The technical merits: Polaris vs. Pascal

[tviceman]

RX 480 (full Polaris 10)
232 mm2
5.7b transistors
Density: 24.57 million transistors /mm2

GTX 1080 (full GP104)
315mm2
7.2b transistors
Density: 22.86 million transistors /mm2


The Technical and Performance Metrics
GP104 is a 36% larger die.
GP104 has 21% more transistors.
Polaris 10 is 7.5% more dense than GP104.
GTX 1070 and 1080 are 70-80% more efficient depending on 1080p or 1440p.
GTX 1080 is 75-85% faster depending on 1080p or 1440p.
GTX 1070 is 50% faster at every resolution.

Perf/$
RX 480 8gb is 66% more cost effective (perf/$) than GTX 1080 (current prices).
Crossfire RX 480 8gb is 17% more cost effective than GTX 1080.
RX 480 8gb is 25% more cost effective than GTX 1070.
GTX 1070 is 9-14% more cost effective than crossfire RX 480 8gb.
RX 480 4gb is 50% more cost effective than GTX 1070 and twice as cost effective as a GTX 1080.



I based these results on techpowerup's RX 480 review graphs between the 1440p and 1080p resolutions, and also Ryan Smith's summary on the last page of the RX 480 preview. It will be interesting to make these comparisons again when GTX 1060 comes out. I will update this thread when it happens if people are interested in continuing a discussion.

 

[kagui]

but what if you throw in, vulkan, dota, linux

http://www.phoronix.com/scan.php?page=article&item=amdgpu-rx480-linux&num=11

do these change your efficiency numbers?

 

[sm625]

According to the techspot review a 1070 only costs less than 10% more in terms of $ per frame.

 

[Silverforce11]

If you base it on DX12 and Vulkan games, then for sure RX 480 looks a lot better.

But so does the 390/X. 390X still leads the RX 480 in DX12/Vulkan.

Sure, a lot less power, but damn.. it would have been nice if RX 480 was a 110W gaming load. Change the entire perception and give it a fighting chance vs the upcoming 1060.

 

[Silverforce11]

According to the techspot review a 1070 only costs less than 10% more in terms of $ per frame.

That's because they are using the WRONG $.

http://www.newegg.com/Product/Produ...N=100007709 601204369&Tpk=GTX 1070&ignorear=1

Look at the 1070s on Newegg. There's ONE that's $399. One horrible one that's $409. The rest are more expensive.

Until the average price of the card drop down to MSRP, they are doing it wrong on their charts. It's actually a damn lie and straight out of NV marketing guide.

"Here guys, these 1070s are only $379... except you can't buy it for that price. It's out of stock everywhere and the average is closer to $439. But who cares, it's just $379!"

 

[digitaldurandal]

Hopefully the current power numbers for Polaris are fixable without castrating performance. Maybe the firmware? It doesn't give me a lot of hope for Vega. Even with ~4000CU and 300w with the same clock speed it seems like it would perform under a 1070.

Am I wrong in thinking that Vega will likely be very similar to Polaris w/ HBM2? Is the architecture going to be much different?

 

[Sweepr]

I based these results on techpowerup's RX 480 review graphs between the 1440p and 1080p resolutions, and also Ryan Smith's summary on the last page of the RX 480 preview. It will be interesting to make these comparisons again when GTX 1060 comes out. I will update this thread when it happens if people are interested in continuing a discussion.

Also based on TPU's review - GP104 has a ~36% lead on perf/mm² @ 1440p.

 

[bystander36]

That's because they are using the WRONG $.

http://www.newegg.com/Product/Produ...N=100007709 601204369&Tpk=GTX 1070&ignorear=1

Look at the 1070s on Newegg. There's ONE that's $399. One horrible one that's $409. The rest are more expensive.

Until the average price of the card drop down to MSRP, they are doing it wrong on their charts. It's actually a damn lie and straight out of NV marketing guide.

"Here guys, these 1070s are only $379... except you can't buy it for that price. It's out of stock everywhere and the average is closer to $439. But who cares, it's just $379!"

They compared it to $400, not the list price. Anyway, there are other places to buy them. Is this one of the horrible ones, because it is tempting for someone wanting to upgrade: http://www.google.com/shopping/prod...33&q=gtx+1070&hl=en&ei=bo90V72VFoemjwP23p6gCw

I'll probably wait until higher end AMD cards show up before I make a choice.

