[VC][TT] - [Rumor] Radeon Rx 300: Bermuda, Fiji, Grenada, Tonga and Trinidad

[Silverforce11]

While picking individual games and showcasing Tonga's improvements do highlight the architectural strides AMD has made this is largely pointless if, in the grand scheme of things, there is little reflection in average performance.

Edit:

Picking games where Tonga does well is fine. But if for every game that Tonga does well, there is a game that tonga does poorly then things balance out.

It looks like you are having a very difficult time understanding the concept involved in this topic of discussion.

R285 is equivalent to a 7950, there's no "full-fat" Tonga on PC desktops yet (only on MACs). Barring the modifications to the front end for more geometry throughput, better tessellation and fillrate, it's just like Tahiti and so, in games where those key aspects are not the limiting factor, it should perform around the 7950 level.

This is what we seen in many games, because they aren't limited by geometry, fillrate or tessellation.

But in games (Thief, Bioshock, FC4 & maybe more I haven't looked at) where those aspects are important, we see the R285 exceed even the 7970 Ghz, which on paper, is a much stronger GPU.

This is the foundation that AMD can work on. Take Tonga, make it better, meaner, give it an entirely new memory subsystem & interconnect that reduces die size and reduces power consumption, give it low latency high bandwidth HBM vram... it's gonna fly.

I hope this has helped some of you guys make sense of the topic at hand.

 

[antihelten]

While picking individual games and showcasing Tonga's improvements do highlight the architectural strides AMD has made this is largely pointless if, in the grand scheme of things, there is little reflection in average performance.

Edit:

Picking games where Tonga does well is fine. But if for every game that Tonga does well, there is a game that tonga does poorly then things balance out.

This just demonstrates that Tonga is a poorly balanced chip, but RS wasn't making a point about Tonga the chip, but rather about Tonga the architecture aka. GCN 1.2.

 

[Silverforce11]

This just demonstrates that Tonga is a poorly balanced chip, but RS wasn't making a point about Tonga the chip, but rather about Tonga the architecture aka. GCN 1.2.

It's actually not poorly balanced at all. In many games, it performs exactly where it should, around the 7950 mark, while being a lower 256 bit bus part with identical specs (which it should in fact be slower than the 7950!).

I think some of them still fail to grasp the simple concept that they are comparing a neutered Tonga versus a full Tahiti in the R280X or 7970Ghz.

The fact it can beat the full Tahiti in some games is down to the improvements in architecture that AMD put in it. It's an iteration that will work out well for 390X.

 

[antihelten]

It's actually not poorly balanced at all. In many games, it performs exactly where it should, around the 7950 mark, while being a lower 256 bit bus part with identical specs (which it should in fact be slower than the 7950!).

I think some of them still fail to grasp the simple concept that they are comparing a neutered Tonga versus a full Tahiti in the R280X or 7970Ghz.

The fact it can beat the full Tahiti in some games is down to the improvements in architecture that AMD put in it. It's an iteration that will work out well for 390X.

True, to be honest I was going on memory and I just looked through one of the latest TPU reviews and the 285 does a fair bit better than I remember from launch reviews. There are still places where it gets beaten by a 7950 (AC:U - unplayable fps for both cards though), but all in all it does quite well. Would be interesting to see what a full Tonga chip is capable of.

 

[Silverforce11]

Just wanted to correct this, but HBM doesn't actually bring any power efficiency improvements compared to GDDR5, quite the contrary actually (at the setups that will be used in GPUs).

I know that most people have probably seen the claims of a 40% reduction in power usage for HBM compared to GDDR5, however that reduction is on a mW/Gbps/pin basis (source), and since HBM features a massive increase in the number of I/O pins, the net result can very easily be an increase in power.

So in the case of Nvidia they will be going from 256 pins, to most likely 4096 pins (4 1GB HBM modules), and from 7 Gbps to 1Gbps. This will result in a 33% increase in power, not decrease (0.58 * 4096pins/256pins * 1Gbps/7Gbps = 1.33).

