AMD Polaris Thread: Radeon RX 480, RX 470 & RX 460 launching June 29th

[oussama-tn]

http://wccftech.com/amd-radeon-rx-480-8-gb-229-usd/

According to wccftech the rx 480 has a compute performance of 5.5 TFLPOs and there is another rumor confirming the 1266 clock speed.

 

[Creig]

So far, easily AMD's best reference blower by far. Quiet & cool. Very NV "Premium"-like cooler design.

Gibbo agrees.

Gibbo can you let slip if AIB cards with aftermarket coolers will be available as well on the 29th?

No custom stuff, ref. only which in my view is fine as AMD nailed the cooler this time, it looks great, is quiet and performs.

https://forums.overclockers.co.uk/showpost.php?p=29653170&postcount=4711

 

[misuspita]

I do wish the NDA lifted BEFORE it went on sale. So we could at least get a review the day before.

Yes, but that way a few good men would argue it is not a truly hard launch, therefore it is a paperlaunch, therefore AMD sux and nVidia doesn't, because they did a real launch, with real live, not woodscrews boards. So, no more cookie for you before lauch!

 

[Yakk]

Yes, but that way a few good men would argue it is not a truly hard launch, therefore it is a paperlaunch, therefore AMD sux and nVidia doesn't, because they did a real launch, with real live, not woodscrews boards. So, no more cookie for you before lauch!

From many sources AMD could indeed do a hard launch right now :

"We have the stock already, all booked in and ready to go on one brand."

https://forums.overclockers.co.uk/sh...postcount=4777

If a seller like Gibbo has enough stock on just brand to ship 500+ cards in just a day, looks like AMD's AIBs all got plenty!

 

[torlen11cc]

So how powerful will the RX 480 be?
faster than GTX 980?
Will two 480's be faster than GTX 1080?

 

[RussianSensation]

Would probably be 250-260mm² in TSCM's process, that's already ~80% of GP104's die size.

And since the chip will never be made at TSMC, this comparison is worthless. I could just as easily pull numbers of thin air and state that at TSMC the chip would be 250-260mm2 but have 1.6-1.7Ghz Boost since TSMC's 16nm is superior to GloFo's 14nm. See how speculation works?

Let's keep going. You also inherently tried to insinuate that AMD's entire GCN4.0 will have similar perf/mm2 scaling (Vega 10/11). Is that why 438mm2 R9 390X is only 24% faster than HD7970? Notice how Hawaii had exponential perf/mm2 scaling? Don't forget that Vega will have HBM2, with a huge boost in memory bandwith.

It's interesting how the perf/mm2 metric is being used with total ignorance as to the real world power usage. After-market 1080 cards are using 190-210W of power, not 120W. Did it occur to you that AMD may have sacrificed perf/mm2 to improve perf/watt?

If it only matches Radeon R9 390X, it's nowhere near 80% GP104's performance.

Ya? The cheapest after-market 1070 is $419.99 USD, nowhere close to RX 480 $229-249 rumoured market price. Here is something for you:

RX 480 = $229-249 -> successor to Pitcairn HD7850 $249
Notice how AMD kept the price similar?

1070 = $419-449 -> successor to a $329 970. Notice how NV raised prices $90-120?

AMD is selling a mainstream chip at mainstream prices.
NV is selling a mid-range chip at high-end prices.

Who is ripping consumers off more?

Nice attempt to use the worthless perf/mm2 metric to hide how overpriced the 1070 is for 1080p 60Hz gaming or otherwise.

The October rumour comes from two forum posts from 3DCenter. AMD itself says Vega is 2017.

That's good. Gives enough time for miners to buy 2x $500 RX 480s, pay them off entirely, sell them in 2017, take the resale value + ether earnings and get a free Big Vega/Pascal. Sounds like an amazing stop-gap solution.

 

[96Firebird]

Did it occur to you that AMD may have sacrificed perf/mm2 to improve perf/mm2?

Wat.

 

[biffosaur]

Wat.

Yeah i got lost here too. Probably because I don't speak russian :sneaky:

Insulting other members will not earn you points here. It is against the rules.

-Rvenger

edit* Can't reply to admins because i don't have enough posts; this was an innocent joke/ play on RussianSensation's name...not meant to be an insult of any kind!

 

[Sweepr]

And since the chip will never be made at TSMC, this comparison is worthless.

It's not because we're probably not far from Polaris 10's real competitor, GP106, which could be smaller and almost as fast or even match it. Doesn't really matter for consumers if they're similarly priced, but still valid from a technical perspective.

Let's keep going. You also used inherently tried to insinuate that AMD's entire GCN4.0 will have similar perf/mm2 scaling (Vega 10/11).

Making up stuff #1

Is that why 438mm2 R9 390X is only 24% faster than HD7970?

Hawaii used newer GCN 1.1 and launched almost 2 years later, so it's more than expected. Quoting you:

''This comparison is worthless.''

It's interesting how the perf/mm2 metric is being used with total ignorance as to the real world power usage. After-market 1080 cards are using 190-210W of power, not 120W. Did it occur to you that AMD may have sacrificed perf/mm2 to improve perf/mm2?

