AMD has been in the headlines for quite a while now since the Computex 2016 all thanks to their Zen reveal, a new architecture for AMD’s microprocessors based on 14nm FinFET with AM4 sockets and also the much anticipated Polaris launch.

Raja-Koduri-SVP-of-Radeon-Technologies-Group-Shows-Off-The-Worlds-Firs...-980x652
source-AMD

The RX480 stole the show when Raja Koduri (Senior Vice President and Chief Architect at Radeon Technologies Group) announced that RX480 will cost mere $199 and will meet the minimum specifications in order to run VR games to fulfill AMD’s ambition of providing cost effective solutions for the “Immersive Era” of gaming and making VR affordable for the masses & help first 100 million customers get VR equipped.

Though not much is still known about the RX480’s specification as AMD gave us limited information but we have started to see leaks reaffirming the speculation that RX 480 is going to provide performance similar to Nvidia’s 980 and may even surpass it, For less than half the price. That’s insane price to performance ratio. However, it doesn’t stop us from speculating about rest of the specification of the GPU. Below you can see the Official Specification AMD provided

amd-rx-480-polaris-4
source-AMD

The Key specification in order to estimate/speculate rest of the details about the card are the Compute Units (CU), Performance in TFLOPS and the fact that it’s GCN based. The RX480 has 36 Compute Units (CU) and due to it being based on GCN, each of those CU will have 64 cores , A simple multiplication tells us the Core count will be 2304 cores.

Now we know the Core count, let’s estimate the clock speeds. In order to do that, first, we need to learn how to calculate the peak performance in FLOPS. There 2 formulas we can use.

1st Formula – (Total number of Cores * 2) * Clock speed in Ghz
2nd Formula-  (ROP * TMU) * Clock speed in Ghz

We know the GPU is at least 5+TFLOPS so putting the value in 1st Formula, we get (2304 * 2) * Clock Speed in Ghz = 5000 GigaFLOPS (5TFLOPS). Simple math will tell you that GPU would need a minimum of 1085mhz on Core in order to achieve 5 TFLOPS of performance with that core count, Similarly, we can check the clock required for the card to hit 6 TLOPS mark and we get 1302mhz. Now obviously if the GPU was capable of 6TFLOPS then AMD would have marketed it 6+TFLOPS but they said 5+TFLOPS which means the Reference GPU on Stock should NOT exceed 1302mhz on Core and should NOT be less than 1085mhz either. We’ve also seen a leak where RX480 was running on 1280-ish MHz so our best guess is 1250 to 1290mhz on Reference Card on Stock.

Now we can move to the ROP and TMU, The GCN architecture has 4 TMU per 1 Compute Unit (CU) and since RX480 has 36 CU, We get to the total number of 144 TMU by multiplying 36 and 4. If you look at our 2nd Formula to calculate FLOPS performance, You’ll see that (ROP * TMU) should be equal to Twice the Core count as seen in 1st Formula, Only then both of the formulae will give the same output.

radeon-480-performance-100663924-orig
source-AMD

So back to Math’s again, 2304 is the core count, Twice of that is 4608. Therefore, (ROP * 144TMU) = 4608. Hence ROP = 4608 / 144 = 32.

So now we have 32 ROP and 144 TMU, with these numbers, We can also calculate the Pixel and Texel Fillrate. Though it’s dependent on Core clock so we’ll calculate them for 1250mhz and 1300mhz to give you some perspective on what we think would be minimum and maximum for the reference GPU on stock frequency. So the Formula for Pixel and Texel fillrate are-

Pixel Fillrate = ROP * Clock speed in Ghz = (32 * 1.250) and (32 * 1.300)
Texel Fillrate = TMU * Clock speed in Ghz = (144 * 1.250) and (144 * 1.300)

So, what do we get? Well, here is a simple table for you to look at.

Untitled_1920x1080
source-AMD

Certainly fascinating, but at the end of the day, It still is mere speculation. We’ll know for sure when the GPU actually launches.

Thanks for reading:

written by Prasoon Tiwari

Published by Utkarsh Tiwari

LEAVE A REPLY

Please enter your comment!
Please enter your name here