For AMD it will be a 7% decrease in power, since they already use a very wide 512 bit bus with Hawaii (0.58 * 4096pins/512pins * 1Gbps/5Gbps = 0.93).

A 33% increase for Nvidia and a 7% decrease for AMD will probably result in the current power gap shrinking by about 10W or so (assuming 4GB of GDDR5 at 5-7 Gbps uses around 20-30 watts).

Now the above is just the memory modules, and obviously the PHY will also change significantly with HBM, which should result in significant power reduction for both AMD and Nvidia. However AMD should once again see the bigger gain, since their current 512 bit interface is probably more power hungry than Nvidias 256 bit interface.

All in all the switch to HBM could on it's own shrink the gap in power usage between AMD and Nvidia, by some 15-20 watts.

It's not that simple because the entire memory subsystem within the GPU is changed. It's not just about i/o pins to the vram. AMD has presented before, that the HBM memory controller is a lot smaller and leaner, saving die space and power. Then the actual memory chip itself also saves power. Its been estimated that the typical power use of a GPU, ~1/3rd of that is due to the memory controller alone. For a R290X, thats ~80W.

 

[Enigmoid]

I understand what RS is saying.

I'm saying that unless this affects enough games this simply isn't going to make a large difference.

Sure Tonga brings improvements but unless these improvements are going to be seen in more than a handful of games so what.

Looking at the relatively small fraction of games where Tonga does well does not excuse the relatively large number where it slots in exactly as it should.

It looks like you are having a very difficult time understanding the concept involved in this topic of discussion.

R285 is equivalent to a 7950, there's no "full-fat" Tonga on PC desktops yet (only on MACs). Barring the modifications to the front end for more geometry throughput, better tessellation and fillrate, it's just like Tahiti and so, in games where those key aspects are not the limiting factor, it should perform around the 7950 level.

This is what we seen in many games, because they aren't limited by geometry, fillrate or tessellation.

But in games (Thief, Bioshock, FC4 & maybe more I haven't looked at) where those aspects are important, we see the R285 exceed even the 7970 Ghz, which on paper, is a much stronger GPU.

This is the foundation that AMD can work on. Take Tonga, make it better, meaner, give it an entirely new memory subsystem & interconnect that reduces die size and reduces power consumption, give it low latency high bandwidth HBM vram... it's gonna fly.

I hope this has helped some of you guys make sense of the topic at hand.

It will fly on the game where that is a problem. On other games unless the 390X brings new improvements it will slot in exactly as expected. You say yourself that this won't happen in most cases.

Stop taking these gains that will apply sometimes and apply them everywhere.

You said this exactly on post #402

I hope you all realize the R285 is actually the cut down Tonga, the full die Tonga is all being sold to Apple for Macs. Thus the 285 is more akin to the 7950, it should have no right beating a 7970Ghz.

But there are improvements in there which when taken advantage of by games (Bioshock, FC4 with enhanced tessellated godrays), lead to a major performance gain compared to Tahiti.

If all R390X is was a 25% bigger Hawaii with Tonga's front end improvements (more throughput & better tessellation), memory compression and HBM (saving die space & TDP) low latency/high bandwidth vram, it should easily be ~50% faster than R290X.

The best case scenario will apply but it will not apply everywhere all the time.

I understand Tonga's improvements. They are quite nice and visible in the real world. But they need to be present most or all of the time to make a huge difference. They are easily seen in a few games but there are many games where they are not.

I would also like to point out that people seem to be counting memory compression and HBM twice. Going from 320 GB/sec to 640 GB/sec doubles bandwidth. Yet the shader count only increases from 2816 to 4096 shaders. You already are pretty much not limited by bandwidth at all from that increase. Adding memory compression does little when you already have so much bandwidth in excess.