Sacrificed something to improve on the same thing? Doesn't make much sense. If you mean perf/watt, no, I don't think they are limiting Radeon RX 480 performance to improve perf/watt on their fastest desktop part. I think that's the best they got out of their mainstream chip, and they priced it accordingly.

1070 = $419-449 -> successor to a $329 970. Notice how NV raised prices $90-120?

They did it because they can. You can bet it would be priced more agressively if Vega was launching instead of Polaris.

Nice attempt to use the worthless perf/mm2 metric to hide how overpriced the 1070 is for 1080p 60Hz gaming or otherwise.

Making up stuff #2

 

[Glo.]

It's not because we're probably not far from Polaris 10's real competitor, GP106, which could be smaller and almost as fast or even match it using a smaller die.

With 2 GPC's(1280 CUDA Cores)?

Nope. It will not match even base clocked P10 XT. GP106 can be a competitor to RX 470. But not for full P10 die.

 

[Sweepr]

With 2 GPC's(1280 CUDA Cores)?

Nope. It will not match even base clocked P10 XT. GP106 can be a competitor to RX 470. But not for full P10 die.

That would be a huge gap between GP106 and Geforce GTX 1070, and we already know the second cut down GP104 part (GP104-150) is a notebook part, not a 'Geforce GTX 1060 Ti' (according to Chiphell). There have been rumours about a 192-bit 6GB GDDR5 part that delivers similar performance to Geforce GTX 970/980. Anyway this is way OT.

 

[Mopetar]

So how powerful will the RX 480 be?
faster than GTX 980?
Will two 480's be faster than GTX 1080?

The 480 should have about 5.8 TFLOPs of performance, which lines up with the 390X. The 480 does have a smaller memory bus though so it might bottleneck in certain workloads, but the general performance should be similar.

The 480 is also going to have a newer GCN core which AMD has said has some improvements such as the primitive discard accelerator which improves performance in certain situations. A good example is when AMD showed off Polaris at their March event running Hitman at 60 fps in 1440 in a part of the game where previous cards would get bogged down trying to draw NPCs that were hidden behind a wall.

They probably have some other special sauce that improves performance in similar situations, but 390X levels of performance seems the most likely for the average case.

Don't forget that Vega will have HBM2, with a huge boost in memory bandwith.

HBM also uses less power which means more juice to devote towards the GPU as well, so even if the bandwidth worked out to be the same, you'd still expect a chip with HBM to have better performance.

 

[el etro]

Yep, it looks unlikely that AMD will compete with Nvidia on the perf/mm2 game. Who will suffer with this is the big Vega, other cards will be fine.

AMD may have some advantages with the WSA contract with GF(and very likely they will fill it because Next Xbox and PS4K can be fabbed at GF), so they may not have a silicon cost disvantage versus Nvidia this time. They just need to match Pascal on the efficiency metric and all will be fine.

 

[AtenRa]

Wat.

Probably he wanted to say sacrifice perf/mm2 to gain perf/watt or perf/$

 

[Glo.]

That would be a huge gap between GP106 and Geforce GTX 1070, and we already know the second cut down GP104 part (GP104-150) is a notebook part (according to Chiphell). There have been rumours about a 192-bit GDDR5 part that delivers similar performance to Geforce GTX 970/980. Anyway this is way OT.

GP104-150 is cut down part from GP104, and it was rumored to be... GTX 1060 Ti.

GTX 1060 would be made from GP106 - 2 GPC setup.

 

[el etro]

AMD needs to match GTX 980 or at least 390x performance, to don't have problems competing with GP106. In case of tie, Nvidia will have the brand advantage to compete with the RX.

 

[Stuka87]

Yep, it looks unlikely that AMD will compete with Nvidia on the perf/mm2 game. Who will suffer with this is the big Vega, other cards will be fine.

The "Perf/mm2 game" has no meaning. It is not relevant to anything. Perf/Watt is something that affects customers. Perf/mm2 has zero impact on anything.

 

[RussianSensation]

Wat.

It's a typo; corrected. In any event, I am going to straight up assume anyone who owned GTX 260/260 216/275/280/285/470/480/570/580/780/780Ti and didn't not complain about perf/mm2 and buy a competing AMD card is straight up trolling.

As I said already, we cannot assume that Vega will have similar perf/mm2 to Polaris 10 because R9 290X/390X didn't either.

Also, it matters a lot more that NV raised prices from $329 to $419-449 with a small die size, but AMD has kept prices at the same level as HD7850, the spiritual predecessor to the RX 480. It's pretty clear which company is padding its gross margins.

Why don't you comment on $229 reference blower RX 480 vs. $449 reference blower 1070. Let me know when a reference 1070 is 96% faster than a reference RX 480

 

[sirmo]

Yep, it looks unlikely that AMD will compete with Nvidia on the perf/mm2 game.

What does that even mean? These perf/mm2 callbacks are embarrassing frankly.

Like seriously I would love to know if anyone on this forum knows the surface area AMD's CU clusters use on the die compared to Nvidia's SM ones. Until you know that there is no point in comparing the perf/area. And even then, different design approaches have different tradeoffs, my understanding is that for higher clocks you usually use more die space. I don't think anyone short of a process implementation engineer who works for some of these companies could even begin to explain it.

