Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

AMD Confirms RX 480 At $199 USD, Other APU & Polaris Announcements

Even INT8 it seems is not proven to be enough, so FP16 would be it. But I can't talk about what I don't know of course.
Like Mago says, some scenarios still require FP64.
But it's a good thing, GPUs keep getting leaner.
 
Mago said:
Allways the same things that require FP64 can be done on a GPU w/o FP64 by just an simple compiler optmization, but running a lot of FP64 operations on an FP32 systems requires 20x more fp32 instructions plus a number of memory registers so hopefully you can get 1:40 FP64 on an FP32-only system.

FP64 maybe not necessary on AI and non HDR video rendering but lots of SCI/ENG apps requires it, as some HI Color HDR rendering too.
Click to expand...
In other words, Nvidia has been brilliant with the GP10x market positioning - GP100 with great FP64 (and a great price tag), and the other GP10x for great FP32 and FP16.

And why is HDR an FP64 application? FP32 should be able to do 20-bit per pixel arithmetic with very good accuracy
 
AidenShaw said:
And why is HDR an FP64 application? FP32 should be able to do 20-bit per pixel arithmetic with very good accuracy
Click to expand...
Pixel are integers (30bpp in HDR, but its usual to work original content in 48bpp) but raytraces require FP operation if you dont want to carry color degradation, as with audio processing if you work on 60khz product your samples should be at 120khz.
 
Mago said:
Pixel are integers (30bpp in HDR, but its usual to work original content in 48bpp)...
Click to expand...
In other words, 10 bits per subpixel to 16 bits per subpixel. FP32 has about 24 bits of precision....

Mago said:
... but raytraces require FP operation if you dont want to carry color degradation, as with audio processing if you work on 60khz product your samples should be at 120khz.
Click to expand...
Someone should tell Pixar that those GTX Maxwells are degrading their color. ;)

Why bring up kHz for audio?
 
ManuelGomes said:
PCIe 4.0 in the works, ....
Click to expand...

PCI-e 4 has been in the works .....

http://www.theregister.co.uk/2014/0...patible_pcie_40_on_track_for_del2015del_2016/

that slide from 2015 to 2016 probably means it missed the window to get included into any CPU designs for the next 6-12 months. If the standard isn't fixed can't particularly tape out. [ The original Xeon E5 products got hung up waiting on PCIe v3 even though nominally done but the testing/infrastuture wasn't flushed out. No solid indications that PCIe 4 isn't in the same state.... given was sliding two years ago. And 1-1.5 years ago no one would sanely bet the farm it was going to be on time having already slid. ]

It is like DisplayPort 4 .... any monitors right now ? No. Have DPv4 "ready" cards.
 
AMD hasn't fixed the problem, they just covered it a bit. The card still goes upwards of the announced 150W.
You call that fixing? I don't. I call it throwing sand into your eyes. Sure, most AIBs have changed the connector to 8pin so that no more out of spec claims are in order, but still current draw is higher than the spec, although most mobos will not burn. they shifted the current draw from one place to the other, so that the slot can be relieved. Come on...

PCIe 4 is now at 0.7 coming in at 0.9 and final next year. OK, no devices for it yet, but once it's out maybe it gets traction.
 
Well, AMD underestimated the core clocks Nvidia would be able to achieve on 16 nm Process with TSMC, and they designed the GPU to run at max 1120 MHz, and 6000 MHz on memory. At that clocks whole GPU is drawing around 110W-125W, because of much lower voltage on core, and 8-9W less on memory.

When they found out, about the core clocks of GTX 1070, they went into panic mode ;). They had to increase the clocks, and pushed the GPUs way out of their comfort zone. They have overdone it, especially with voltages. You can already increase the performance of the GPU, and lower the power consumption just by lowering the core voltage.

It is actually quite funny that nobody have spotted that there is so far no mobile GPUs based on Polaris architecture, available anywhere ;).

Compare this. 1024 GCN core GPU - power consumption 35W.
896 GCN core GPU - power consumption 80W.

