Official AMD Polaris Review Thread: Radeon RX 480, RX 470, and RX 460 *RX 470 LIVE*

http://www.purepc.pl/karty_graficzn...strix_gaming_chlodny_i_cichy_polaris?page=0,4


Interestingly here we have the very same charts of the RX460 "leak" without the RX460 numbers, down to the highlight on the RX470 since it's a RX470 review. Videocardz most probably stumbled upon some fakes.

 

Could be they got both the RX 470 and RX 460 cards at the same time, were tested at the same time, but the RX 460 numbers were "accidentally" included in the RX 470 results. Good way to get some extra traffic on their website.

If they are fake it looks like someone did a reasonable job of guesstimating factory overclocked RX 460 performance. The numbers seem believable for a card with only 44% of the shaders of the RX 470 and 2GB of memory.

 

Slower than a 950 is to be expected, the 460 is only 39% of a 480 spec wise after all, and a 950 is 48% the performance of a 480.

The really interesting thing here will be the power usage (to me at least). Unfortunately we are looking at aftermarket cards and full system power, which makes it a bit harder to compare. AMD originally showed off a PC system with a 460 using 54W less than a similar system using a 950. PurePC only gets a 34W difference, but the aforementioned issues makes it a bit fuzzy to compare these things. It will be nice to get some numbers where the GPU power usage is isolated.

 

The RX 460 efficiency isn't impressive in this test. Crysis 3 was used for power testing. The GTX 950 was 34% faster in Crysis 3 than the RX 460, so the performance/watt of the RX 460 probably wasn't much better than the GTX 950. The higher GTX 950 performance probably increased the power consumption a bit on the CPU and other system components. That would make the system power difference somewhat larger than the actual video card power difference as well.

 

It's worth noting that the 950 in question is unusually fast. Comparing it to the 480 nitro and extrapolating to the numbers TPU have for reference models, that 950 is a good 50% faster than a reference 950.

 

Basically matches a 750ti....

 

According to an actual TPU review of a Gigabyte GTX 950 WindForce OC we should expect that the PurePC Gigabyte GTX 950 isn't all that much faster than a reference GTX 950. No factory overclocked air-cooled Maxwell card is going to be anywhere near 50% faster than its reference counterpart.

 

Take in care something:
- The card shown was the full version. The one with the leaked performance has less SP, so less performance
- Seems that the PCI only power way is not the way AMD meant to go with the card.
- It won't OC as many can expect.

However seems that is not targeting the 950 after all, but only the GT tier cards like 940 or the older GTX tier cards like 750Ti.

The tier that nVIDIA didn't care to cover and Intel can't still catch up.

 

it's behind the 750 ti and even the 260x in some tests.. ouch.
overall it's a little faster than both, but not that much...

 

The RX 460 is underwhelming. Expected better from all of the initial demos of P11 claiming better perf/watt than GTX 950

 

How could it possibly be slower than the 260X? Both the 260X and 460 have 896 shaders, 16 ROPs, and a 128-bit memory bus. The Polaris 11 card has faster core and memory clocks than its Bonaire counterpart. Any architectural changes should be to Polaris's advantage. Are we still looking at immature drivers at this point?

The fact that AMD made the RX 460 a cut SKU is also puzzling. Was there a huge amount of demand for uncut P11 for mobile/AIO, or are GloFo yields still terrible?

 

Numbers say it's 18% faster than 750ti
950 is 17% faster than it

If the Vulkan and DX12 numbers wern't (seemingly) borked, it would look a tad better. (numbers show it slower than a 260x, which is 'impossible')

 

The ref model could be power limited. I'll have to guess that 950 model has external power for instance.

 

That's my whole point. The 950 Windforce that PurePC has is performing much faster than it realistically ought to.

In PurePC's tests the 950 Windforce is 67% the performance of the 480 Nitro. A 480 Nitro is roughly 8% faster than a reference 480. As such if the 950 Windforce is 67% of the Nitro card it would be 72% of a reference RX 480 (67% * 1.08 = 72%). TPU shows a reference 950 as being 48% of a reference 480, thus PurePC's 950 Windforce would be 50% faster than a reference 950 (72% / 48% = 1.5).

So either PurePC's 480 Nitro numbers are messed up, their 950 Windforce numbers are messed up, or they have the most incredible golden sample 950 GPU the world has ever seen.

