Pascal vs Maxwell **same clocks**

[nashathedog]

I think his logic is very basic. Most people could come to similar conclusions by themselves, but the truth is architecture is extremely complicated. Pascal seems to react very differently to Maxwell in DX12 for example. I think Pascal is similar to Maxwell in many ways, but to say they're basically the same is probably way of the mark.

+1,

It may have many similarities (no-one seems to mind the GCN iterations) but it clearly excels in certain ways such as VR and syncronous compute when compared to Maxwell, And why shouldn't it be similar? Maxwell excelled in many areas offering the best overall products through it's lifespan and so far Pascal is doing the same with obvious improvements in the features of the moment.
It may be Maxwell on speed under certain conditions but who really cares if it's still number 1?

 

[Jono8]

I watched a video that basically concluded the reason Nvidia Maxwell and Pascal are more efficient than GCN cards is because GCN has a lot of hardware features that even though they don't get used still require power, Pascal and Maxwell lack these and thus don't require the extra power required to run.

This may explain why no hardware support of Async etc, however should Nvidia add this support to hardware their power efficiency will take a hit like AMD has and you may see closer power consumption between manufacturers.

They concluded by saying if Nvidia add hardware support for a lot of features like AMD have on Volta then expect the power consumption to be higher and the cards to be less efficient.

But even with a sync and in AMD's flagship dx12/async game HItman, the 1070 is still 30% faster than a 480 (stock for stock) and still has around the same power consumption.

http://www.kitguru.net/components/graphic-cards/zardon/sapphire-rx-480-nitro-oc-4gb-8gb-review/30/

 

[Tommy_Here]

My mistake I've just realised my mistake that the Ti also has more rops but then the data compression Pascal has over Maxwell will benefit the 1080 take that of what you will

 

[Doom112]

This is the reason why Polaris will never get OEM wins or laptops wins. Efficiency is the main key for laptops and OEM success, which AMD do not have.

 

[hominid]

If you drive a Ferrari and a Nissan Micra at 60Mph they are both as fast as each other. Save your money, don't buy a Ferrari, a micra is just as fast.

That's one of the most useless comparisons I've ever seen, what the OP posted, and what the video shows, is nothing that can be compared like that AT ALL..

A better comparison would be two car engines ran at the same RPM produce similar power, but one has a higher RPM limit so ultimately can produce more power. Not ideal, but a much better comparison than that rubbish statement you spat out.

Really it is the most pointless comparison possible. Nvidia spent millions on critical path analysis and optimization in order to maximize the instructions per second, which is the only relevant indicator of performance. Critical path optimization will routinely lower IPC.

And still only made it a very small amount faster clock for clock. The real next gen technology from Nvidia comes with Volta.

 

[eddyr]

A bit of bandwagon jumping going on in the thread without watching the video.
The video isn't a criticism of Nvidia, in fact somewhat the opposite. The last few minutes (after methodology and results) are positive to Nvidia and Pascal. He re-identifies the positives currently going for AMD as significant threats to AMD once Nvidia implement them with Volta (namely HBM and even more versatile async support) considering the efficiency of Pascal.

Other bits to note:

He does go on to state why clockvsclock is not a fair comparison.

His intent was to gauge shader throughput which was why he equalised tflops in order to compensate for the reduced number of cores/SMX in the 1080.

Granularity at 1 second doesn't matter so long as it is representative. Tough to say whether it is however.

 

[Rroff]

This is the reason why Polaris will never get OEM wins or laptops wins. Efficiency is the main key for laptops and OEM success, which AMD do not have.

I imagine it is coming - I don't think it is coincidence that the 480 exhibits everything that was rumoured to be problematic about 16nm FF+ a bit over 6 months before Pascal release - ironically looks like 14nm at GF is actually ~6 months behind TSMC in maturity and not the other way around as some insisted.

If AMD play their cards right they could make a killing on the seasonal period sales.

 

[rossiboy001]

Yes it would be interesting to see. A clock for clock performance test would show if the extra performance of the 1080 is more down to the higher speed, rather than all the others factors. To many thinking of upgrading from a 980ti would know if it was the case that really what they are spending money is just on something that clocks faster but not better archtecture.

 

[SiDeards73]

Interesting. Got a link to that vid?

Here ya go buddy

https://www.youtube.com/watch?v=Xuea5nt4TnY

Was a pretty good watch, i dont like the guy who presents it, style, something about his accent just grates me up the wrong way lol, but its a decent video.

 

[Seanspeed]

Yes it would be interesting to see. A clock for clock performance test would show if the extra performance of the 1080 is more down to the higher speed, rather than all the others factors. To many thinking of upgrading from a 980ti would know if it was the case that really what they are spending money is just on something that clocks faster but not better archtecture.

I feel like so many are completely oblivious to the fact that clock speeds and architecture are directly linked.

It wouldn't clock as high as it does if the architecture wasn't improved.

It may not be a full architectural jump ala Kepler->Maxwell, but we knew that already. But it is incorrect to say it is nothing more than shrunk Maxwell. It would not have improved DX12 performance, if you need proof of that. A die shrink alone would not cause that.

