Why Tonga and its successor fail?

Full unlocked Tonga(Antigua XT) released couple days ago and it suprasses GTX 960(actually R9 380 also defeats it in most situations) but there's no huge difference within two GPUs. This is not bad users, because the card has good value but there's a very sad story for AMD. With 5000 billion transistors and 359 mm^2 die size, we are entitled to expect more performance. (GTX 960 has 227 mm^2 die size)

How can we analyze this failure? What's the reason/reasons of this?

My assumptions:

*Number of ROPs is the weakest point. There's 1/64 ROP/shader ratio and it seems unbalanced. In full unlocked Maxwell 2.0 cards, it's 1/32. And according to official sources new cards use bandwith power more efficiently. In standard INT8 format, fillrate is bound to ROPs. It seems like more ROPs could improve the performance.

*Maybe color compression doesn't work as it should be. NVIDIA's solution looks like more effective in real world.

*Optimization is a problem but i am happy with AMD's drivers. They don't mess up the whole thing for performance gain in popular games. (like NVIDIA do)Their drivers has less fatal problems.

*There's something wrong with memory access patterns.

*Their lack of FP16 textures are big problem on games. Perhaps nowadays it becomes a bigger issue.

What do you think about this? What's wrong with this GPU?
 
It's been years since there was any useful data presented anywhere to understand what's really going on. I'd go so far as to say it's a waste of time even thinking about this.

e.g. there is no performance at different resolutions for shadowing on/off in a single game across multiple cards.

etc.

One could also argue that games are far too complex now to isolate performance for each component of the rendered frame. And there can easily be 10s of percent variation in performance with new games solely due to drivers. Long after the window of opportunity to make comparisons has past. So no one revisits.
 
I'm not sure a die size comparison is relevant since a fair bit of Tonga apparently remains disabled, at least 128 bits of its memory bus and perhaps 16 additional ROPs. Beyond this, I'd say that Maxwell can reasonably be considered a generation ahead of GCN as it currently exists.
 
Full unlocked Tonga(Antigua XT) released couple days ago and it suprasses GTX 960(actually R9 380 also defeats it in most situations) but there's no huge difference within two GPUs.
Yes, because both R9 380 and R9 380X have exactly the same core clock and exactly the same memory clock. The only difference is in the number of active CUs. In fact every preceding "X" model was clocked higher than the "non-X". The diference was 50 MHz for R9 390(X) and Fury-(X), 67 MHz for R9 280(X), 125 MHz for HD 7970/7950 and 140 MHz for HD 7870/7850. But it's zero for R9 380(X).
 
Hang on, if this card is faster than a nvidia card of the same price how is it a failure ?
I might have read it wrong, but it sounded like OP was comparing die size (GTX960 ~ Tonga perf & price bracket), so there's a lower implied margin. Or something.

Also comparison of architecture/density/trannie efficiency.
¯\_(ツ)_/¯
 
Tonga apparently remains disabled, at least 128 bits of its memory bus and perhaps 16 additional ROPs. Beyond this, I'd say that Maxwell can

I'm very curious about it. One year ago they showed a possible Tonga block diagram which was consisted of 384 Bit and 48 ROPs. Was it only a speculation? Or is this card not fully unlocked. 384 Bit and 48 ROPs could change everything.

But if this is true, there's another failure story. It has a great potential but AMD can't use it somehow. How much is the yield decrease if they produce full fat hypothetical GPU like this. There must be a very big problem.

As far as i see 1/8 extra shaders don't benefit too much. If it's fillrate or geometry limited, extra ROPs and BW could solve the one of weak spot.
 
Why would anyone think 48ROPs? Consider the precedence with CGN.
Because…
A generic crossbar hasn't been there since prior to R600 (which had the ring bus, but was dropped subsequently). Tahiti was an exception
If only Tahiti was the exception, than Tonga / Antigua should have 48 ROPs. Are you implying, that Tahiti wasn't the only exception, or that Tonga's additional two 64bit interfaces aren't connected to its 32 ROPs?
 
Well it definitely seems that he is implying that it does not have 48 ROPs at least. :)
 
Beyond this, I'd say that Maxwell can reasonably be considered a generation ahead of GCN as it currently exists.

I was just wondering if anybody has any idea as to why a Maxwell based card with 1024 processing elements can keep up with an GCN card with 2048 processing elements? At one point I was going to start a new thread in the architecture forum but put in here instead. I was reminded about this with the 280x reviews, and would really appreciate some insight. I mean I just can't imagine any feasible explanation as to how something with half the processing elements and half the width memory bus can be in the same ballpark performance wise. It's been bugging me for a while and was wondering if anyone can offer up an explanation beyond the abstract contention of architectural efficiency?
 
What is the normalized specs taking into account the speed of the elements? Does one run at significantly higher speed or have significantly lower latency on completing instructions?
 
Does not Maxwell sacrifice double performance a lot?

They have completely remove them.. and in addition release sku who was ( initially ) only intended for goes on game retail channel.


What is the normalized specs taking into account the speed of the elements? Does one run at significantly higher speed or have significantly lower latency on completing instructions?

In addition, what is the games, software tested, resolution, settings ? .. If you take a game like fallout4, Assassin Creed etc, who out of the box offer 30 too 50% performance gain on the same Nvidia gpus .. well you see what i mean... ( furyX is equal to a 980 in Fallout4, when in other games the old 390x ( 290x Hawaii ) match it.

The problem with Tonga, is it is just too late on the market... Itt was allready the case witth the 285, released way too late ( after Hawaii ), and it is the same problem with this TongaXT ...

Tonga have never been anything more of the place it should take on the market lineup.. but have allways fall in delay with the market.

If this gpu will have been released during the Kepler time, it will have been excellent.. now.

I still use my good old 2x 7970 ... basically a FuryX ( 8Tflops SP / 2TFlops DP ).. even with Intensives render with cycles ( Blender + openCL ), Luxrender ( Blender + opencl ), Vray ( 3Dmax + OpenCL), i still dont see the need to upgrade to FuryX ( i will need 2, or 3, for increase the performance of my system ).

Now that i look on Corona render, this could change ( AMD FireRay OpenCL raytracing engine ) .. As the performance should be at least equal or better of Luxrender with 3Dmax and Maya.
 
Last edited:
It feels like a slightly better Radeon 7970, and it's priced like a 290. I think that sums up why the 380X is boring or disappointing.
Now if you're on linux and bet on AMD, a 380 or 380X is the card to get because GCN 1.2 will get better driver support. (eventually)

What's missing is the Iceland GPU (low end) : I guess it was done ages ago but there is a backlog of Oland chips first.
What's also missing is a decent CPU from AMD to pair with Tonga, it's at least one year away (Zen).

If it's just for gaming, I guess a 380X and a 1080p 144Hz monitor make for something decent at least but a current FX or Athlon 860K doesn't cut it on some of the games.
 
Iceland has been shipping for ages. It has no displays so it won't end up in a desktop add-in card...
 
Lots of notebooks ship with a discrete GPU, but they do it in a with switchable graphics - i.e. when looking at the desktop, Word, Excel, etc., it will render on the integrated graphics to save power; however when it is on AC or a high intensity 3D app is used it can switch to the discrete GPU. In these notebooks the displays are routed via the integrated graphics, rendering any display pipes on the discrete GPU inactive. Display-less GPU's specifically for the switchable graphics notebook market have been designed for several generations.
 
Back
Top