Edit: just noticed that PurePC is actually running their 480 Nitro in quiet mode, so their 950 Windforce would only be 32% faster than a reference 950, not 50%. Still way above normal.

Edit 2: I just tried looking at their 950 Windforce numbers vs. their 370 Gaming numbers. The 370 Gaming is running in silent mode, which means it is basically running stock clocks. The 950 Windforce is about 10% faster, which is roughly what you would expect vs. a reference 370, so it would appear that the above issue is due to PurePC's 480 Nitro numbers being way too low, not the 950 numbers being too high. Of course that arguably puts all of their Polaris numbers into question.

 

Lol where are all those people who said it would be as fast as 950/960? AM sure some people replied to me saying that when i said it will be 20-30% slower than 950.Even read a lot of folks saying as fast as R9 270X ! Even i was wrong tbh,it is even more slower than 30%.Slower than even R7 370.Worse perf/watt than Maxwell 950.I don't see the 2.8X more perf/watt anywhere in those charts.Yet Another lie from AMD?
Its not completely their fault though.Polaris just isn't as efficient as we all thought it would be,even less efficient than Maxwell.It will take them another generation or two to even match Maxwell in efficiency.
As fast as 750Ti when 750ti costs less than $100,i don't see what the fuss is all about.Most people would be better off with $110 950 because $110-130 for 2gb RX460 and $130-150 for 4gb makes no sense.GTX 950 crushes it while costing less.

 

Either those leaks are false, or amd completely lost their mind releasing a cut down useless card.

 

I wonder which one of these cards showcased against the GTX 950 in the Battlefront was it. (142w vs 85w) Anyone takes a guess?

 

I can't believe AMD managed to deceive even the most experienced members of Anandtech. And we say AMD cannot do marketing(lying)

 

That's correct, overclocked cards top out at being over 20% faster then reference just like in the case of Asus Gold Platinum which is basically a special edition of ROG Matrix or zotac gtx 980 Ti AMP! EE
which is a lot and basically a whole new tier of performance. That's actually more then what separates the top 80 series model from 70 series model.
970>>>980 = about 15%
1070 >>> 1080 = about 20%

 

Wait for the reviews tommorrow?? No point going off one leak which might or might not be true.

But,I would at least expect an third party RX460 to match or beat an HD7850?? The HD7850 has 14% more shaders but the RX460 should have 20% more clockspeed. Memory bandwidth would be less,but the RX460 has higher speed GDDR5 and the memory compression thingy too,and the R7 260X does not have that.

If it can't even get to HD7850 level with a third party card,something has definitely gone wrong.

 

Different gpus and apis put different loads on cpu. So I think best result is total system power than just gpu power. Is there any site that tests total system power? Sorry for being off topic.

 

So performs worse than the no connector 75W GTX 950s and draws more power?

 

GloFo gonna GloFo.

 

GloFo isn't to blame for it all.

 

Why is it whenever an AMD product underperforms it's GloFo's fault? 14LPP isn't as good as 16FF+, but the performance delta at the process level is like 10-15% tops.

 

10%-15% is quite a lot, it would put the reference 480 at 980/Fury level.

Indeed, Polaris under-performing could be placed on Glofo, and I'm sure they have some part in it. However I think that people are putting too much blame on GloFo.

TSMC made both Hawaii and Maxwell, and we saw what happened there. Things look quite similar to how they looked in the previous generation.

 

TPU only measures the GPU, Chiphell graph is for the entire system (109W in not the RX 460 alone ).

Also RX 460 compete against the GTX 750Ti not the GTX 950.

Custom RX 470 4GB from Gigabyte and MSI currently available in newegg at $199,99, at that price they are way better products than current GTX 960 4GB in that price range.

http://www.newegg.com/Product/Produ...85&IsNodeId=1&bop=And&Order=PRICE&PageSize=30

 

Pricing will determine this, GTX 950 starts at $134 right now. From the leaked benches Polaris 11 barely outperforms a 2.5 years old 28nm planar 750 Ti and uses more power.

 

I don't think AMD agrees with you on that one, given that they themselves have compared it to the 950.

I don't think we can conclude yet whether or not the 460 uses more power than a 750 Ti. As AtenRa mentioned the ChipHell numbers are system power, and since we don't have comparable numbers for the 750 Ti, we can't really compare yet.

