AMD: Speculation, Rumors, and Discussion (Archive)

[kalelovil]

This probably goes some way towards explaining the really poor overclocking results by some, potentially more variability in boost clock attainability between cards than we are used to:

Using AMD's new overclocking tool, we raised the power threshold to 150 per cent, but only achieved a stable, reliable core clock increase of four per cent, taking us up to 1315MHz.
...
The core clock increase of just four per cent is interesting bearing in mind the larger increases overall. The fact is that with our review card at least, our RX 480 sample doesn't always reach its max 1266MHz during gaming, often levelling out at 1190-1230MHz. The overclock seems to overcome that limitation and then gives you the additional OC on top of the max 1266MHz stock boost clock.

I wonder how this card would have performed with more ROPs, or if the bottleneck is more complicated than that. It seems to be do relatively better at 1080P (a cynic might see that at odds with its VR-heavy promotion).

It is disappointing to see power consumption not improve as much as I was expecting, especially considering AMD appears to have made considerable architectural changes aimed at power-efficiency this time (unlike previous GCN versions).

 

[Esrever]

With how much power this thing already draws at stock, can't really oc it on just a 6 pin.

 

[Erinyes]

Quite disappointing..after all the hype..it basically matches or is slightly ahead of the GTX 970..both in performance AND Perf/watt.

Frequency is definitely a big aspect here..the difference in clock speeds between Pascal and Polaris is the highest ever between NV and AMD. Some part of this may be due to GF's 14nm process being inferior to TSMC 16FF+, but that dosen't explain the entire difference (~35%).

NV already had a ~15% clock speed advantage with Maxwell. For 980 - >1080 the boost clock increased from 1214 mhz to 1733 mhz..about 42%.

For AMD..for 390x -> 480 it went from 1050 mhz to 1266 mhz..barely 20%.

This huge disparity in clocks is going to hurt AMD big time..GP106 should be able to match Polaris 10 pretty easily..and given that its less than 200 mm2, NV will easily be able to compete on price.

Also..AMD setting the minimum DRAM speed for Rx480 at 7 Gbps is a crap move.

 

[Orion]

Power issues disappoint a bit.
Regards performance:

Continued checking reviews
http://www.gamersnexus.net/hwreviews/2496-amd-rx-480-8gb-review-and-benchmark-vs-gtx-970-1070/page-4
In doom difference between a 480 and a 390x is 10~% at 4k ultra and 1440p ultra, at 1080p ultra it shrinks to 5~%.

http://www.hardwarecanucks.com/foru...9-radeon-rx480-8gb-performance-review-18.html
At 1080p highest setting in hitman 480 is 16% faster than a 980, at 1440p highest setting in hitman 480 is 30% faster than 980

 

[CSI PC]

Running dual 480s could be a bit of a worry from a mainboard voltage-current perspective and maybe something that needs more emphasis by publications because the average person was thinking of using 2 of these to compete with a 1070.
They may see-read it going over spec with a single GPU and not think too much of it, without realising that it is shared with other devices including the mGPU cards they were intending to do.
Breakdown at Tom's Hardware:

That is a lot of demand over the mainboard if going dual GPU.
Interestingly Nvidia take a different route with 1070/1080 and focused most of the power requirement through the PCIe connector, at Tom's Hardware they measured a peak of 62W and average of 40W for a 1080FE with the mainboard 12V.
And sadly that also puts into perspective that Nvidia has a strong edge in terms of performance/watt with this generation of cards as Tom's Hardware shows for the same game the 1080FE draws only 10W more on average/peak in total.

That said OCing puts the Pascal cards well out of shape in terms of efficiency-consumption, and so will be interesting with Polaris (expecting the same tbh).
Cheers

 

[silent_guy]

6 months of hype and then this?

My expectations were already pretty low, but Polaris is even worse than that.

One wonders how AMD could have marketed this thing differently. When engineering gives you something like this, there's really not much you can do. Fermi was at least the performance leader...

At least we now know why Vega will improve power efficiency: Polaris just didn't deliver.