 

[Gregster]

Here ya go buddy

https://www.youtube.com/watch?v=Xuea5nt4TnY

Was a pretty good watch, i dont like the guy who presents it, style, something about his accent just grates me up the wrong way lol, but its a decent video.

A good watch that and he does pick on some valid points. He also picks up on the API overhead that AMD suffer with on DX11 and with more cores, it doesn't give the performance that they should get and power efficiency on Maxwell and Pascal being so far ahead.

Cheers.

 

[Seanspeed]

Here ya go buddy

https://www.youtube.com/watch?v=Xuea5nt4TnY

Was a pretty good watch, i dont like the guy who presents it, style, something about his accent just grates me up the wrong way lol, but its a decent video.

It's the very 'pitchy' out-loud reading voice(for me at least). It's normal to pitch words more when reading something out loud in an effort to provide emphasis and color so you dont end up with a boring, droning commentary, but many people way overdo it. It ends up feeling highly unnatural and while you may know something is scripted, you ideally dont want it made quite so painfully obvious. It is preferable to try and at least disguise it a bit to make it feel more flowing and natural.

 

[SiDeards73]

A good watch that and he does pick on some valid points. He also picks up on the API overhead that AMD suffer with on DX11 and with more cores, it doesn't give the performance that they should get and power efficiency on Maxwell and Pascal being so far ahead.

Cheers.

Yeah, its interesting that AMD took such a huge risk with GCN by putting all their eggs into that particular basket, i guess they played the long game, while it has crucified them in the interim...

Now with the new API's they can actually leverage their hardware, while this might sell new GPU's i bet a lot of people will hesitate to upgrade if they get a new lease of life from old cards, the only sticker though is VRAM amounts, which will invariably force peoples hands anyhow.

Wonder why AMD decided to not support Command Lists as well as Async in the hardware? perhaps there is a conflict there?

 

[eddyr]

Thanks for the vid SiDeards from HarwareUnboxed on the marginal contribution from GCN async in current implementations - where in the reality its the alleviation of AMD's driver overhead in Vulkan/DX12 that's responsible. Should help add some context to all the async back and forth.

In a similar vein, AMDTV ...errr.... AdoredTV also analysed Doom Vulkan results and identified a possible ~5% contribution from AMD's Async implementation under Vulkan.
Sadly his bias did make him say:

@13.55

"All this stuff, Vulkan getting rid of the CPU overhead? Polaris should have done that anyway without the need for Vulkan. AMD has known about GCN's CPU overhead issue for 4.5 years and yet Polaris still has it. For me that's just not good enough. RX480 should be sitting at 80-90 fps anyway, even in OpenGL.......So really, AMD, you really just need to get the finger out and get these things right."

 

[David Bisset]

Indeed. It has taken AMD nearly two years just to bring something out that matches the 970/980 in perf per watt.

It is slightly concerning in terms of what they can do to counter GP102 but hopefully vega has even further improved efficiency over POlaris and isn't just a scaled up Polaris. - HBM will obviously be used to alleviate the efficiency disadvantage as well (like with the fury range).

Nano beat the 970 in perf/watt

 

[Jono8]

Nano beat the 970 in perf/watt

Yes ,because of HBM, as i mentioned in my post.

Polaris, using conventional GDDR5, has only managed to match ~970/980 perf per watt which is a shame considering the 970 and 980 came out 22 months ago.

 

[Tommy_Here]

Yes ,because of HBM, as i mentioned in my post.

Polaris, using conventional GDDR5, has only managed to match ~970/980 perf per watt which is a shame considering the 970 and 980 came out 22 months ago.

And it's on 28nm

 

[Telecaster]

I wish AMD would get their arse into gear and release something proper, to give Nv an excuse to actually TRY.

 

[D.P.]

That's one of the most useless comparisons I've ever seen, what the OP posted, and what the video shows, is nothing that can be compared like that AT ALL..

A better comparison would be two car engines ran at the same RPM produce similar power, but one has a higher RPM limit so ultimately can produce more power. Not ideal, but a much better comparison than that rubbish statement you spat out.

Your comparison is utterly flawed. Under-clocking cards has absolutely no relevance to performance of the final product, so comparing 22 cars being forced to travel at the same speed is absolutely perfect analogy to forced under-clocking.

And still only made it a very small amount faster clock for clock. The real next gen technology from Nvidia comes with Volta.

What the hell has performance per clock got to do with anything? The only relevant factor is instructions per second, whcih Nvidia has icnreased massivle due to critical path optimization. I'm actually sursied the IPC of PAscal is as high as it is.



The bottom line is NVidia achieved a 25% performance increase in the Pascal architecture outwith any benefits from the shrunk node. AMD official made 15% improvement increases, the rest coming from the new process node. That is why the performance per watt advantage Nvidia has has increased to the point where the 1070 uses less power than the RX480 for nearly twice the performance.


If Nvidia's 25% improvement is not worthy of being called a new architecture then what is AMD's 15% improvement?

 

[Tommy_Here]

I wish AMD would get their arse into gear and release something proper, to give Nv an excuse to actually TRY.

I totally agree at least the 8gb 480 isn't appealing at all at stock it's a 970 on steroids overclock them both and they're both equal with the same performance:watt 28nm vs 14nm 2014 vs 2016 £275 vs £249 for the 480