 

How are you so quick to dismiss that Glofo is not to blame. AMD is basically pumping a lot more voltage than necessary as to qualify a lot more dies as working Rx 480. The fact that you can undervolt and get better performance due to reduced clock throttling shows that the process has significant variance which is forcing AMD to use more voltage to qualify the maximum dies possible.

https://translate.google.com/transl....de/2016-06/radeon-rx-480-test/12/&edit-text=

dude 10-15% higher transistor performance is quite close to a half node improvement in terms of transistor performance . Full node transitions bring 30-35% higher transistor performance at same power or 50% lower power at same transistor performance. The more damning aspect is GF 14 LPP is nowhere close to a robust implementation of Samsung 14LPP with significant process variance. In fact AMD has mentioned they are running actual product through Samsung and I would guess if GF does not resolve the process issues they will force GF to reduce their WSA commitments and have Samung fab chips for them. The Q2 2016 GF wafer purchases were very low and the 2016 WSA has still not been revealed.

http://seekingalpha.com/article/399...-results-earnings-call-transcript?part=single

Total wafer purchases from GLOBALFOUNDRIES in the second quarter was $75 million, and year-to-date, we have purchased $259 million.

Stacy Aaron Rasgon - Sanford C. Bernstein & Co. LLC

Got it. If I could ask one more really quick. I was just a little surprised at how much your wafer purchases at GLOBALFOUNDRIES came down quarter-over-quarter, given the increase in notebooks as well the timing of the Polaris launch. Should we read anything into that?

Devinder Kumar - Chief Financial Officer, Treasurer & Senior VP

I think basically the purchases of the wafers are in line with product demand and mix of business and as I said in the prepared remarks, year-to-date we've purchased about $260 million of wafers and we are getting into the back half of the year, which as you can see, with our revenue guidance in Q3, is pretty strong.


Joseph L. Moore - Morgan Stanley & Co. LLC

Great. Thank you. I was also curious about the GLOBALFOUNDRIES being so low in Q2. And I guess, have you guys finalized the 2016 wafer supply agreement, and can you talk more generally about how you're deciding which products are allocated to GLOBAL and TSMC and is there anything that's exclusive to one or the other?

Devinder Kumar - Chief Financial Officer, Treasurer & Senior VP

Yeah, several parts to it. The working relationship, as Lisa said, with GLOBALFOUNDRIES is very good. We continue to work through the 2016 WSA and that part's now finalized. We are in the process of negotiating that. In the meantime, we continue to get delivery of wafers for the products that we need in line with the product demand and mix of business. And relative to your – which products on which foundry, we typically do not share the source of foundry for any of our wafers or products.

http://www.forbes.com/sites/patrick...14nm-manufacturing-with-samsung/#3e0024a750fb

 

That comparison was for perf/watt at 60fps cap, also that RX 460 they used back then could be the full chip and not the cut down version they are using for the RX 460 today. The product itself today (RX 460) compete against the GTX 750Ti.

 

We have the leaked PurePC results where RX 460 is hotter and uses a bit more power:

http://videocardz.com/63033/gigabyte-radeon-rx-460-windforce-2x-performance-leaked

Tomorrow we find out if the much better perf/watt than Maxwell they promised earlier this year turns out to be true or not. And no, the full chip at lower clocks wouldn't do miracles so that excuse is not valid.

 

R7 260X
14 Compute Units and 896 shaders
Core Clocks = 1100MHz
56 TMUs 16 ROPs
128bit memory and 104GB/s bandwidth

RX 460
14 Compute Units and 896 shaders
Core Clocks = 1090MHz / Boost 1200MHz
56 TMUs 16 ROPs
128bit memory and 112GB/s bandwidth

Any benchmarks that has the RX 460 bellow R7 260X are subject to question.

 

That's an aftermarket 460 and 750 Ti though, so the reference numbers are still up in the air. With that being said though, it certainly looks like the efficiency claims AMD have been making are optimistic to put it mildly.

 

It's the same custom model from Gigabyte (Windforce 2X), it is comparable. And I'm not concluding anything, I clearly said 'from the leaked benches'.

 

That's why I am saying it is power limited, and the clocks are dropping below base due to hitting the 75W limit. When this was originally rumored to be a 50 W part.

 

The RX460 only has 896 shaders - the full Polaris 11 GPU has 1024 shaders,which the WX4100 has. Even some of the early AMD marketing material said it was a 1024 shader card.

