AMD: Speculation, Rumors, and Discussion (Archive)

[ninelven]

feigning disappointment

Please don't pretend to know MY emotional state. --respect

A dead AMD is not anything we should be interested in.

Given as I have family with stock in AMD, I am not. I am still clinging to a Phenom II FFS. Have to say this discourse has also been pretty disappointing....

As for the card itself, well it just went from an instant buy for me to let's give it a couple of weeks, see if there are some driver issues, what NV has on the horizon, etc...

AFAICS, the performance of the card is OK (a little inconsistent but whatever). I do also care quite a bit about power consumption (mainly due to heat and noise which I abhor). I do like the 8GB of memory, and the price is very good. But the fact remains that, instead of putting in my order for one tomorrow as I had planned, I will now be thinking the matter over a bit more...

 

[Orion]

AMD said 150W typical board power at Computex. What hype do you believe you saw from them, honestly? The missing piece was performance, which we didn't know about in full until today. And it's more than just on par with their 28nm. Yes it's not as good as Pascal, but that's not what was promised.

With the exception of the power issue, I agree it offers pretty decent performance. Depending on the title it can be above the 980 some sites suggesting significantly above in some titles at some settings(strangely enough that significantly above doesn't include ashes.).

e.g.
http://segmentnext.com/2015/06/09/nvidia-gtx-980-ti-vs-gtx-980-is-it-a-worthy-upgrade/
980 ti 24% faster than 980 at 1440p ultra in hitman.

http://www.hardwarecanucks.com/foru...9-radeon-rx480-8gb-performance-review-20.html
480 according to some sites can be 30~% faster than a 980 on hitman at highest settings.

Of course as these are different sites, these might not be directly comparable. But it suggests that at least in some apps and settings the performance might near 980ti

 

[3dilettante]

That's fine, but they basically left anyone who is middle to high end and above waiting. Combine that with the 2.8x figure that wound up flat out being not true in practice across the entire GPU(Probably is true in individual transistors)....

It's an up-to figure and I think AMD is picking specific clock bands where their features are more effective and they don't hit an architectural and process brick wall wall. Then there might be an adjustment factor for possibly missing their goals by 1-2 speed grades.

 

[Benetanegia]

I'm not being obtuse. Show me the slide that's driven your disappointment so we can figure out where AMD went wrong. They'll read this thread, maybe you'll help them.

Edit: Posted this to show the actual slide, which I think no one did, or maybe I'm blind. And to show that AMD claimed that they would improve efficiency with architecture improvements way beyond what 14FF could achieve.

 

[Rys]

Please don't pretend to know MY emotional state. --respect

Given as I have family with stock in AMD, I am not. I am still clinging to a Phenom II FFS. Have to say this discourse has also been pretty disappointing....

Sorry, I thought it was clear I was talking about the general state of the reaction to the product, not yours in particular. I mean for no argument or ill will.

 

[Triskaine]

Yes, I agree that they're now further behind. They have been behind for years. The fallacy is expecting them to catch up, and then feigning disappointment when it inevitably doesn't happen. Everything we discuss around here and have done for 15 odd years should have led us to something better.

The question is whether RX 480 is a bad product unworthy of the asking price. That you can even buy it at that asking price is worth a small cheer, given the FE bullshit. I contend the answer is no, and that it's worth the money at the perf/W and perf/$, and that AMD might even make some money. That's the thing that'll help them stop slipping further behind. A dead AMD is not anything we should be interested in.

As of now Nvidia can set prices to whatever they want, like Intel, and their products still get bought. What's is the point of AMD existing at all if the only benchmark where they are even slightly competitive is price? The circumstances of the last 10 years have put AMD in a hole which they can't leave under their own power. Let them sink, I say.

 

[3dilettante]

I am really happy about the architectural changes AMD did. It seems that console devs have given them lots of feedback about the choke points.

I'm curious if it's specifically a console customer that guided some of these changes.
The prioritization and CU reservation in particular addresses concerns brought up in the context of the PS4's development..

Native int16/fp16 support is also great in reducing the register pressure. Most of my integer math could be ported directly to int16.

