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

[Zarniwoop]

GV100 is quite a chip, 120 TFLOPS of dedicated Tensor operations, woot!? And some GPGPU improvements. But Nvidia's' secret sauce is once again in the software, where the magic is created.

How much this will improve graphics is to be seen. GV100 seems to be 42% more efficient. Performance per Stream processor per clock is same as with Pascal architecture. But there are 43% more of stream processors on GV100 compared to GP100, and 42% more computational power, but same 300W TDP. So, Volta is more efficient and has a new Tensore operations, which is used mainly for machine learning. Did I miss something else? Oh yeah, it could be that the size matters with Volta, and so with smaller chips with less cores and higher clocks the efficiency is lost. But we'll see.

I have a fear, that Vega is going the path of Fury and Polaris... AMD is forced to overclock it to compete with Nvidia, and it'll lose the efficiency... once again.

AMD needs an OS that supports HSA... but Linux is too small market, and Apple seems to continue with Intel... so... no HSA, bad for AMD, bad for competition.

 

[SoyCapitanSoyCapitan]

RX 480 (unsupported driver) killing the GTX 1080 (beta driver) in OpenCL video rendering tasks by a wide margin

 

[koyoot]

https://videocardz.com/69424/amd-to-hold-a-press-conference-on-may-31st

Last year, at similar conference AMD announced RX 480 GPUs.

 

[ManuelGomes]

Seems there will be a dual GPU Vega card, liquid cooled. Spotted in Linux drivers.
A monster too probably, but at what power draw?

 

[koyoot]

Seems there will be a dual GPU Vega card, liquid cooled. Spotted in Linux drivers.
A monster too probably, but at what power draw?

300W TDP.

Expect around 20 TFLOPs of compute power FP32, and 40 TFLOPs FP16.

 

[Asgorath]

300W TDP.

Expect around 20 TFLOPs of compute power FP32, and 40 TFLOPs FP16.

I'm not sure how you're getting those numbers. The high-end Vega is supposed to be around 12 TFLOPs, right? So, 20 is nearly double that, which means a single high-end Vega must be (12/20) * 300 = 180W? I find that hard to believe, given that the RX 580 already uses that much power.

 

[koyoot]

12.5 TFLOPs is ONLY for MI25 GPU.

 

[Asgorath]

12.5 TFLOPs is ONLY for MI25 GPU.

Okay, but in order to hit 20 TFLOPs at 300W they're going to have a card that does 10 TFLOPs at 150W? I still find that hard to believe. What makes you think they're going to have 20 TFLOPs at 300W exactly?

 

[koyoot]

Okay, but in order to hit 20 TFLOPs at 300W they're going to have a card that does 10 TFLOPs at 150W? I still find that hard to believe. What makes you think they're going to have 20 TFLOPs at 300W exactly?

R9 Nano has 7 TFLOPs in 185W power consumption envelope. And this is on 28 nm process.

This is all I will say about this.

 

[Asgorath]

R9 Nano has 7 TFLOPs in 185W power consumption envelope. And this is on 28 nm process.

This is all I will say about this.

Okay, thanks for posting information without any data to back it up then. A quick search suggests the RX 580 gets 6.17 TFLOPs at 185W as well, using the 14nm process. So, somehow AMD has found a way to create a less efficient GPU going to the next process node? That doesn't give me any confidence in your suggested 20 TFLOPs at 300W using 2 Vega GPUs, so I'll just ignore your speculation.

 

[koyoot]

Okay, thanks for posting information without any data to back it up then. A quick search suggests the RX 580 gets 6.17 TFLOPs at 185W as well, using the 14nm process. So, somehow AMD has found a way to create a less efficient GPU going to the next process node? That doesn't give me any confidence in your suggested 20 TFLOPs at 300W using 2 Vega GPUs, so I'll just ignore your speculation.

HBM2 consumes at peak power 8W of power, and memory subsystem (memory controllers, TSV's, etc) consume 15W Total. Memory cells in RX 580 consume 36-38W alone. Memory subsystem can consume up to 50W, alone(there is a reason why MSI afterburner shows 130-140W power consumption for GPU die, alone).

Vega design is different than Polaris. Both in power design, voltage/frequency curve, Register File handling. Lets just say that they do not have a lot in common.

