 1MHz on an AMD card is about 1.5MHz on an Nvidia competitor - roughly.

this is true..on amd you get more per 1mhz  overclock than you do with nvidea

[Sakkura]

Not only double precision - anything not directly related to gaming is cut. CUDA core count went down from Kepler, everything went down - transistors. Maxwell is a "fake" efficiency gimmick carried by Nvidia's wizardry software. Simple.

 

No. The Titan X has more shaders than the Titan Black (full GM200 vs. full GK110). The GTX 980 has more shaders than the GTX 680 (full GM204 vs. full GK104). GM200 has more transistors than GK110 - 8 vs. 7 billion. It's also bigger - 601 vs. 561 mm².