• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Radeon RX 480 "Polaris" Launched at $199

kalniel said:
No. 1266Mhz is a higher clock speed than 1050, not a cut. Fewer shaders, but they run at higher frequency. IPC stands for instructions per cycle, so is lower on the 480 once clock speed is equalised.
Click to expand...

390X: 2816 Shaders at 1050Mhz
480: 2304 Shaders at 1266Mhz

2304 Shaders + 22.2% = 2816 Shaders
1050Mhz + 20.5% = 1266Mhz

So bar 2% the performance from Shaders + Mhz vs Shaders + Mhz is a 390X.
But that's assuming no architectural changes, which we know Polaris has a lot, that doesn't mean it will be any faster per Shader / Mhz, but if not, why bother?

it could also go wrong and end up much slower.
 
Tris1066 said:
You know I had no idea that's how TFLOPS were calculated (shaders*clock*2). If that's the case then my 1070 is hitting a hair over 7.9TFLOPS :)

Can't wait to see how far the 480 can be pushed :D
Click to expand...

Technically it's operations per cycle * cycles per second = operations per second. Does what it says on the tin :p

When talking about single precision floating operations, all modern shaders can carry out two per cycle, hence the shaders*2 bit, giving the equation above. When people specify double precision OPs the figure can drop massively as cards perform significantly fewer operations per cycle, sometimes less than one.
 
humbug said:
390X: 2816 Shaders at 1050Mhz
480: 2304 Shaders at 1266Mhz

2304 Shaders + 22.2% = 2816 Shaders
1050Mhz + 20.5% = 1266Mhz

So bar 2% the performance from Shaders + Mhz vs Shaders + Mhz is a 390X.
But that's assuming no architectural changes, which we know Polaris has a lot, that doesn't mean it will be any faster per Shader / Mhz, but if not, why bother?

it could also go wrong and end up much slower.
Click to expand...
There won't be any architecture changes that affect FLOPS, it's too fundamental and the figures match what AMD have already supplied (5+).

Now that doesn't say ANYTHING solid about game performance - TFLOPs never has. But as to why bother? Well you couldn't buy a 390X in volume at the price the 480 is, nor does it fit in the same TDP envelope. So even if there are no architectural improvements that result in better game performance, you're getting a cheaper, cooler and smaller card for the same performance.
 
Last edited:
it is interesting how deep folks are willing to swim in the name of speculation, and how well debated it can become with the tiniest glimmer of light.. well, i say interesting, what i really mean is interesting and thankful.. nothing helps pass the time, and keep me grounded, like a well balanced discussion.
 
D.P. said:
No, it just the clock speed difference:
390x = 2816*1050*2= 5.9TFLOPS
480 = 2304*1266*2=5.8TFLOPS

IPC is lower on the 480.
Click to expand...

That would be horrible.


Here's a 390x steamvr performance result:

vr1.png



I tried doing an underclock to test relative performance if the underclocked 480 theory proves true, but steamVR keeps forcing my default clock even when I do so.
 
kalniel said:
Technically it's operations per cycle * cycles per second = operations per second. Does what it says on the tin :p

When talking about single precision floating operations, all modern shaders can carry out two per cycle, hence the shaders*2 bit, giving the equation above. When people specify double precision OPs the figure can drop massively as cards perform significantly fewer operations per cycle, sometimes less than one.
Click to expand...

Still think it's awesome that when I was young (waaay back in the 80s) I read about the world's most powerful supercomputer having just over 2tflops peak compute performance and thought that was amazing!

Not exactly the same since that would have been cpu not gpu but still cool ;)

Edit* it might even have been gflops it was a loooong time ago!
 
Last edited:
D.P. said:
No, it just the clock speed difference:
390x = 2816*1050*2= 5.9TFLOPS
480 = 2304*1266*2=5.8TFLOPS

IPC is lower on the 480.
Click to expand...

Hmm...ok, but the cut in TMUs and Compute Units whilst still rumoured to get 390x performance surely signifies that the RX480 is doing more with less and at less power?
 
kalniel said:
There won't be any architecture changes that affect FLOPS, it's too fundamental and the figures match what AMD have already supplied (5+).

Now that doesn't say ANYTHING solid about game performance - TFLOPs never has.
Click to expand...

I know, TFlops is just a theoretical performance calculation based on Shaders and Mhz.
Its not the be all and end all of performance.

Look at the 980 for example, it has 5 TFlops to the 390X 5.9 and yet they are very similar in performance.
The difference is the 980 has better tessellation culling, it has a better front end so can run more of its Shaders A-synchronously, at least in DX11.
 
Last edited:
cib24 said:
Hmm...ok, but the cut in TMUs and Compute Units whilst still rumoured to get 390x performance surely signifies that the RX480 is doing more with less and at less power?
Click to expand...

TMUS:

176*1050 = 184800
144*1266 = 182304



The memory bandwidth reduction is quite significant, 512bt down to 256bit, but clocks speeds are 335 faster. I expect there is a lot of technologies like delta compression going no that are improving memory bandwidth utilization. AMD ave always required greater memory bandwidth than nvidia so there was obviously some additional optimizations to be done. Even with Pascal Nvidia have made some good improvements, compare the 1080 memory bandwidth to the 980ti. The 1080 is a much faster card but gets away with less bandwidth due to improved delta color compression and L2 cache.
 
Last edited:
D.P. said:
Try telling that to people that think pascal is just a die shrunk Maxwell:p
Click to expand...

Besides some front end changes for compression and their new VR guff, GP104 Pascal appears to be no different from Maxwell in terms of the CUDA core layout. The only Pascal part that diverges to any great extent is GP100.

Nvidia have always been able to feed their cores well, which is where most of the GCN4/Polaris improvements revolve around. Although we still need to hear what some of the enhancements do and what they help with.
 
SkeeterUK said:
Btw will these 480s be any good in dx11 titles or is nvidia still better with dx11 titles? As usually amd was lacking in scores vs nvidia in dx11 games.
Click to expand...

A few of the front end enhancements added to GCN4/Polaris are to help with single threaded performance and subsequant shader utilisation. So it should perform better in DX11 games compared to equivalent GCN1-3 parts, although it has been hinted that it will perform best in DX12/vulkan etc.
 
Last edited:
Just seen that another retailer has listed the 8GB model at just over £217. Not sure how reliable a source that will prove to be, but it's about what was expected I guess.
 
I think Pascal is a slightly enhanced Maxwell, nothing wrong with that, Maxwell was and still is a good architecture.
 
Kaapstad said:
I can't wait for the 29th June so we can see some proper reviews and find out the real performance.
Click to expand...

+1 tbh, cant wait for the whole drama to be over and to see if the cards are worth buying or not, desperate to get rid of my 290 for something else from AMD, even more so now i just bought a Freesync screen..

Got a feeling im going to be waiting for Vega though
 
You must log in or register to reply here.
Back
Top Bottom