 

[antihelten]

It's not that simple because the entire memory subsystem within the GPU is changed. It's not just about i/o pins to the vram. AMD has presented before, that the HBM memory controller is a lot smaller and leaner, saving die space and power. Then the actual memory chip itself also saves power. Its been estimated that the typical power use of a GPU, ~1/3rd of that is due to the memory controller alone. For a R290X, thats ~80W.

I know, and I did touch upon it a bit in the second to last paragraph, although I didn't mention memory controllers specifically.

This paper seems to suggest that memory controllers are responsible for 15-25% of the power usage btw

 

[Silverforce11]

I understand what RS is saying.

I'm saying that unless this affects enough games this simply isn't going to make a large difference.

Sure Tonga brings improvements but unless these improvements are going to be seen in more than a handful of games so what.

Looking at the relatively small fraction of games where Tonga does well does not excuse the relatively large number where it slots in exactly as it should.

It will fly on the game where that is a problem. On other games unless the 390X brings new improvements it will slot in exactly as expected. You say yourself that this won't happen in most cases.

Stop taking these gains that will apply sometimes and apply them everywhere.

You said this exactly on post #402

The best case scenario will apply but it will not apply everywhere all the time.

I understand Tonga's improvements. They are quite nice and visible in the real world. But they need to be present most or all of the time to make a huge difference. They are easily seen in a few games but there are many games where they are not.

I would also like to point out that people seem to be counting memory compression and HBM twice. Going from 320 GB/sec to 640 GB/sec doubles bandwidth. Yet the shader count only increases from 2816 to 4096 shaders. You already are pretty much not limited by bandwidth at all from that increase. Adding memory compression does little when you already have so much bandwidth in excess.

As been stated, there are improvements in Tonga that if applied in games, lead to major performance boosts. I don't know for certain that more future games will be geometry, fillrate or tessellation heavy. But I do know that games will inevitably get more complex, and one of the easiest ways is via higher geometry models or heavy tessellation. FC4's godrays using tessellation is a novel application of that feature, so its not far fetched to see future Ubi titles at least use tessellation to enhance not just models, but lighting.

These are weak points of GCN, that Tonga has addressed. It can only get better.

Keep in mind regarding the HBM advantage, I know from mining for awhile on many GPUs that the shaders are very sensitive to latency. Decreasing vram latency but keeping bandwidth the same or even reducing it actually leads to massive performance gains. The theory is it keeps the shaders fed more (less stalling waiting for data to crunch), thus, higher uptime boosts performance/efficiency.

Is it far fetched to see a 25% bigger Hawaii incorporating Tonga's improvements, with a new memory subsystem and HBM performing very well? I don't think so.

 

[antihelten]

I understand what RS is saying.

I'm saying that unless this affects enough games this simply isn't going to make a large difference.

Sure Tonga brings improvements but unless these improvements are going to be seen in more than a handful of games so what.

Luckily these improvements are in fact seen in more than a handful of games.

In the latest TPU review with 285 numbers, the 285 beats the 7950 by at least 5% in 17 out of 20 games, most often by around 10%, and with several titles being 15 to 40% faster (COD:AW, Civ:BE, GRID 2, Metro:LL, Watch Dogs and Wolfenstein), a level of increase I would consider quite significant.*

*I looked at the 1440p numbers btw

 

[Enigmoid]

Luckily these improvements are in fact seen in more than a handful of games.

In the latest TPU review with 285 numbers, the 285 beats the 7950 by at least 5% in 17 out of 20 games, most often by around 10%, and with several titles being 15 to 40% faster (COD:AW, Civ:BE, GRID 2, Metro:LL, Watch Dogs and Wolfenstein), a level of increase I would consider quite significant.*

*I looked at the 1440p numbers btw

The 7950 numbers in TPU are the 7950 without boost.

1792 shaders @ 800 mhz.

Based on clockspeed alone you would expect Tonga to be ahead.

 

[antihelten]

The 7950 numbers in TPU are the 7950 without boost.

1792 shaders @ 800 mhz.

Based on clockspeed alone you would expect Tonga to be ahead.

