• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

GCN 1 vs GCN 3 vs GCN 4

LordOfChaos

Member
Computerbase.de has made an interesting comparison between the Radeon R9 280X, Radeon R9 380X and Radeon RX 470.
All clocked at the same frequency and configured with similar memory bandwidth. All three have 2048 shader units.

It's in German but chrome auto translate works pretty well with it as the German-English sentence structure is so similar. Plus, graphs explain themselves.

https://www.computerbase.de/2016-08/amd-radeon-polaris-architektur-performance/

On some games the difference is huge, like The Witcher 3 which shows 41% improvement over GCN 1, all at the same core count and clock speed.

which, ya know, makes me even sadder that the most money I could throw at a macbook would get me a GCN 1.0 GPU, which is hardly an update from my mid 2014...Grumble grumble
 

Thraktor

Member
which, ya know, makes me even sadder that the most money I could throw at a macbook would get me a GCN 1.0 GPU, which is hardly an update from my mid 2014...Grumble grumble

A new line of MBPs with Polaris GPUs is due soon enough, and if you're using a mid 2014 model then you can quit your grumbling, I'm on an early 2011 15" and it's still got a good few more years of life in it before I put it out to pasture.

Back on topic, though, it's pretty interesting to see the real-world performance impact of the various architectural improvements over the years. I'd be very interested to see them dive into the graphical settings in the games to find out where the performance gains were largest, but I'm aware that would be a hell of a lot of work.
 
It was interesting to note as they did how much bandwidth the latest Poralris needs to maximise its ALU utilisation. It seems like the RX480 is a bit unbalanced in that regard.
 

LordOfChaos

Member
A new line of MBPs with Polaris GPUs is due soon enough, and if you're using a mid 2014 model then you can quit your grumbling, I'm on an early 2011 15" and it's still got a good few more years of life in it before I put it out to pasture.



I was planning on reselling it and losing a few hundred to upgrade, of course there are years of life left in it before it's driven into the ground :p

I guess I'm just surprised they never stopped this multiyear gap by ordering a very slightly customized GCN 1.3 (now GCN 3) part for the right wattage, as they already did slightly custom GPUs for the Mac Pro.

I guess though that just like the still on Haswell CPU, they were waiting for a single larger marketable upgrade.
 

Inuhanyou

Believes Dragon Quest is a franchise managed by Sony
NEO should benefit even more from this, as PS4's GPU pipleline was based on GCN 1.1(which was basically GCN 1 with tighter intergrated GPGPU customization)

So GCN 1.4 having that kind of impact architecture wise, the tflop difference from 1.84 to 4.2 should have a greater impact than just doubling PS4's GPU and calling it a day
 

LordOfChaos

Member
NEO should benefit even more from this, as PS4's GPU pipleline was based on GCN 1.1(which was basically GCN 1 with tighter intergrated GPGPU customization)

So GCN 1.4 having that kind of impact architecture wise, the tflop difference from 1.84 to 4.2 should have a greater impact than just doubling PS4's GPU and calling it a day

Though only Sea Islands/1.2/2 on Desktop got the 8 ACEs of the PS4, so it was a bit borrowed from the future (kind of like Xenos).

They're now moving even beyond that to 4 ACEs plus two HWS (hardware schedulers) which do the work of 2 ACEs each, or something like that I think.
 

Inuhanyou

Believes Dragon Quest is a franchise managed by Sony
Though only Sea Islands/1.2/2 on Desktop got the 8 ACEs of the PS4, so it was a bit borrowed from the future (kind of like Xenos).

They're now moving even beyond that to 4 ACEs plus two HWS (hardware schedulers) which do the work of 2 ACEs each, or something like that I think.

Yep. Its hard to belive we've been through 4 GPU generations since the PS4 launched(including the non consumer 8000 series)
 

entremet

Member
Confusing since ATI(AMD)also designed the Gamecube graphics processor, which was used for three generations.
 

LordOfChaos

Member
Confusing since ATI(AMD)also designed the Gamecube graphics processor, which was used for three generations.