You get the picture, what happened with AMD down the line.
 
koyoot said:
Well, AMD underestimated the core clocks Nvidia would be able to achieve on 16 nm Process with TSMC, and they designed the GPU to run at max 1120 MHz, and 6000 MHz on memory. At that clocks whole GPU is drawing around 110W-125W, because of much lower voltage on core, and 8-9W less on memory.

When they found out, about the core clocks of GTX 1070, they went into panic mode ;). They had to increase the clocks, and pushed the GPUs way out of their comfort zone. They have overdone it, especially with voltages. You can already increase the performance of the GPU, and lower the power consumption just by lowering the core voltage.

It is actually quite funny that nobody have spotted that there is so far no mobile GPUs based on Polaris architecture, available anywhere ;).

Compare this. 1024 GCN core GPU - power consumption 35W.
896 GCN core GPU - power consumption 80W.

You get the picture, what happened with AMD down the line.
Click to expand...

This is silly. The RX 480 was never going to compete with the GTX 1070. The excessive power of the RX 480 is the clearest sign that it didn't meet AMD's performance/efficiency targets. They had to increase the clocks to compete with the GTX 970.
 
  • Like
Reactions: tuxon86
AMD was targeting RX 480 at GTX 1060. They know how to estimate performance of Nvidia GPUs, Nvidia knows how to estimate performance of AMD GPUs, and Intel knows how to estimate performance of AMD CPUs. Its called reversed engineering.

Like M. Papermaster have said: "AMD is paranoid about competition".

RX 480 is faster than GTX 970 in DX11, it is much faster in DX12, and has much more compute power than GTX 970. Even with lower clocks it would be faster.
 
koyoot said:
RX 480 is faster than GTX 970 in DX11, it is much faster in DX12, and has much more compute power than GTX 970. Even with lower clocks it would be faster.
Click to expand...

And yet they're still embarrassingly behind in terms of perf/watt compared with Pascal.
 
Asgorath said:
And yet they're still embarrassingly behind in terms of perf/watt compared with Pascal.
Click to expand...
RX 480 - 35 GFLOPs/watt.
GTX 1060 - 39 GFLOps/watt.
GTX 1080 - 44 GFLOPs/watt.
Titan X - 44 GFLOPs/watt.
GTX 1070 - 47.5 GFLOPs/watt.

So GPUs from similar performance/price bracket are divided by just 4 GFLOPs/watt. And I did not included RX 470 which, reference model, has exactly the same efficiency as GTX 1060.

But, you are free to be entitled to your opinion(s).
 
koyoot said:
RX 480 - 35 GFLOPs/watt.
GTX 1060 - 39 GFLOps/watt.
GTX 1080 - 44 GFLOPs/watt.
Titan X - 44 GFLOPs/watt.
GTX 1070 - 47.5 GFLOPs/watt.

So GPUs from similar performance/price bracket are divided by just 4 GFLOPs/watt. And I did not included RX 470 which, reference model, has exactly the same efficiency as GTX 1060.

But, you are free to be entitled to your opinion(s).
Click to expand...

You can keep quoting those theoretical numbers, but they don't line up with reality in real applications. The Polaris GPUs have way more raw horsepower than the Pascal GPUs, yet somehow the Pascal GPUs win in real-world tests (often by large margins).
 
Asgorath said:
You can keep quoting those theoretical numbers, but they don't line up with reality in real applications. The Polaris GPUs have way more raw horsepower than the Pascal GPUs, yet somehow the Pascal GPUs win in real-world tests (often by large margins).
Click to expand...
In compute? They do show that those "theoretical" numbers are actual numbers. But as I have said, you can believe whatever you want to believe. I can bring even benchmarks, but I already know what you will say about them. You have already stated your view on them in the past.

As for gaming benchmarks it is question of optimization of drivers, and APIs. Few months ago R9 390X was not even close to GTX 980 Ti, and yet there is more and more games, that show that both GPUs are very close to each other(5-10%). How come? Because both GPUs have similar compute power.
 
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back