Didn't think of that, my bad. Assuming a performance difference between 7950 and 7950 boost of about 5% (boost review), then that would reduce the number of games in which the 285 is at least 5% faster from 17 out of 20, to 12 of 20, still a substantial amount. And of course there are still games where the 285 is up to 30% faster.

 

[ocre]

You are now the 3rd poster in this thread for whom the architectural comparisons of 285 vs. 7970 flew well over your head when the discussion had 0 to do with 285's perf/watt vs 7970 or their respective average performance standing.

As I said to others who clearly missed the whole point of Tonga
vs. 280x/290x comparison that you take some time to carefully read the last 3 pages of this thread to actually understand what's being discussed. You should actually read what's stated and not assume things and ideas that aren't even being debated.

He was 100% incorrect since his posts couldn't at all comprehend how AMD could improve with R9 300 series on the same node. He doesn't even acknowledge real architectural advancements in Hawaii's memory controller or Tonga's tessellation, colour/memory compression. His rebuttal of linking 285's perf/watt is a waste of time because 285 is not a fully unlocked product and because 390X will have a completely different memory controller and memory type. Basically he wasted space by posting 285's perf/watt as some "proof" that AMD cannot improve perf/watt with 390 cards by resting his argument on the point that 285 couldn't improve upon 7970. What happens when AMD takes all the architectural advancements in Tonga, add far more efficient HBM on a bandwidth/mm2 and watt basis and incorporates other changes they didn't have time to add to Tonga? Do you honestly believe 390X won't have superior perf/watt to a 290X? Since 290X uses about 270W, this is a done deal unless you think 390X will only be 11% faster?

The thing with that poster is that he never posted even 1 positive thing about any AMD products since he joined. Most people who followed his posts for the last 2.5 years pretty much know he will post nothing positive regarding anything AMD. In every thread on AMD he posts negativity.

----
I am interested in seeing how AMD will address the active DP transistors for gaming. NV has solved this issue but it seems this has burdened Hawaii and Tahiti significantly. If AMD manages to improve finer voltage granularity and keep those DP units idle in games, even with no changes to the actual architecture they will already achieve a good increase in perf/watt.

You changed your post, guess that one way to win. :thumbsdown:

Anyway, I am not in disagreement when you say AMD can improve their architecture. I also can see there are improvements in tonga but i dont support cherry picking no matter how much you justify it. You also changed your post so I guess its kind of pointless for me to even respond now.

If we want to talk about architectures and improvements, we have to see what they mean in the real world. A synthetic or a random case here or there does isn't enough. Others have tried to say this but then they too get attacked, "you just don't understand" or " you fail to comprehend the word has been spoken"

If you want to talk about an architecture, than do it right.

I think AMD set out to improve with tonga but the results are a mixed bag for sure. But most importantly, if we talk about why tonga was brought up then we are talking about efficiency. So, back on that topic....

I wasn't impressed at all with the real world results of tonga when it comes to efficiency, especially when you look at how hard it was being pushed. Efficiency was the center of its marketing.

When it comes to tonga efficiency, I was mostly disappointed. I have seen plenty of data showing tonga to be less efficient than Tahiti. Sometimes its on par but to say its this huge improvement was a complete lie. But it was being passed of as such.

When we look at new architectures, it takes more than one game or one case to fully appreciate it. That's the only way to really have a worthy discussion.
Tonga was a let down to me. Its not only that, all this memory compression and improvements made to help out the chip with limited bandwidth......
I have to ask how those improvements really matter much to a chip that will be using HBM stacked memory. Bandwidth is not gonna be an issue for Fiji. This doesn't mean I don't think its important for AMD to work on that stuff, cause it really is especially for their lower skus, dgpus, and such.. I am just saying its not something that matters to Fiji, I really don't think like e thought of using tonga to pump up Fiji.. I just don't think it does and there are far better chips accomplishments we can talk about.