It was on-die for the Wii U but that at least moved to a unified shader design from the HD4000 era. The Gamecube GPU wasn't the hardware running Wii U games.
 

Larogue

Member
What GCN Scorpio gonna get? Sounds like each TF worth more in real life performance as more advancement is made in the core technology.

Can we expect 6 TF Scorpio (new core) = 8-9 TF of XB1 GCN in real life performance.
 

wachie

Member
What GCN Scorpio gonna get? Sounds like each TF worth more in real life performance as more advancement is made in the core technology.

Can we expect 6 TF Scorpio (new core) = 8-9 TF of XB1 GCN in real life performance.
Based on those gains, more like 6.6 TF of XB1.
 
Also RX 470 is only 4% less than a RX 480 at the same clocks, AMD seems to just have lowered the clocks to get a lower the tdp, as both cards seem to get around the same clocks on overclock. Maybe people waiting for stock on a RX 480 or the mythical 4GB RX 480 should just get a "discounted RX 480" aka RX 470.
 

Inuhanyou

Believes Dragon Quest is a franchise managed by Sony
What GCN Scorpio gonna get? Sounds like each TF worth more in real life performance as more advancement is made in the core technology.

Can we expect 6 TF Scorpio (new core) = 8-9 TF of XB1 GCN in real life performance.

We're expecting Vega, so who knows. But its going to be a decent jump from Southern Islands/Sea islands architecture.
 
The PS4 Neo should benefit greatly from Polaris's(GCN4) Lossless Delta Framebuffer Compression as well -- it should increase effective memory bandwidth significantly.
 

LordOfChaos

Member
We're expecting Vega, so who knows. But its going to be a decent jump from Southern Islands/Sea islands architecture.

Vega pretty much is a larger Polaris (GCN 4) with HBM2 I believe. The Scorpio version has GDDR5X instead, so basically a large Polaris die. Per clock per core then I'd expect similar to GCN 4 in the above article, so up to 40% more performance per paper flop than GCN 1.

The PS4 Neo should benefit greatly from Polaris's(GCN4) Lossless Delta Framebuffer Compression as well -- it should increase effective memory bandwidth significantly.

True, the only 20% higher memory bandwidth threw me for the over double more (paper) GPU execution power, but the compression seems so significantly better it may be enough.

It's almost like they thought of this stuff and I should stfu!
 

Proelite

Member
This test is misleading since the GCN1.1 or lower is bottlenecked by low bandwidth, where as the newest GCN has the bandwidth compression. They should redo the test with GCN1.1 or lower test run with the highest possible bandwidth so as to not bottleneck any of the configurations.

Run the tests with only half the CUs and ROPS enabled.
 

Inuhanyou

Believes Dragon Quest is a franchise managed by Sony
This test is misleading since the GCN1.1 or lower is bottlenecked by low bandwidth, where as the newest GCN has the bandwidth compression. They should redo the test with GCN1.1 or lower test run with the highest possible bandwidth so as to not bottleneck any of the configurations.

Run the tests with only half the CUs and ROPS enabled.

But that's part of the architecture change to give significant boosts. That's like saying a GPU test is borked because the older GPU doesn't have more CU's to offset the architecture advancements, that's the point
 

LordOfChaos

Member
This test is misleading since the GCN1.1 or lower is bottlenecked by low bandwidth, where as the newest GCN has the bandwidth compression. They should redo the test with GCN1.1 or lower test run with the highest possible bandwidth so as to not bottleneck any of the configurations.

Run the tests with only half the CUs and ROPS enabled.

How does that make sense to do? GCN 1 had a bottleneck somewhere, GCN 4 removed said bottleneck and now its cores are able to do 40% better per core per clock. That doesn't seem misleading to me.

It's like when Fermi cards were limited by memory capacity, you could say one should do tests where they aren't, but the fact is that memory capacity limited them early, so regardless of where the bottleneck is, performance will be X in testing.
 

