[MindBlankTech]Clock for clock 1st vs 4th gen GCN - Radeon 280X vs RX 470

[PontiacGTX]

 

[Yakk]

The 280x can still hold it's own 4+ years later on top tier games at resolutions it was originally targeted for. Impressive.

The original 7970 was released Dec. 2011 making it 5 years old. The 280x respin brought AMD better binning (which was already very good with the 7950 mostly being 7970 chips with just the bios reducing speed due to lack of "bad" chips) and allowed higher clocks to be more easily attained. Why compare it to a 470, probably should be against a 480 imo.

In any case 280x uses a bit more power and 4 years later can still be considered mainstream. Was an awesome buy considering where its original competition now stands.

 

[AtenRa]

Nice but the 280X is a full Tahiti chip when RX 470 is a cut down Polaris 10. They had to compare the RX 470 to 7950 or 280, or 280X vs RX 480.

 

[.vodka]

Nice but the 280X is a full Tahiti chip when RX 470 is a cut down Polaris 10. They had to compare the RX 470 to 7950 or 280, or 280X vs RX 480.

Nah, cut down P10 has the same specs as full Tahiti, so you can make a clock for clock comparison. Main difference is the memory subsystem, and that's still going to be in favor or P10 because of memory compression even if you match bandwidth on both.


Anyway, nice gains on the same basic architecture after all these years.

 

[PontiacGTX]

The 280x can still hold it's own 4+ years later on top tier games at resolutions it was originally targeted for. Impressive.

The original 7970 was released Dec. 2011 making it 5 years old. The 280x respin brought AMD better binning (which was already very good with the 7950 mostly being 7970 chips with just the bios reducing speed due to lack of "bad" chips) and allowed higher clocks to be more easily attained. Why compare it to a 470, probably should be against a 480 imo.

In any case 280x uses a bit more power and 4 years later can still be considered mainstream. Was an awesome buy considering where its original competition now stands.

the 7950 were the version with 1792 but it didnt unlock to 7970. unlike Hawaii/Fiji do and the reduced clock speed was due to effiency and make a difference in performance given at 900MHz the difference would be very low

the 470 is the similar gpu to the 280x compute wise at same clock aswell SP count

Nice but the 280X is a full Tahiti chip when RX 470 is a cut down Polaris 10. They had to compare the RX 470 to 7950 or 280, or 280X vs RX 480.

the 470 is the similar gpu to the 280x compute wise at same clock aswell SP count

if they compare them that would require to underclock the 480 to achieve similar performance levels and probably woildnt be the same as comparing card with same CUs count

Nah, cut down P10 has the same specs as full Tahiti, so you can make a clock for clock comparison. Main difference is the memory subsystem, and that's still going to be in favor or P10 because of memory compression even if you match bandwidth on both.


Anyway, nice gains on the same basic architecture after all these years.

Not only memory, also ROPs,DMA,GCP,HWS+ACEs,Geometry engine,multimedia engine,

 

[AtenRa]

Ok, they opted for the same CUs as well.

There are some differences though, Tahiti had 384bit memory controllers with higher memory bandwidth than RX 470 at the same 7GHz memory but RX 470 has 4GB of memory vs 3GB for Tahiti.

Nice to see the performance gains at the same CUs/closks but it would be even more eye opening if they would also include power consumption. Because Tahiti at 1150MHz will be extremely power hungry and RX 470 still has more than 10-15% higher clocks to use.

 

[AtenRa]

double post

 

[dogen1]

Not only memory, also ROPs,DMA,GCP,HWS+ACEs,Geometry engine,multimedia engine,

470 has double the geometry processors of 280x. But it does show you that improving geometry performance and instruction prefetching can increase GPU utilization quite a bit.

Should be even better with Vega's improved load balancing. Hopefully there's another ~4TFlop card to do another comparison with.

 

[happy medium]

So 17% better performance on average in 4 years with the same gpu specs?

Why does it sound like this guy thinks that's good?

 

[PontiacGTX]

Ok, they opted for the same CUs as well.

There are some differences though, Tahiti had 384bit memory controllers with higher memory bandwidth than RX 470 at the same 7GHz memory but RX 470 has 4GB of memory vs 3GB for Tahiti.

Nice to see the performance gains at the same CUs/closks but it would be even more eye opening if they would also include power consumption. Because Tahiti at 1150MHz will be extremely power hungry and RX 470 still has more than 10-15% higher clocks to use.

also there is another comparison with GCN3 https://www.computerbase.de/2016-08/amd-radeon-polaris-architektur-performance/2/ there could be a 384 bit Tonga but only there are 256bit card on the market

 

[Bacon1]

So 17% better performance on average in 4 years with the same gpu specs?

Why does it sound like this guy thinks that's good?

Looking at the newer games they were mostly 22%+ improvement. Mafia 3 @ no change brings down the average

 

[happy medium]

Looking at the newer games they were mostly 22%+ improvement. Mafia 3 @ no change brings down the average

Yea your right, I've noticed with some newer games the 280x falling behind the 380(x) and even the gtx960 4gb! and not because of the 3gb it has.