Hawaii was a great chip. It really was a great accomplishment of engineering. One that everyone should have been proud of. But AMD sabotaged it right out of the gate. A massive misstep that I still have a hard time understanding. Just a few post ago someone was claiming that pro nvidia people point to Hawaii reference to dog on it, but this was AMDs blunder. All the way! Its not so much that they released such a terrible reference cooled design, its the fact that there was no other option for so so long. This really hurt the image of an otherwise very impressive product.

So when we look back, I think Hawaii or Tahiti is much more worth talking about than tonga. The 5870 is another really great accomplishment. But it is not so much the cards, they are important but its how you play them. AMD is fully capable of designing great architectures and they have proven this time and time again. But what they can't afford is any missteps. Not at this point. The stakes are very very high.

So, there is more to this discussion when we talk a out Fiji. We have seen that great chips aren't enough. Just as important is how AMD plays their hand. This is where we should all be concerned...
At least, that is what I think.

AMD needs to really sit down and have a huge discussion. Their engineers can make great chips, we just need AMD to be firing on all cylinders this next round

 

[Enigmoid]

Didn't think of that, my bad. Assuming a performance difference between 7950 and 7950 boost of about 5% (boost review), then that would reduce the number of games in which the 285 is at least 5% faster from 17 out of 20, to 12 of 20, still a substantial amount. And of course there are still games where the 285 is up to 30% faster.

I think the better comparison is really Pitcarin. Similar bandwidth (which is why the 285 does badly in a couple games) makes it much easier to isolate performance. Furthermore, Tahiti is quite unbalanced (ROP limited) making comparisons to Hawaii difficult (Hawaii has fixed this bottleneck).

As far as efficiency and performance/shader goes, tahiti is on the bottom of the stack. Furthermore, the many of the gains that Tonga has over Tahiti will not be present in GCN 1.1 Hawaii. You can't really take a number of improvements over the worst version of GCN 1.0 and expect them to apply equally when comparing to GCN 1.1.

Tonga is nice, however, given the timeframe (nearly 3 years after Tahiti) I feel that it is nothing terribly impressive. Lets face it, more than 2.5 years later I was expecting something radical which tonga decidedly is not (not nearly to the extent VLIW4-> GCN was or Fermi -> Kepler -> Maxwell). I expect the 390x to shake things up but its really behind schedule.

 

[Silverforce11]

Tonga is nice, however, given the timeframe (nearly 3 years after Tahiti) I feel that it is nothing terribly impressive. Lets face it, more than 2.5 years later I was expecting something radical which tonga decidedly is not (not nearly to the extent VLIW4-> GCN was or Fermi -> Kepler -> Maxwell). I expect the 390x to shake things up but its really behind schedule.

That, we can agree upon.

AMD needs 390X last month, not in June.

 

[3DVagabond]

The 7950 numbers in TPU are the 7950 without boost.

1792 shaders @ 800 mhz.

Based on clockspeed alone you would expect Tonga to be ahead.

Those benches would have been updated since the original 7950 release reviews. I know W1zzard isn't exactly the best with keeping his numbers current, but I really really doubt he's still using original 7950 numbers.

 

[3DVagabond]

You changed your post, guess that one way to win. :thumbsdown:

Anyway, I am not in disagreement when you say AMD can improve their architecture. I also can see there are improvements in tonga but i dont support cherry picking no matter how much you justify it. You also changed your post so I guess its kind of pointless for me to even respond now.

If we want to talk about architectures and improvements, we have to see what they mean in the real world. A synthetic or a random case here or there does isn't enough. Others have tried to say this but then they too get attacked, "you just don't understand" or " you fail to comprehend the word has been spoken"

If you want to talk about an architecture, than do it right.

I think AMD set out to improve with tonga but the results are a mixed bag for sure. But most importantly, if we talk about why tonga was brought up then we are talking about efficiency. So, back on that topic....

I wasn't impressed at all with the real world results of tonga when it comes to efficiency, especially when you look at how hard it was being pushed. Efficiency was the center of its marketing.