 

[Silverforce11]

For those interested: https://www.youtube.com/watch?v=24rFeHo1S-o

Interview with Raja from PCPER, 1hr. Start at ~8min onwards.

Ryan gets him drunk and Raja just talks and talks... it's a good watch!

 

[SPBHM]

Hopefully the current power numbers for Polaris are fixable without castrating performance. Maybe the firmware? It doesn't give me a lot of hope for Vega. Even with ~4000CU and 300w with the same clock speed it seems like it would perform under a 1070.

Am I wrong in thinking that Vega will likely be very similar to Polaris w/ HBM2? Is the architecture going to be much different?

firmware? I think AMD has done a lot of work with the power management hardware/software or whatever, to improve the current cards they can drop clocks/voltage more (and lose performance), I think it will take new board designs and probably the process maturing to get power down a bit...

but, the RX480 seems OK in terms of power draw, same level as a 970, not great, but I don't see how that's a huge problem, proportionally it will be far behind a 1060 in perf/W but, perf/$ is king for this price range, and it's not like it uses 250W like the 290s... as I said I think proportionally it can look very bad, but at the end of the day 150-160W is not anything new for this price range.

also, isn't Vega supposed to use HBM2? that could make a big difference competing with 8-10GHz GDDR5 cards (1070/1080), like the Fury line perf/W was a big improvement over Hawaii in great part because of HBM!?
but yes, Pascal is looking so good, it will be difficult

 

[trane]

GCN 4 is solid, forward thinking, future-proof. Pascal is just Maxwell on speed. It's not going to hold up for DX12/Vulkan workloads, which will be omnipresent by next year. VR is a bit of a toss up, not sure how that will pan out.

GP104 is much better manufactured. They have really honed in on TSMC 16nm, pushed clocks and power efficiency to uncharted heights. AMD have struggled with GloFo 14nm. Which is not surprising, considering this is a brand new architecture on a brand new process which has never seen as discrete GPU built at a fab which has never made a discrete GPU. As a result, it clocks very, very low. The optimum clocks for Polaris 10 is probably in the ~1 GHz range, while GP104 is in the ~1.5 GHz range. That's an unbelievable 50% advantage. Yes, at 28nm Nvidia had a clock speed advantage, but it was more like ~20%.

14nm never brought about the expected advantage in density either. In fact, on 28nm, AMD was more than 7.5% dense than Nvidia, so they have actually lost on 14nm.

On the other hand, by foregoing perf/W with 480, AMD have been able to push performance while bringing the cost way down and retaining availability. They have turned off power management features, pushed voltages way high to meet the lowest common denominator - push as many P10s capable of 480 clocks as possible. So, AMD can afford to sell P10 at $200 with a relatively wide availability. 470 will get perf/W back in line.

So, it's a bit of a mix and match. P10 would have been much better at TSMC 16, but the move to GloFo had to happen. It'd be reasonable to assume they'll work it out by Vega time. On a pricing level though, $200 is still great for 480, and I imagine 470 will be even better.

 

[Cookie Monster]

GCN 4 is solid, forward thinking, future-proof. Pascal is just Maxwell on speed. It's not going to hold up for DX12/Vulkan workloads, which will be omnipresent by next year. VR is a bit of a toss up, not sure how that will pan out.

GCN is future proof because this is the strategy AMD has taken for the last 4 years. Instead of relying on massive architectural changes with each generation like they use to, they've gone out and convinced the industry by having them adopt APIs that are beneficial to their architecture. This is much cheaper (especially when your R&D budget is no where near your competitions) and its paying off in some situations especially with the aging bit. I would imagine in return AMD would be supplying their GPUs for cheap to console manufacturers (Sony/MS) which also happen to have the biggest influence when it comes to APIs.. not to mention that most games if not all are developed soley with "console first PC later" or to work on all 3 platforms mantra.

GP104 is much better manufactured. They have really honed in on TSMC 16nm, pushed clocks and power efficiency to uncharted heights. AMD have struggled with GloFo 14nm. Which is not surprising, considering this is a brand new architecture on a brand new process which has never seen as discrete GPU built at a fab which has never made a discrete GPU. As a result, it clocks very, very low. The optimum clocks for Polaris 10 is probably in the ~1 GHz range, while GP104 is in the ~1.5 GHz range. That's an unbelievable 50% advantage. Yes, at 28nm Nvidia had a clock speed advantage, but it was more like ~20%.