Also AMD and Nvidia are on two different fabs. Different features entirely and different economies of scale. It's the most meaningless topic I've ever seen discussed on any technical forum ever.

 

[IllogicalGlory]

I'm expecting closer to R9 390X performance at stock, though admittedly that's a fairly small difference. At the default 1266 MHz boost clock, it should have almost as much raw power (TFlops) as a 390X, so even discounting any major architectural improvements since GCN 1.1, it should be as good as 390X at a minimum.

Overclocking headroom should be better, since the 390X is basically a factory OC'd card to begin with. Hawaii usually couldn't overclock higher than 1100-1150 MHz (less than 10% over 390X), while we've already seen leaks indicating that P10 should be able to hit 1500 MHz without much trouble (~18% over stock RX 480). In practice, I expect that to translate to about 10%-12% additional performance.

Also, a lot of people are placing too much stock in 3DMark scores. While it's not insanely vendor-biased (nothing like, say, Project CARS in that regard), it does tend to lean a few percentage points towards Nvidia. R9 390X beats GTX 980 at 1080p in TPU's combined real-world gaming benchmarks, while it falls behind by a couple percent in 3DMark. Perhaps more importantly, this is just a single number, while a major real-world consideration is minimum FPS. Based on the Hitman demo and AMD's mention of a "primitive discard accelerator", there's some reason to hope that minimum FPS might see improvements on Polaris. That means a better gaming experience at given settings even if average FPS is similar to older products. I'm interested to see what TechReport has to say about frame latencies on this card (once they work through their review backlog).

As for perf/watt, I think it will come reasonably close to GP104 (GTX 1070/1080) and beat GP106 when that comes out. GTX 1080 peaks at roughly 185W during gaming. R9 390X is about 60% of 1080's performance, so if we assume RX 480 also lands there, then it would need peak gaming power of 111W or lower to match 1080's perf/watt. Reference R9 270X only peaked at 122W in gaming, and with this chip being similarly sized and having reasonably conservative stock clocks compared to what the silicon is capable of, I would not be surprised to see similar results. This would mean probably about 90% of GTX 1080's perf/watt.



That's an oddly specific prediction about the TPU score. Just to be sure I double-checked their site and found nothing. Even tried manually formulating a URL to see if they maybe had it posted but just not linked, but got a 404.

I don't see how P10 is an engineering failure. Similar die size and power consumption to Pitcairn, with roughly 2x the performance in both compute and gaming, plus a full complement of modern video processing features. What's the problem? If GP106 is half a GP104 (the most likely scenario, IMO) then P10 should beat GP106 in both performance and perf/watt.



Expecting that P10 would match 1070 was always a stretch. At this point it's fairly clear that isn't happening, at least not at this time. AMD never promised any such thing.

To be clear, I'm completely making this up based on the leaks. I have no insider info. I'm just posting it so I can be embarrassed when I'm wrong.

What I mean by engineering loss is that GP106 will be more efficient and faster at the same die size. Pascal > GCN4.

Everything you're saying might be true, but I just have a bad feeling about this card.

 

[sirmo]

To be clear, I'm completely making this up based on the leaks. I have no insider info. I'm just posting it so I can be embarrassed when I'm wrong.

What I mean by engineering loss is that GP106 will be more efficient and faster at the same die size. Pascal > GCN4.

Everything you're saying might be true, but I just have a bad feeling about this card.

Pure conjecture.

 

[krumme]

To be clear, I'm completely making this up based on the leaks. I have no insider info. I'm just posting it so I can be embarrassed when I'm wrong.

What I mean by engineering loss is that GP106 will be more efficient and faster at the same die size. Pascal > GCN4.

Everything you're saying might be true, but I just have a bad feeling about this card.

Well if gibbo is to be trusted and they have 500 cards for day one. Half the amount of 1080 for all days. Its the engineering miracle of this decade imo. Gf turning out finfet in spades. A company that was 3 years behind tsmc.
I dont really care if perf or eff is plus minus 10%. If this is a sign what is in front of us it will give the biggest boost to compettition and price perf for consumers in a decade. The importance of gf beeing able to use the samsung process can not be understated.

 

[IllogicalGlory]

Pure conjecture.

Absolutely it is. Not claiming anything else.

 

[Stuka87]

To be clear, I'm completely making this up based on the leaks. I have no insider info. I'm just posting it so I can be embarrassed when I'm wrong.

What I mean by engineering loss is that GP106 will be more efficient and faster at the same die size. Pascal > GCN4.

Everything you're saying might be true, but I just have a bad feeling about this card.

Based on what information? Or is this just a guess?

 

[AtenRa]

Just a consideration,

There are only two things unchanged from Hawaii to Polaris as seen in the slide bellow and those are the Rasterizers and ROPs. So with the same amount of Rasterizers (4x) and ROPs (64x) count on both Hawaii (R9 390X) and Polaris RX 480, the performance in games/synthetics that are Raster/ROP limited will be the same.

I believe this is what we see in 3D Mark FireStrike.