When it comes to tonga efficiency, I was mostly disappointed. I have seen plenty of data showing tonga to be less efficient than Tahiti. Sometimes its on par but to say its this huge improvement was a complete lie. But it was being passed of as such.

When we look at new architectures, it takes more than one game or one case to fully appreciate it. That's the only way to really have a worthy discussion.
Tonga was a let down to me. Its not only that, all this memory compression and improvements made to help out the chip with limited bandwidth......
I have to ask how those improvements really matter much to a chip that will be using HBM stacked memory. Bandwidth is not gonna be an issue for Fiji. This doesn't mean I don't think its important for AMD to work on that stuff, cause it really is especially for their lower skus, dgpus, and such.. I am just saying its not something that matters to Fiji, I really don't think like e thought of using tonga to pump up Fiji.. I just don't think it does and there are far better chips accomplishments we can talk about.

Hawaii was a great chip. It really was a great accomplishment of engineering. One that everyone should have been proud of. But AMD sabotaged it right out of the gate. A massive misstep that I still have a hard time understanding. Just a few post ago someone was claiming that pro nvidia people point to Hawaii reference to dog on it, but this was AMDs blunder. All the way! Its not so much that they released such a terrible reference cooled design, its the fact that there was no other option for so so long. This really hurt the image of an otherwise very impressive product.

So when we look back, I think Hawaii or Tahiti is much more worth talking about than tonga. The 5870 is another really great accomplishment. But it is not so much the cards, they are important but its how you play them. AMD is fully capable of designing great architectures and they have proven this time and time again. But what they can't afford is any missteps. Not at this point. The stakes are very very high.

So, there is more to this discussion when we talk a out Fiji. We have seen that great chips aren't enough. Just as important is how AMD plays their hand. This is where we should all be concerned...
At least, that is what I think.

AMD needs to really sit down and have a huge discussion. Their engineers can make great chips, we just need AMD to be firing on all cylinders this next round

You really don't get it. You have just said the same thing you've been saying all along. Sorry if that makes you feel like I'm attacking you, but it really isn't intended as such.

RS simply showed situations that took advantage of the improvements. Then he went on to theorize that if other bottlenecks are removed the improvements can be more wide spread. He never said that Tonga was better on average. Others have later set that bar which wasn't his position at all. That bar has been set to circumvent his arguments rather than take them at face value. If anyone has been attacked here, it's RS.

 

[Enigmoid]

Those benches would have been updated since the original 7950 release reviews. I know W1zzard isn't exactly the best with keeping his numbers current, but I really really doubt he's still using original 7950 numbers.

Nope its the 7950 non boost. Otherwise it would be closer to the 7970.

7970 is 11% faster than the 7950 Boost, 18% faster than non-boost.

Yet in later graphs, the 7970 is 19% faster than the 7950.

Its the non-boost.

 

[antihelten]

I think the better comparison is really Pitcarin. Similar bandwidth (which is why the 285 does badly in a couple games) makes it much easier to isolate performance. Furthermore, Tahiti is quite unbalanced (ROP limited) making comparisons to Hawaii difficult (Hawaii has fixed this bottleneck).

I don't really see how Pitcairn (or Curacao) is a better comparison, sure you have the same bus width, but the number of shaders and TMUs are completely different, so you just move the uncertainty from one area to another. Furthermore Tahiti may very well be ROP limited, but then so is Tonga (same ratio of ROPs for 7950 and 285, and in theory for 7970 and 285x), unless there have been architectural improvements with GCN 1.2 (the thing RS was alluding to in the first place).

As far as efficiency and performance/shader goes, tahiti is on the bottom of the stack. Furthermore, the many of the gains that Tonga has over Tahiti will not be present in GCN 1.1 Hawaii.

You are right the gains are not necessarily present in Hawaii, but then this thread is about Fiji not Hawaii, and it is probably safe to assume that Fiji will not be GCN 1.1, but rather 1.2 (or 1.3?)

You can't really take a number of improvements over the worst version of GCN 1.0 and expect them to apply equally when comparing to GCN 1.1.