 

[gamervivek]

Kyle was on point.

Kyle has another point now.

AIB partners are reporting to me this morning that with extremely good custom air coolers the 480 GPUs are seeing from 1480MHz to 1600MHz clocks, but tell me it is a "lottery draw" on GPUs.

https://hardforum.com/threads/first-gaming-benchmarks-up.1903628/

Decent 1080p card and nothing more. Guru3d has the best one for AMD and PCGH has the best one with overclocking where it tops the MSI 390X in every game at 1080p even going past Fury X on occasion.

 

[Razor1]

Ya know what if they used a 4gb version of the rx 480 for this launch and then came out later with the 8gb versions, I think it would have been better, their power consumption might have been just a tad lower.

 

[msia2k75]

So, the card has 3072 sps?

 

[Rys]

I had posted a much stronger rant about the upthread armchair analysis on the perf/W, swiftly removed to rethink how I make the point. It boils down to this: GPUs are the most difficult semiconductor device type to create these days, because of their size, complexity, feature set and everything else. So there are going to be differences baked into the designs of GPUs from two different vendors, almost across the board, from the microarchitecture all the way through to the physical aspects, never mind the software.

So I don't know what miracles everyone expects. Heck, I don't even know why I'm trying to argue it out. It's just disappointing, knowing how difficult it is to design a functioning GPU that can be shipped into the market, to see armchair analysis take big swings at the hard work of a company that has limited resources and can't afford to take giant risks in the microarchitecture.

We've been discussing GPUs here at Beyond3D for over a decade. We mostly know how it goes. The hype from the vendor ahead of time doesn't really factor in to things. That's marketing. The leaks from the myriad industry-ruining piece of shit news websites doesn't really factor in. They're toxic.

Given the foundry tech available and how their design works, and their inability to just throw it all out, I was expecting the transistor tech to take them most of the way across the line on this one. That's what we got with Pascal. And I think that's what we have here, just with what looks like GF14 doing quite a bit worse for them. That's sad, but that's possibly not their fault given capacity and their plans for volume.

The end result is something with a nice set of features and good perf/$, at competitive perf/W with the existing chips in the segment, that's worth buying.

 

[AlNom]

So, the card has 3072 sps?

? It's 2304

 

[CSI PC]

Kyle has another point now.



https://hardforum.com/threads/first-gaming-benchmarks-up.1903628/

Decent 1080p card and nothing more. Guru3d has the best one for AMD and PCGH has the best one with overclocking where it tops the MSI 390X in every game at 1080p even going past Fury X on occasion.

Here is a big worry, how many of those getting good results including OCing measured a breakdown of the power interfaces individually.
When it came to OC the 480, this is what Tom's Hardware said:

We skipped long-term overclocking and overvolting tests, since the Radeon RX 480’s power consumption through the PCIe slot jumped to an average of 100W, peaking at 200W. We just didn’t want to do that to our test platform.

I also wonder how much demand was also required from the mainboard 12V, and in general how the lottery draw affects these performance characteristics.
Cheers
Edit:
There is a fair bit of leeway I agree with the PCIe connector as the Molex connector spec is designed for a minimum of 8A for the pins (peak if gauge wire is also correct is 192W), even if the spec for PCIe is 2-3A from the 2004 reference spec book.
But still, I would not rely upon a budget board and PSU to guarantee this or a modern board-PSU to sustain high figures.

 

[homerdog]

But is has DX12AsyncCompute guyzzz!!!!

Polaris would have been a decent competitor against Maxwell (lol). This does put into context just how good Pascal/Maxwell are. My hopes of a price drop on the 1070 have been thoroughly dashed even though realistically the 1070 (cut down 314mm^2 chip) should be competing with RX480 (full 232mm^2 chip), the 1070 is way out of the 480's league.

 

[Malo]

But is has DX12AsyncCompute guyzzz!!!!

