SFF.Network [SFF Network] AMD Announces the Radeon RX 480: 5 TFLOPS and 4GB for $199
- Thread starter PlayfulPhoenix
- Start date
If you can read anything into the 5 TFLOPS performance, the RX 480 costs $40 per TFLOP compared to $58 per TFLOPS of the GTX 1070 and $66 for the 1080.
FLOPS doesn't necessarily translate into real performance. For instance, when fused multiply-add (FMA) ALUs were being added, a lot of processor vendors claimed 2x FLOPS, but the workload would have to be a theoretical one of exclusively FMA operations.
A nice card with an "ITX" sized PCB and only 6-pin power requirement. I'm glad they're putting the focus on Crossfire, that way my R9 290X project is not an effort in futility. Although the 150W vs 275W for basically the same performance is a little harder to swallow for that ye olde card.
So going by the figures in your article
RX 480 – 1267 * 2304 * 2 = ~5.8 TFLOPS which compares to the R9 380 – 970 * 1792 * 2 = ~3.5 TFLOPS pretty favourably.
With additional per clock improvements in performance in GCN 4 it may be near double the performance of the previous $199 card. Pretty nice. Hopefully we'll get an ITX version close to launch.
RX 480 – 1267 * 2304 * 2 = ~5.8 TFLOPS which compares to the R9 380 – 970 * 1792 * 2 = ~3.5 TFLOPS pretty favourably.
With additional per clock improvements in performance in GCN 4 it may be near double the performance of the previous $199 card. Pretty nice. Hopefully we'll get an ITX version close to launch.
Not sure where AMD got their efficiency improvement factors from. A R9 Nano has 175W TDP, while this RX 480 has 150W. R9 Nano seems slightly faster, although we still need benchmarks to confirm. However, based on those numbers, it seems to me that the shrink to 14nm gained them nothing?
- Feb 22, 2015
- 1,052
- 1,990
You'd need to know the die size to make a direct comparison. My presumption (which I think is a pretty safe one) is that AMD's gone with a significantly smaller die (saving money), that's then pushed to the absolute limit in order to dump all of the process improvements into performance rather than efficiency.
NVIDIA, comparatively, focused on performance but also translated some of the efficiency gains into energy savings - the consumption per core of Pascal is lower than Maxwell, but not by much, and the TDP of the 1080 (for example) is actually a tad higher than that of the 980.
Yeah the die of the RX 480 is quite small. Some numbers I found.
R9 Nano: 596 mm² die size, and 8,900 million transistors, 175W TDP
RX 480: 232 mm² die size (rumoured), ? million transistors, 150W TDP
GTX 1080: 314 mm², 7,200 million transistors, 180W TDP
I think what you say makes sense, although I am not sure how the physics formulas work here. I had expected lower TDP for the RX 480, given the performance benchmarks, and these 1.7x 2.0x, 2.5x, 2.8x efficiency factors they have been giving. It isn't particularly obvious how this generation cards live up to those numbers.
R9 Nano: 596 mm² die size, and 8,900 million transistors, 175W TDP
RX 480: 232 mm² die size (rumoured), ? million transistors, 150W TDP
GTX 1080: 314 mm², 7,200 million transistors, 180W TDP
I think what you say makes sense, although I am not sure how the physics formulas work here. I had expected lower TDP for the RX 480, given the performance benchmarks, and these 1.7x 2.0x, 2.5x, 2.8x efficiency factors they have been giving. It isn't particularly obvious how this generation cards live up to those numbers.
I'm guessing if they clocked these much higher, they lost some of that efficiency. I think they usually say stuff like "2.0x efficiency at same clock"
- Feb 22, 2015
- 1,052
- 1,990
Where did you find these die sizes? Can you link to the source?
Had AMD kept the clocks the same, then that level of efficiency gain would have been realized. But if they're clocking chips very aggressively, then they're throwing it all away and boosting TDP in order to increase performance, at low/no cost to them (at least from a fabrication perspective). As Puma says:
AMD is basically betting that people won't give a darn about efficiency if they can deliver a card with ~70% of a GTX 1070's performance at ~50% the cost. I think that's a pretty good bet.
If they keep the promises they made with that card, it could be my next one. Having a short PCB there makes sense in my book, they maybe don't want to fall into the same trap they did with the Nano where they already made the shortest version of that card themselves.
With the short PCB, they can offer a good reference design and manufacturers only have to put their already existing short thermal solutions on top of it to make a whole bunch of ITX-sized cards which don't compete with AMDs "perfect" own solution. It seems like in the renderings, the PCB is a bit longer than ITX, though, not sure what's up with that.
With the short PCB, they can offer a good reference design and manufacturers only have to put their already existing short thermal solutions on top of it to make a whole bunch of ITX-sized cards which don't compete with AMDs "perfect" own solution. It seems like in the renderings, the PCB is a bit longer than ITX, though, not sure what's up with that.
Well in that case the TDP would be even lower, something in the realms of the GTX 960? Which would be even better.
I just googled them. The die size for RX 480 is rumoured to be leaked by some Chinese AMD former employee who included the die size on his LinkedIn (probably no one wants to hire such a "leaker"). The other ones can be found on, for example, TechPowerUp.
http://videocardz.com/58237/amds-project-f-is-232mm2-discrete-gpu-made-in-14lpp-process
https://www.techpowerup.com/gpudb/2735/radeon-r9-nano
You are right, it's referred to as "power", which leaves the possibilities open. For some reason they want to keep certain numbers secret for a while more, maybe until that NDA lifts on June 29th. Hopefully for good surprises. Personally, I feel quite attracted to these Polaris cards, since they seem to offer great value for the money. If AMD would disappear, it would be a disaster for the consumers. So I am likely to buy a Polaris card. Provided the benchmarks aren't too bad, or something else is left shady. Worried about coil whine and "leaf blower noise" though. Releasing a $200 card probably means they aren't using premium components...
There is some speculation that there'll be a higher tier GPU around the 299$ price that will be competing against the 379$ GTX 1070. Basically because the RX 480 seems to be a 90W TDP card and it would be silly of AMD to cater to people that use 250W or less PSUs. So it's probably going to be releasing an 8-pin or dual 6-pin card with the "full Polaris 10" with a higher TDP and higher performance.
Yeah that seems to make sense. Though one has to wonder why AMD didn't mention or show this more powerful $299 Polaris card at Computex then, to possibly prevent people from buying the GTX 1070. I haven't fully figured out how AMD is thinking, but I get the feeling they haven't showed everything yet. Hopefully.
GDDR5X + higher clock rates might help close the gap on the 1070. I thought the RX 480 was already using full Polaris 10 though?
Similar threads
- Article
- Article