 

[crisium]

How does this compare to Intel CPU IPC improvements? Sandybridge to Skylake is a similar time frame (close to 4.5 years).

It's nice to have IPC improvements, but they need to get past the clock barrier and high power consumption. Typical Tahiti could hit 1.15GHz+ and a golden sample Tahiti chip could hit 1.3GHz on air, iirc. Can a typical Polaris 10 even hit 1.4GHz? Not much improvement. I wouldn't ask AMD to sacrifice the marginal IPC gains, but they really need to find a way to push past the ~1.3GHz barrier.

 

[PontiacGTX]

How does this compare to Intel CPU IPC improvements? Sandybridge to Skylake is a similar time frame (close to 4.5 years).

It's nice to have IPC improvements, but they need to get past the clock barrier and high power consumption. Typical Tahiti could hit 1.15GHz+ and a golden sample Tahiti chip could hit 1.3GHz on air, iirc. Can a typical Polaris 10 even hit 1.4GHz? Not much improvement. I wouldn't ask AMD to sacrifice the marginal IPC gains, but they really need to find a way to push past the ~1.3GHz barrier.

Polaris has some tweaks to improve minimum clock speed(1.2GHz vs 950MHz) but maybe the process needs further improvement to get higher clock speed also for Vega there could be planned something to get higher clock speed while getting less power used on idle SPs

 

[Valantar]

So 17% better performance on average in 4 years with the same gpu specs?

Why does it sound like this guy thinks that's good?

Because that's essentially architecture only - without counting that the 280X is OC'd, the 470 is UC'd, and that the 470 - even OC'd - doesn't come close to the power consumption of the 280X. ~20% clock-for-clock gains in 4 years while significantly reducing power consumption and increasing max clocks, that's a very good result.

Nice but the 280X is a full Tahiti chip when RX 470 is a cut down Polaris 10. They had to compare the RX 470 to 7950 or 280, or 280X vs RX 480.

That makes no sense. Wether a chip is cut-down or not makes no difference to its actual performance, relative to its actual functional components. What he's doing is comparing as similar chips as possible in terms of specs, while clocking them the same too. Which is exactly what you would want him to do. What's the point of comparing dissimilar GPUs simply due to them being, for example, the fully enabled SKU of the mid-range chip of that generation? Then you're not discussing things that make any difference to actual users.

 

[PontiacGTX]

Because that's essentially architecture only - without counting that the 280X is OC'd, the 470 is UC'd, and that the 470 - even OC'd - doesn't come close to the power consumption of the 280X. ~20% clock-for-clock gains in 4 years while significantly reducing power consumption and increasing max clocks, that's a very good result.


That makes no sense. Wether a chip is cut-down or not makes no difference to its actual performance, relative to its actual functional components. What he's doing is comparing as similar chips as possible in terms of specs, while clocking them the same too. Which is exactly what you would want him to do. What's the point of comparing dissimilar GPUs simply due to them being, for example, the fully enabled SKU of the mid-range chip of that generation? Then you're not discussing things that make any difference to actual users.

some benchmarks form anandtech shows really small difference between 7970 and 380x given the clock speed maybe in some games/graphics settings the difference is noticeable but other games arent but in computerbase review there were games which 470 had a lead even over 5% avg anandtech shows

 

[RussianSensation]

The 280x can still hold it's own 4+ years later on top tier games at resolutions it was originally targeted for. Impressive.

The original 7970 was released Dec. 2011 making it 5 years old.

7970 is a legend even though many owners of the opposing brand will never admit to it. 9700Pro and 8800GTX/Ultra are amazing GPUs and will go down in history as some of the best, but no GPU* ever made until 7970 could play many AAA games at 50-60 fps at 1080p over the span of 5 years for as long as HD7970 OC did. 7970 punched way above its class, over the same 5 years outperforming both the 680, 770, GTX960 as well as coming dangerously close to the GTX780 that cost close to double when R9 280X refresh came out.

Computerbase.de shows an 18% increase in IPC from GCN 1 to 4, which is consistent with the review in the OP:
https://www.computerbase.de/2016-08/amd-radeon-polaris-architektur-performance/2/

The best part of all is that HD7970 managed to do all of this while also making $ on the side due to mining, essentially paying for itself multiple times over and allowing the HD7970 gamer to upgrade to R9 290/290X and/or Fury/X, etc.

* Arguably the HD7950 OC is even more special since it could overclock up to 56% (from 800mhz to 1250mhz), thus exceeding the performance of stock HD7970Ghz and GTX680, for a fraction of their price.

So 17% better performance on average in 4 years with the same gpu specs? Why does it sound like this guy thinks that's good?

You do realize AMD has been spending hundreds of millions of dollars and 4+ years designing Zen CPU on the side too, right?

GCN 4 will show greater IPC increases in newer demanding titles and especially in DX12 & GWs games titles which benefit much more from the primitive discard accelerator and more efficient geometry engines.