To me that's actually more important, its efficiency with concurrent execution in modern titles which Nvidia still seems to be lacking in but isn't a big factor for them (yet). In the coming month I'm hoping to see some real in-depth analysis and benchmarks of the 470 as well (is that releasing around same time?). I'm fine with my 970 in my main PC for now and skipping this generation but I think for building 1080p gaming PCs and looking forward to many PC/console developed games coming out, it's going to be a real contender.

Just greatly worried about the power though. Idle twice as high as a 1080? Drawing way above PCIe slot limits? ugh

 

[homerdog]

Radeon has had D3D feature advantages over GeForce many times in the past and it always amounted to nill. I don't expect much to change on that front.

 

[gamervivek]

RX 490 rumors floating around.

 

[Infinisearch]

Is the excess power draw from the PCIe slot dangerous in anyway?
Also any review tell whether or not the 480 has conservative rasterization?
edit - and other D3D12 features?

 

[ninelven]

It's just disappointing, knowing how difficult it is to design a functioning GPU that can be shipped into the market, to see armchair analysis take big swings at the hard work of a company that has limited resources and can't afford to take giant risks in the microarchitecture.

Please.... AMD did this to themselves with all the absurd efficiency claims leading up to the launch. It is marginally better or on par with chips produced on 28nm; it is not. AMD had to know the numbers. The solution is simple... don't hype up your weakest point.

https://www.techpowerup.com/reviews/AMD/RX_480/25.html

Back to the card itself and performance.... I am a bit perplexed at the scaling. It seems to be doing worse than I would have thought at 1080p (what I game at) and better (relatively) with the higher resolutions. I would have thought the higher base clock (compared to 390) would have skewed this the opposite way. There also appear to be minimum frame-rate issues. I'm guessing/hoping these are mostly driver related?

 

[Orion]

Kyle has another point now.



https://hardforum.com/threads/first-gaming-benchmarks-up.1903628/

Decent 1080p card and nothing more. Guru3d has the best one for AMD and PCGH has the best one with overclocking where it tops the MSI 390X in every game at 1080p even going past Fury X on occasion.

http://www.guru3d.com/articles_pages/ashes_of_singularity_directx_12_benchmark_ii_review,7.html
Here's a bench from the end of february with 390x being about 30% faster than a 980ti at 1440p crazy settings, that's about the average gap between the 1080 and the 980ti in most applications. The 480 is within 5-10% sometimes even higher than the 390x depending on the application.

Yet some of the new benches are strangely showing even a regular 980 above the 480 in ashes at crazy settings(take note that the 480 appears up to 30% faster than 980 in other dx12 apps like hitman at 1440p), and many reviewers are ignoring crazy settings(the performance gap between nvidia and amd seems to increase significantly going from high to crazy, at least in the past). That's very strange considering ashes was the title with the greatest amd performance vs nvidia. Has there been massive optimization for this particular title on nvidia's side?

Because unless there was a serious flaw in the past guru3d bench or there's been massive optimization the following

Here's a bench from the end of february with 390x being about 30% faster than a 980ti at 1440p crazy settings, that's about the average gap between the 1080 and the 980ti in most applications. The 480 is within 5-10% sometimes even higher than the 390x depending on the application.

Would've likely put 480 performance near 1080 in this application at these settings, and at 1600Mhz it would have put 480 above 1080, again in this application at these settings.

Stranger still from what I've heard the difference between dx12 and dx11 performance on 1080 is minimal, this includes ashes.
http://www.pcworld.com/article/3071...badass-graphics-card-ever-created.html?page=8

So it seems something has occurred that gives a substantial jump in dx11 performance in apps like ashes for nvidia, a jump so high as to match that obtained by amd from async.

 

[el etro]

Kinda hard to admit it, but RX 470 carries the damage control responsibility now. If RX 470 does as marketed with the performance/watt numbers, there will be a new chance to Polaris shine.

Performance of the card is not the problem(altough it dissapoints me a bit, i wanted to see GTX 980/390x performance at GM206 power levels), problem really is the power consumption of the card. Does anybody familiar with processes technology know if going for the IBM 14 High Performance(SOI process on track at IBM) can help Polaris efficiency?