No, and no one ever claimed that to be the case. It would of course be optimal if we could compare GCN 1.2 directly to GCN 1.1, but the specs between Tonga and Hawaii are simply too different, to say anything conclusive.

Nope its the 7950 non boost. Otherwise it would be closer to the 7970.

7970 is 11% faster than the 7950 Boost, 18% faster than non-boost.

Yet in later graphs, the 7970 is 19% faster than the 7950.

Its the non-boost.

It doesn't really matter whether it's boost or non-boost, the performance difference is small enough between the two versions that it doesn't really change the fact that the 285 outperforms the 7950 in the majority of the tested games, and sometimes by large margins, so obviously something has happened architecturally speaking.

 

[Enigmoid]

I don't really see how Pitcairn (or Curacao) is a better comparison, sure you have the same bus width, but the number of shaders and TMUs are completely different, so you just move the uncertainty from one area to another. Furthermore Tahiti may very well be ROP limited, but then so is Tonga (same ratio of ROPs for 7950 and 285, and in theory for 7970 and 285x), unless there have been architectural improvements with GCN 1.2 (the thing RS was alluding to in the first place).

You are right the gains are not necessarily present in Hawaii, but then this thread is about Fiji not Hawaii, and it is probably safe to assume that Fiji will not be GCN 1.1, but rather 1.2 (or 1.3?)

No, and no one ever claimed that to be the case. It would of course be optimal if we could compare GCN 1.2 directly to GCN 1.1, but the specs between Tonga and Hawaii are simply too different, to say anything conclusive.

It doesn't really matter whether it's boost or non-boost, the performance difference is small enough between the two versions that it doesn't really change the fact that the 285 outperforms the 7950 in the majority of the tested games, and sometimes by large margins, so obviously something has happened architecturally speaking.

Tonga's ROPs are likely at the very least slightly modified given the massive increase in fillrate. I'm also talking about this in the performance increase over Hawaii which is not ROP limited (no gains there).

Sure this thread is about Fiji but the comparison is being made to Tahiti and Hawaii. There is nothing wrong with making that comparison in speculation about Fiji's performance. Fiji will for sure be GCN 1.2 possibly 1.3 (I hope) but you cannot look at the differences between gcn 1.2 and 1.0 and expect the same difference between 1.2 and 1.1. I'm comparing with Hawaii because that is how everyone on this thread is estimating performance (ie + 40% over 290X).

You know that the 7950 non boost runs at 800 mhz? Well below the 285 at 918 mhz boost. The 7950 boost runs at 850 mhz with 925 mhz boost. An increase in core clock of nearly 16% is nothing insignificant from the 7950 to its boost version. Furthermore, comparing the 285 to the 7950 non boost the 285 is expected to be ahead given its 15% higher FLOP rating.

 

[antihelten]

Sure this thread is about Fiji but the comparison is being made to Tahiti and Hawaii. There is nothing wrong with making that comparison in speculation about Fiji's performance. Fiji will for sure be GCN 1.2 possibly 1.3 (I hope) but you cannot look at the differences between gcn 1.2 and 1.0 and expect the same difference between 1.2 and 1.1. I'm comparing with Hawaii because that is how everyone on this thread is estimating performance (ie + 40% over 290X).

I absolutely agree that you can't extend the difference between 1.2 and 1.0, to 1.2 (1.3?) versus 1.1, you'll find no argument from me there. However I would think that RS is perfectly aware of this, and if a suitable comparison between Hawaii and Tonga could be made, then he would have done that instead, but again such a comparison can't really be made due to the differences in spec. So all in all the comparison between Tonga and Tahiti, might not be optimal, but it is arguably the only one that we can meaningfully make.

You know that the 7950 non boost runs at 800 mhz? Well below the 285 at 918 mhz boost. The 7950 boost runs at 850 mhz with 925 mhz boost. An increase in core clock of nearly 16% is nothing insignificant from the 7950 to its boost version. Furthermore, comparing the 285 to the 7950 non boost the 285 is expected to be ahead given its 15% higher FLOP rating.