So it makes me wonder why it has 896 shaders?? 1024 shaders at a lower clockspeed is probably more efficient.

 

Dr

How convenient, comparing a new 14 nm card with the previous generation 28 nm card from your competitor. It had better win, and should win by a lot more than it does.

Actually, the 460 will "compete" with whatever the market determines, not some artificially constrained scenario posited by AMD to make the chip look more competitive than it really is. Basically, it "competes" against any card of similar price that does not require a six pin connector. The 460 seems to me by far the worst of the Polaris line. Its only saving grace is that it offers semi-competent 1080p performance without a six pin connector. Yes, it beats nVidia's previous generation 750 Ti (but not by as much as it should) and probably would lose to a 75 watt 950, although I consider that to be a sort of fringe card, much like the "green" 9800 GTs that came out a few years ago. And we haven't seen the true competitor yet, which will be the nVidia 10xx equivalent. Given the significant performance/watt advantage nVidia still owns, I would expect that they could squeeze a lot more performance into a sub 75 watt card, should they care to. I am also hoping nVidia (or later AMD) chooses to put out a card at around a hundred watts with performance intermediate between the 460 and 470.

 

Its more worrying AMD can't sell the RX460 as a 1024 shader card. 1024 shaders at 1.1GHZ+ would get closer to a GTX950. Polaris 11 does not look a very big GPU either. The R7 370 has nearly 15% more shaders.

 

Why are you so quick to dismiss uarch and IC design? Why do GloFo always have to get the blame?

GP104 would only run around 1375Mhz if it wasn't for a lot of extra work.



There is a reason why Pascal does even better in mobile than it does on desktop, unlike Polaris. R&D matters.

 

Agreed. Even more, I feel that the 460 was in fact the rumored 450 that has that ammount of shadders for the lowest tier market and OEMs.
At least is something new... that NVIDIA can't deliver since tons of years ago a new brand low tier card since Kepler and the only card in that range that has Maxwell was the GT 930 and that is because that card have 3 models: a good model (Maxwell at 15W), a meh! model (Kepler at 25 W) and a abomination (Fermi at 50W!)
http://wccftech.com/nvidia-geforce-gt-930-launch-q1-2016-maxwell-kepler-fermi/
https://forums.anandtech.com/threads/nvidia’s-low-end-geforce-gt-930-rumored-to-launch-in-q1-2016.2458367/
And the rest of the low tier were just mere rebrands. If nVIDIA doesn't deliver something new with Pascal, seems that AMD would win big time even with that mediocre card.

Also. Seems that AMD is reserving the 450 and the 465 models for something. The full model maybe will come later and then a even more cut down version with passive cooling.

 

Just because they are both Windforce cards, doesn't necessarily mean that they are tuned/overclocked in a comparable fashion. Either way though, whether or not the 460 ends up above or below the 750 Ti, the efficiency claims AMD were making looks a long way off.

 

I'm not so sure there is anything to worry about. AMD might send full Polaris 11 to laptops (as the 480m) and workstations where the margins are much higher.

 

Well the competitor doesnt have a new 16nm GTX 750Ti, same thing as GTX 1080 vs Fury X

Lets wait and see what MSRP RX 460 will debut and its performance and then we can talk about it, but it is obvious its not a GTX 950 competitor. A full Polaris 11 chip could be a GTX 950 competitor, im guessing AMD waits for NV to release its new 16nm GTX 750Ti and then they will release a full Polaris 11 chip as the RX 465 ??

 

There's a review of RX 470 on Serbian Benchmark.rs web site but I dunno if accidentally or on purpose, they've put results of Saphire Nitro rx 460 4gb.

http://www.benchmark.rs/artikal/test_sapphire_radeon_rx_470_nitro_8_gb-4169/6 .

Seems that it does around Gtx 950 performance, but in DX12 same as or higher then Gtx 960.

Can't believe I did Sign up on Anandtech forum just so everyone can check this out.

 

From this review the cut down Polaris 11 is as fast as the GTX 960 in DX-12. Lets wait and see more reviews and watch if this trend will continue.

Thanks for the link.

 

Interestingly the GTX960 used is a Gainward Phantom which has a massive cooler too. Seems to have very variable performance - matches an overclocked GTX960 in Hitman:Absolution,Ashes of the Singularity and GTA V but then gets thrasjed in other titles.