Total War Warhammer = 19%
Rise of the Tomb Raider = 19%
Black Ops 3 = 20%
AC Syndicate = 23%
Doom = 23%
GTA V = 24%
Mirror's Edge Catalyst = 24%
Just Cause 3 = 24%
The Division = 24%
XCom 2 = 24%
Fallout 4 = 25%
MGS V = 28%
Ashes of Singularity = 29%
The Witcher 3 = 41%

R9 280X/7970Ghz is still faster than GTX960 4GB or GTX1050Ti 4GB in modern games and can be had for $80-90.

Polaris architecture is a nice step-forward for AMD into closing the efficiency gap (perf/clock per shader) and perf/watt with NV. All this should translate into a very competive RX 490 series and hopefully a nice flagship Vega 10 card later on.

If this is true^, my GTX1070s may be my most short-lived GPU setup.

However, NV isn't sweating that Paxwell is just Maxwell+. They learned a great deal from GPU clock optimization and perf/watt over Kepler/Maxwell/Pascal generation and thus DX12 Volta should be NV's next architectural monster, and that's coming within 18 months. AMD better be ready with GCN 5 and another 20-25% IPC increase to offset what Volta may bring to the table.

 

[Valantar]

some benchmarks form anandtech shows really small difference between 7970 and 380x given the clock speed maybe in some games/graphics settings the difference is noticeable but other games arent but in computerbase review there were games which 470 had a lead even over 5% avg anandtech shows

Sorry, what are you saying? I seriously can't understand you. Punctuation might help somewhat.
Youre saying:
a) Anandtech showed small differences between the 7970 and 380x.
b) Computerbase reviewed the RX 470 and it performed ~5% better than ... something? What?
Sorry, but I don't understand your objection to my post. And how does this relate to the video?

 

[poofyhairguy]

So 17% better performance on average in 4 years with the same gpu specs?

Why does it sound like this guy thinks that's good?

Much better than the leap from the 280X to the 380X.

 

[PontiacGTX]

Sorry, what are you saying? I seriously can't understand you. Punctuation might help somewhat.
Youre saying:
a) Anandtech showed small differences between the 7970 and 380x.
b) Computerbase reviewed the RX 470 and it performed ~5% better than ... something? What?
Sorry, but I don't understand your objection to my post. And how does this relate to the video?

that in some benchmarks anandtech has the 380x doing barely 5% better but some other testing has been done on computerbase shows a bigger difference maybe games needs ot be programmed towards GCN3 to get the fullest

 

[Valantar]

that in some benchmarks anandtech has the 380x doing barely 5% better but some other testing has been done on computerbase shows a bigger difference maybe games needs ot be programmed towards GCN3 to get the fullest

I still can't quite grasp the relevance of what you're saying, as neither of those cards are in the video. Sure, the 7970 is essentially the same as the 280X, just an older revision. But it's clocked lower too. Also, no one here is discussing generational gains up to gcn 1.3. We all know it was rather limited. What we're discussing is how Polaris kicked that seeming trend to the curb and shot past previous cards. The 380X is essentially a 285, which didn't outperform the 280X at all at launch (although it performed very similarly with less power consumption).

Also, of course different games perform slightly differently on different architectures. That's just the nature of PC gaming.

None of this either throws shade on or somehow puts into perspective the gains of Polaris. AMD did a great job with GCN 4.

 

[tential]

The problem is we can only figure out what AMD GPUs are good in hindsight due to AMDs poor driver optimization. AMD remains competitive with Nvidia through driver optimization rather than blowing by them by optimizing drivers from the start. As a result, we can look back and say WOW Hawaii is amazing, but if AMD has simply released Hawaii with a good cooler and decent drivers from the start......

That's why AMD is in a bad spot. Because no matter what they do, there is always a deficiency holding them back.

 

[tviceman]

The problem is we can only figure out what AMD GPUs are good in hindsight due to AMDs poor driver optimization. AMD remains competitive with Nvidia through driver optimization rather than blowing by them by optimizing drivers from the start. As a result, we can look back and say WOW Hawaii is amazing, but if AMD has simply released Hawaii with a good cooler and decent drivers from the start.....

I disagree somewhat. Past driver overhead had been a sore spot for AMD, but having xb1 and ps4 running 100% AMD hardware made more if a difference over time, IMO. Once cross platform games started coming out regularly, Kepler tanked and Maxwell's lead shrank. When looking only at MMO's, PC exclusives, and games that were released on only 1 of the two big consoles + PC, performance generally didn't exhibit the large shifts that favored AMD. As of now, I think the advantages of the consoles and hard-coded asynchronous compute are fully realized so that AMD's newer products won't "out live" Nvidia's newer products like Tahiti and Hawaii did over Kepler. In other words, I think the GTX 1060 and RX 480 will remain as competitive in 12-18 months as they are now.

Anyways back on topic... comparing clock for clock performance of different generations of cards is only an exercise in curiosity. It's silly to say that throughput only went up X percentage when not taking into account shader functionality and power consumption.

 

[dogen1]

shader functionality

What do you mean by shader functionality?

 

[tviceman]

What do you mean by shader functionality?

Having a larger feature set, being able to take a bigger advantage of DX12 or more graphics renderings techniques.