 

[ManuelGomes]

Hope you're right about Vega.
I also have a hard time with those numbers but what you say makes sense of course.

 

[Zarniwoop]

Between Fury Nano and Fury X there was just around 8% speed difference, but 60% power consumption difference... same chip, one is overclocked without much power gating, another is run on optimal settings.

Same is with Polaris products... AMD marketing dept. is demanding factory overclocking to get few bucks more out of the chip. The result is couple of FPS more in some games and a lost efficiency.

If Vega is not able to catch Nvidia, we're going to see same thing happening again...

 

[koyoot]

There are leaked two SKUs.

If top Vega has 1600 MHz, with 4096 GCN cores, that makes it 13.2 TFLOPs of compute power, at around 250W TDP.
How much power you will get from that GPU if you will lock it to 150W?

 

[ManuelGomes]

Possibly around 8TFlops? Not too shabby but it might not scale like that though.
Let's see if those numbers hold.
[doublepost=1494592598][/doublepost]Will we finally have details on Vega at the end of the month at Computex? May 31st seems to be the V (Vega) day.

 

[Asgorath]

RX 480 (unsupported driver) killing the GTX 1080 (beta driver) in OpenCL video rendering tasks by a wide margin

Hey guess what, Compressor isn't using the NVIDIA drivers at all, it's doing the transcoding on the CPU only. I would bet that it's using the AMD driver they developed for the nMP though, assuming that their video encoders didn't change much between the D*00 and Polaris.

 

[cube]

The RX 560 is out.

So far there seems to be no benchmarks of the 1050 against the 560 2GiB, but the RX 560 4GiB is only $5 more expensive than the 1050 (2GiB).

In this case, the 1050 is winning with traditional APIs, while the 560 is winning with the new low level APIs.

The 560 is beating the 1050 Ti (4GiB) in some cases. The latter is quite a bit more expensive.

http://www.gamersnexus.net/hwreviews/2908-sapphire-rx-560-pulse-oc-review-vs-gtx-1050

 

[Stacc]

We are starting to get more rumors about Vega. Rumors have 3 versions of Vega coming in the near future.

"Core" at $400
"Eclipse" at $500
"Nova" at $600

Going on prices alone compared to nvidia's offerings we can guess that the core sits between the GTX 1070 and 1080. The Eclipse offers better performance for the same price as the GTX 1080 and the Nova is either competitive or slightly worse than the 1080 Ti.

Rumors have the announcement as May 31st and the release June 5th. Lets hope one of these shows up in a mac sometime soon.

 

[ManuelGomes]

And this:
https://videocardz.com/69475/amd-radeon-vega-spotted-with-16gb-memory-and-1600-mhz-clock

HSAIL?

 

[Stacc]

And this:
https://videocardz.com/69475/amd-radeon-vega-spotted-with-16gb-memory-and-1600-mhz-clock

This is probably the HPC version of the card, especially given the 16 GB of memory. It is promising though that it can hit 1600 Mhz.

 

[owbp]

I would bet that it's using the AMD driver they developed for the nMP though,

+1
And that's why OpenCL video rendering test would be much more realistic if done in some third party NLE, not FCPX.

I've just moved from R9 280X to GTX 980 in my cMP4,1 (R9 died on me) and i can say that editing in FCPX is making my GTX980 "spike" in load the same way it does when watching YT video on my monitor and at the same time projecting a movie on TV. So it really does almost nothing, while R9 280X and older HD7950 were almost always at full load during editing work (especially background rendering and exporting).

So that video is great representation of Apple's optimization at work, not beta vs non-supported drivers, IMHO.

 

[Asgorath]

So that video is great representation of Apple's optimization at work, not beta vs non-supported drivers, IMHO.

Or Apple intentionally making changes to make the NVIDIA cards look bad (e.g. forcing CPU encoding despite there being a massively powerful GPU sitting right there).

 

[cube]

Clean sweep here. Improved NVIDIA driver or OC differences?

 

[cube]

I wonder is the games programmed with low-level APIs are already using the extra RX 560 Compute Units properly.

 

[Asgorath]

I wonder is the games programmed with low-level APIs are already using the extra RX 560 Compute Units properly.

It doesn't work like that.