I don't really care what clock rate the different cards run at since performance does not scale linearly with frequency, as clearly indicated by actual benchmarks. So no the 285 is not expected to be ahead of the 7950 simply because of higher FLOP rating, since again performance is not necessarily proportional to FLOP rating (case in point, the 290X has a 22% higher FLOP rating than the GTX 980)

 

[ocre]

You really don't get it. You have just said the same thing you've been saying all along. Sorry if that makes you feel like I'm attacking you, but it really isn't intended as such.

RS simply showed situations that took advantage of the improvements. Then he went on to theorize that if other bottlenecks are removed the improvements can be more wide spread. He never said that Tonga was better on average. Others have later set that bar which wasn't his position at all. That bar has been set to circumvent his arguments rather than take them at face value. If anyone has been attacked here, it's RS.

I don't get it?

You should read my post again. He changed his post so maybe its that your not getting.

My post is to him and I wanted to share my thoughts about tonga efficiency, and tonga in general. Originally that is where the conversation was, we were talking about efficiency and how some people don't think AMD can improve their situation. RS was claiming they were wrong to think like that and offered several designs as proof AMD can improve their situation and not to count them out. He brought up tonga and that is what I have to say about it. The fact that his words are now different (edited), that doesn't change how I feel.

My post is to RS and its how I feel about tonga and AMDs situation. I appreciate other architectures much more and I think that tonga is one of the worst ones I would use when talking about AMD doing great things

 

[Enigmoid]

I don't really care what clock rate the different cards run at since performance does not scale linearly with frequency, as clearly indicated by actual benchmarks. So no the 285 is not expected to be ahead of the 7950 simply because of higher FLOP rating, since again performance is not necessarily proportional to FLOP rating (case in point, the 290X has a 22% higher FLOP rating than the GTX 980)

Did I say there was linear scaling? No I did not. Needless to say, raising the core clock will increase the performance as long as you are not completely memory bandwidth bound (which the 7950 is not). However, raising core clocks will increase performance on the 7950 which obviously has happened.

Well no crap you can't compare FLOPS across radically different architectures, however, Tonga and Tahiti are relatively similar, furthermore, this is our goal; namely to compare performance normalized to shader count and frequency to observe architectural improvments.

If you don't care about the FLOP rating (which is a function of the number of shaders, the operations each shader can do per cycle, and the clockspeed) or the core speed then in effect you do not care about the number of shaders which makes a comparison impossible.

 

[antihelten]

Did I say there was linear scaling? No I did not. Needless to say, raising the core clock will increase the performance as long as you are not completely memory bandwidth bound (which the 7950 is not). However, raising core clocks will increase performance on the 7950 which obviously has happened.

Raising the clock from 800 MHz on the 7950 to 925 MHz on the 7950, a 15.6% increase, provided a 5.3% performance boost (TPU).

My point is that starting to talk about FLOPS, when we have actual benchmarks available is a bit silly, especially when you already know how unreliable that metric is to begin with.

If you don't care about the FLOP rating (which is a function of the number of shaders, the operations each shader can do per cycle, and the clockspeed) or the core speed then in effect you do not care about the number of shaders which makes a comparison impossible.

It's not that I don't care about FLOPS or frequency or number of shaders or any of the other specs of a given GPU, it's that I don't care about any of them in isolation, which is basically what did when you said that we could except something solely based on an increase in FLOPS.

 

[Enigmoid]

You have to put things in isolation to determine architectural improvements. How can you compare performance per shader if you don't know (or care) about then number of shaders or the clockspeed of the items being compared?

There is no guarentee that an increase in FLOPS will necessarily increase performance, however, there is a general correlation, a correlation which becomes quite strong when comparing very similar architectures.

 

[Cloudfire777]

A little bird tweeted this today. Its been posted in several sites so it could be legit.

380x 319 달러
390x 499 달러

달러 = Dollar