Originally Posted by Mahigan
I was talking about Vega and not Polaris. I speculated that Polaris 10 would output around FuryX performance. The RX 480 appears to be around 390x performance which means that the 480x will probably hit the Fury X performance figures.
There are two Vega SKUs. One which is set to compete with GP104 and the other with GP102. Think about it... RX 480 has around 256GB per sec of memory bandwidth and is able to keep up with a card that has around 384GB per sec of memory bandwidth performance. The RX 480 has 2304 cores and is able to compete with a card which has 2816 cores. So I mentioned a 30 to 40% IPC increase for Polaris 10 (a particular user here disagreed with me). AMD showed a 1.8x performance per watt figure due to Finfet followed by a 2.8x figure due to architectural improvements. Basically... AMD are agreeing with what I was saying. You can see this by comparing the amount of memory bandwidth needed for both.
5500 Gflops with 256GB per second vs 5900 Gflops with 384GB per second. That is 21.5 Gflops per 1GB per s of memory bandwidth matching 15.4 Gflops per 1GB per s of memory bandwidth. That means an architectural efficiency boost of 140% or 40% boost. That is HUGE. It allows AMD to circumvent one of their largest architectural flaws (the reason Hawaii was using a 512bit bus in the first place). All that for under 150W (a single 6 pin with motherboard power offers around 150W max). The full Polaris 10 (RX 480x) will come with 2560 cores vs 2304 cores for the RX 480.
Big Vega will likely pack around 4096 cores whereas the smaller Vega (RX 490 and 490x) will pack less (between 2560 and 4096).
So AMD have released a very competitive part. At $199 ($229 for the 8GB model)... WOW.
You are misquoting me a little, mate, but whatever. Let's start:
1) RX480 is under 390x performance. For full part to hit Fury X we talk about ~20%+ difference between cut and full chip. Not even 1080-1070 difference is THAT massive. Note that it works off of assumption that RX480 is cut chip. Somewhat verified by Dr Lisa, but nonetheless.
2) Vega is not coming until 2017, just saying. Now onto, your statement about bandwidth: Rx480 with 256MB/sec of bandwidth so far is slightly under GTX 980 with 224MB/sec of bandwidth. Not so impressive now, huh? Next, RX480 has 5.8TFLops of peak compute and slightly loses to 6TFLops of peak compute (390x) in 3dmark11 (don't get me started on GTX980).
3) So far your IPC increase speculation does not quite work out, but we'll see when drivers are ironed out.
4) Perf/watt claim disagree with their own TDP figure. This card would have under 130TDP if it really was a 2.8x perf/watt improvement over any currently produced AMD GPU. I dare say AMD are pulling JHH here with stating figure for niche cases (like 10x performance for Pascal, remember that?)
5) Rest are basic facts, so no point contesting these.