Unless we know for certain that Glofo's 14nm is rubbish which I will give them the benefit of doubt because it ain't proven yet, architectures play a big role (if not the primary role) in determining the clock frequency and power consumption. It makes no sense to compare clock speeds of the GPUs because their architectures are vastly different.

Comments bolded above.

 

[LTC8K6]

Not sure it's wise to look at performance per dollar right now. Prices are still adjusting, and other cards are coming soon that will surely change things.

 

[Alaa]

How do you calculate perf/$? Does it even take playable framerates into consideration? Because for me, it seems 480 8gb at 240$ has 10% extra perf/$ at 1080p over 1070 at 440$.

 

[beginner99]

Looking at the specs performance/transistor also looks pretty bad for AMD. GP104 only has 26% more transistors yet offers 1.8x of the performance.

Really makes you wonder what AMD is doing? Either they are still hiding the fact 480 is a cut-down die (from 40 CUs?) and only full chip actual would use all 5.7b transistors or it just really sucks. In fact it still looks bad even if full chips had 40 CUs.

 

[khon]

So in short the NVIDIA cards are better, but AMD lowered their prices to remain competitive.

 

[know of fence]

Looking at the specs performance/transistor also looks pretty bad for AMD. GP104 only has 26% more transistors yet offers 1.8x of the performance.

To adjust for transistor delta, divide performance by transistor ratio:
1.8x / 1.26 % = 1.43x more performance

1120/1266 MHz (Radeon 480X)
1607/1733 MHz (GTX 1080)
Base/Boost
1.43x /1.37x times higher clocks

It's funny how the result is exactly 1.43x.

 

[Flapdrol1337]

Afaik glofo and amd had a deal where amd had to pay, production or not. At least they now have something glofo can produce that people might actually want to buy.

 

[TheRyuu]

Pascal is just Maxwell on speed.

I'm not sure it's fair to paint it in quite this broad of a stroke. Sure I can't really disagree with that statement since I suppose it sums up the situation pretty well but Pascal did have some (albeit small) improvements to the workloads which you're referring to. It should, at least theoretically, stop the negative scaling we saw with Maxwell.

 

[Erenhardt]

Lets look at the memory bandwidth.

We have Polaris 10 and GP104 both using 256bit bus memory.

I know, 1080 uses gddrx memory, but even if we scale it back to rx 480 clocks, their compression gives 133% increased bandwidth, compared to only 36% on polaris 10.
That means with the base at 194 GB/s uncompressed texture, polaris can peak at 264GB/s, and pascal 453GB/s. This is a HUGE deal.

Crypto mining shows how badly rx 480 is held back by memory bandwidth. You can downclock the core from 1266 down to 950mhz without any impact on performance. Overclocking memory yields huge boost in hashrates. Gaming scenarios are a bit different workload, but the bottleneck still remains.

What is very worrying is AMDs first shot at memory compression with 285/380 gives 27% increase in bandwidth. With new generation their are at 36%. To reach 133% improved bandwidth benefit they will need 10 more generations with the same improvement rate (10% per generation). That is beyond 2025 just to match 2016 arch.


It is not all bad. Theoretical peak bandwidth for 380x and 480 is the same 256 GB/s yet in the test 480 is 41% faster uncompressed and 51% compressed performance. 480 reaches 75% of its theoretical bandwidth (uncompressed) - the same level as pascal, while 380x reaches only 54%.


EDIT: Their table is wrong. 380X doesn't have 8gbps memory.
http://techreport.com/review/30328/amd-radeon-rx-480-graphics-card-reviewed/5

 

[Rannar]

It is not all bad. Theoretical peak bandwidth for 380x and 480 is the same 256 GB/s yet in the test 480 is 41% faster uncompressed and 51% compressed performance. 480 reaches 75% of its theoretical bandwidth (uncompressed) - the same level as pascal, while 380x reaches only 54%.

Might want to check your numbers. 380x does not have the same 256GB/s theoretical peak. Most 380x cards are equipped with 5.7Gbps GDDR5 memory.

 

[TheRyuu]

What is very worrying is AMDs first shot at memory compression with 285/380 gives 27% increase in bandwidth. With new generation their are at 36%. To reach 133% improved bandwidth benefit they will need 10 more generations with the same improvement rate (10% per generation). what is beyond 2025 just to match 2016 arch.