Proelite

Member
But that's part of the architecture change to give significant boosts. That's like saying a GPU test is borked because the older GPU doesn't have more CU's to offset the architecture advancements, that's the point

Yes, but this doesn't mean that a 4 teraflop GCN 1.3 card will perform like a 6 teraflop GCN 1.1 card, which is what some people trying to infer for the Neo.

The NEO gpu will mostly perform like a 4.2 teraflop GCN 1.0 / 1.1 GPU with no memory bottleneck.

Case in point, look at R290x vs RX 480. The GCN 1.1 teraflops are performing about the same as the GCN 1.4 teraflops.
 

LordOfChaos

Member
The NEO gpu will mostly perform like a 4.2 teraflop GCN 1.0 / 1.1 GPU with no memory bottleneck.

Which would have performed 40% better without the memory bottleneck...So...Ok, I guess I'll take that statement.

Given that the PS4s memory bandwidth is about the 7850s plus the 20GB/s CPU bandwidth, I don't think this is a terribly assumptive thing to say about the Neo vs the PS4. The 7850 saw benefit from the bandwidth removal in the next 200 series, the PS4 will see the benefit from bandwidth limit removal in the Neo with compression.
 

Proelite

Member
Which would have performed 40% better without the memory bottleneck...So...Ok, I guess I'll take that statement.


The 280x that's out on the market is not bottlenecked. It comes with a 288GB/s bus.
The test crippled it to 224 GB/S.

Not seeing the point of the tests besides find out which games are really sensitive to memory bottlenecks.
 

Madao

Member
I thought this was a GameCube thread.

Still have no clue why GCN is its abbreviation.

i think because the neo geo pocket color already had NGC as its abbreviation

probably an unofficial explanation but good enough for me

Wasn't it officially the GAMECUBE NINTENDO?

a great man once said:


blowing on that endo, gamecube nintendo.

Nintendo had to use GCN instead of NGC because National Geographic Channel owned "NGC" at the time.

in japan, they did use NGC for the Gamecube.
 

Cheerilee

Member
i think because the neo geo pocket color already had NGC as its abbreviation

probably an unofficial explanation but good enough for me

Neo Geo Pocket had NGP and Neo Geo Pocket Color had NGPC. They never had NGC (Neo Geo Color).

Nintendo had to use GCN instead of NGC because National Geographic Channel owned "NGC" at the time.

in japan, they did use NGC for the Gamecube.
There were a half dozen different companies using "NGC" at the time. It doesn't matter. NOA actually applied for a trademark on NGC and they were granted it, but then they abandoned it. The only reason people think it matters is because the World Wildlife Fund and the World Wrestling Federation got into a nasty fight over who was allowed to have wwf.com.

NCL wanted to get off using "Nintendo" as a brand, because failures like the Nintendo 64 were damaging it. They preferred something that stands on it's own like GameBoy or Wii, because those names are disposable, while Nintendo is forever.

NCL apparently told NOA to drop the "Nintendo" part of the name, but NOA thought that Nintendo was in trouble and it was no time to start getting conservative with their branding, so NOA wanted to include it. They came up with a "compromise" of NOA being allowed to use "Nintendo" in GameCube's name, but NCL ordered NOA to de-emphasize it, making it literally GameCube Nintendo.

And then NCL stopped giving a shit and started calling it a Nintendo GameCube (contrary to their idea of trying to avoid using the word Nintendo).
 

Dezeer

Member
Vega pretty much is a larger Polaris (GCN 4) with HBM2 I believe. The Scorpio version has GDDR5X instead, so basically a large Polaris die. Per clock per core then I'd expect similar to GCN 4 in the above article, so up to 40% more performance per paper flop than GCN 1.



True, the only 20% higher memory bandwidth threw me for the over double more (paper) GPU execution power, but the compression seems so significantly better it may be enough.

It's almost like they thought of this stuff and I should stfu!

Polaris, GCN 4, is GFX IP 8.0 / 8.X. Vega is rumored to be GFX IP 9.0 which would make it GCN 5.
 
Top Bottom