It is not all bad. Theoretical peak bandwidth for 380x and 480 is the same 256 GB/s yet in the test 480 is 41% faster uncompressed and 51% compressed performance. 480 reaches 75% of its theoretical bandwidth (uncompressed) - the same level as pascal, while 380x reaches only 54%.

Just keep in mind that this is with an ideal "all black" texture. I can't say I know a lot about real world graphic texture's but I'm not sure how often you'll actually see that ideal situation. It may be that AMD's cards can't match the theoretically best case but they also probably don't fall off a cliff when things aren't ideal. That being said Nvidia's texture compression is almost certainly superior even with real world textures so your point is still sort of valid (although that 10 generations number is almost certainly BS).

I think it's also important to keep in mind the requirements of each of the architectures here. Nvidia's cards have been somewhat bandwidth starved in the Maxwell era so it probably needs the texture compression to pay off more than AMD's cards do/did.

 

[Erenhardt]

Might want to check your numbers. 380x does not have the same 256GB/s theoretical peak. Most 380x cards are equipped with 5.7Gbps GDDR5 memory.

Good catch. I was going by their table:
http://techreport.com/review/30328/amd-radeon-rx-480-graphics-card-reviewed/5
Which seems to be wrong.

Just keep in mind that this is with an ideal "all black" texture. I can't say I know a lot about real world graphic texture's but I'm not sure how often you'll actually see that ideal situation. It may be that AMD's cards can't match the theoretically best case but they also probably don't fall off a cliff when things aren't ideal.

Very true and something to have in mind. It is like double edged sword. Going by their claims thou, it seems very good at finding compressible stuff:

 

[ocre]

That's because they are using the WRONG $.

http://www.newegg.com/Product/Produ...N=100007709 601204369&Tpk=GTX 1070&ignorear=1

Look at the 1070s on Newegg. There's ONE that's $399. One horrible one that's $409. The rest are more expensive.

Until the average price of the card drop down to MSRP, they are doing it wrong on their charts. It's actually a damn lie and straight out of NV marketing guide.

"Here guys, these 1070s are only $379... except you can't buy it for that price. It's out of stock everywhere and the average is closer to $439. But who cares, it's just $379!"

It's been really tiring to see so many attack on Nvidia and their ceo on their pascal 1080 event. It was an absolute bashing party around here. Of course there were statistics and figures that were best case scenarios but I bet I read 10,000 post of negative blasting and claims Nvidia lied. It was ridiculous to read, people going on and on about things like the 1080 faster than 980 sli. Ridiculous because there were plenty of cases where the 1080 was faster than 980 sli, so many in fact that there were several reviews that showed on average the 1080 was indeed faster that 980 sli. But, ever still, this was still blasted-nvidia lied, it's only a few percent faster so it don't count. Or other claims, well it's only because sli scaling so that means Nvidia lies. Best case scenarios have been used for ever know, I remember slides from amd of a bunch of games with truly obscure and bizarre settings that they used to prop up the appearance that their card was faster than an Nvidia card. Only when that card launched, reviews found a completely different result. Yet at those obscure settings, with no af on a strange set up of test systems...It could have been true that they found some rare special cases where their card actually was faster than that Nvidia card. This, which had very different results from reviewers, this they got very little flack over.

The point, my point... things like 1080 faster than 980 sli was not just some obscure and super rare case set up special for a 1080 win. There were reviewers finding the 1080 faster than 980 sli on many counts, and even averaging faster...but Nvidia lied because it was only a few percent.

I honestly am surprised by 480 power consumption. Amds claims of 2.8x and nearly every single person here was positively sure that amd was gonna be so far ahead of pascal in perf per watt that Nvidia was gonna be in serious trouble. Surely you were convinced that the 480 was gonna be more efficient than the gp104 too.

But yet there is crickets when it comes to that. Really, let's try to not be so critical of one while ignoring some seriously way out there stuff that comes from another.

 

[oussama-tn]

If you base it on DX12 and Vulkan games, then for sure RX 480 looks a lot better.

But so does the 390/X. 390X still leads the RX 480 in DX12/Vulkan.

Sure, a lot less power, but damn.. it would have been nice if RX 480 was a 110W gaming load. Change the entire perception and give it a fighting chance vs the upcoming 1060.

Lol i feel your disappointment about the power usage of the rx 480, you probably mentioned that in every post since the first benchmarks appeared