GCN3 at an ISA level already exposed the ability to work at this granularity. Perhaps this is something different in terms of specific instructions rather than setting mode flags and fiddling with operand bits?

 

[Rys]

As of now Nvidia can set prices to whatever they want, like Intel, and their products still get bought. What's is the point of AMD existing at all if the only benchmark where they are even slightly competitive is price? The circumstances of the last 10 years have put AMD in a hole which they can't leave under their own power. Let them sink, I say.

You've answered your own question. If there's no competition and they can set whatever prices they want, expect prices to go up.

 

[swaaye]

What's up with the silly thread merges? This thread is crazy land.

Anyway, this launch reminds me of when Barts launched with the 6870 name and everyone was like "oh man it's barely faster than the 5870". But they were launching the bigger chip in a few months (6970). Maybe Vega will be more interesting and stir up the high end a little.

It does seem likely that the manufacturing process isn't as refined as it could be though. Shocker.

 

[Esrever]

The card is fine and dandy but the architecture is disappointing this being a new node and amd claiming getting more from then architecture than just a node jump. Currently it seems it barely even seems like a node jump when compared to Tonga and especially Fiji. Now we can kind of guess why Vega is so far up the graph in perf per watt from amd's roadmap.

Just hope amd can catch up sooner or later because nvidia hasn't been shy with the price gouging.

 

[Triskaine]

You've answered your own question. If there's no competition and they can set whatever prices they want, expect prices to go up.

I'm saying that AMD's continuing abstinence from competitiveness has already been priced in to a large extent by Nvidia and Intel. AMD could stop existing tomorrow and prices wouldn't change that much.

 

[Dave Baumann]

FYI - 110W is the ASIC power limit for this configuration of board. This is the BIOS value that PowerPlay uses to manage the ASIC power to.

 

[silent_guy]

The fallacy is expecting them to catch up, and then feigning disappointment when it inevitably doesn't happen.

It would be a falacy had they not built their whole initial launch around it. The december thing about Polaris was about perf/W and perf/W only. You can't fault us for that.

The question is whether RX 480 is a bad product unworthy of the asking price.

That's a good question for a which-GPU-should-I-buy thread. And the answer is: yes, if you don't have a GPU in this class, it is an excellent buy. Just like most other AMD GPUs. But it's of low relevance in an architecture thread.

Depending on the price and performance of the 1060, it will sell good to great, so there's that.

 

[sir doris]

Having taken a look over some other reviews (previously only looked at Anantech), it doesn't seem too bad as far as performance goes. There are some nice results in there as I assume some of the lower benchmarks can only go up with driver maturity?

I guess different websites bench different parts of games, because Anandtech has the GTX970 beating the RX480 in RotTR @ 1080p, while Tech Powerup have the roles reversed but quite a margin.

It's a tad noisy, but blowers always are. Any idea when the custom cooled boards will arrive?

 

[I.S.T.]

It's an up-to figure and I think AMD is picking specific clock bands where their features are more effective and they don't hit an architectural and process brick wall wall. Then there might be an adjustment factor for possibly missing their goals by 1-2 speed grades.

That... actually makes a lot of sense.

Consider that part of my post retracted.

 

[Razor1]

Edit: Posted this to show the actual slide, which I think no one did, or maybe I'm blind. And to show that AMD claimed that they would improve efficiency with architecture improvements way beyond what 14FF could achieve.

*Snip*

That is just marketing hype, this is why you never believe in marketing. Take what AMD said in a presentation and drop it down a notch for what you end up with.

 

[Rys]

I'm saying that AMD's continuing abstinence from competitiveness has already been priced in to a large extent by Nvidia and Intel. AMD could stop existing tomorrow and prices wouldn't change that much.

Good luck convincing Nvidia shareholders to leave money on the table.

 

[Rys]

I'm sorry about the thread merging, @BRiT should never have bowed to the pressure to do that. I'll try and unmerge it later so the individual aspects can be discussed properly.

 

[Erinyes]

Regarding power draw (and speculations about process tech, I guess) the Computerbase undervolting experiment may be interesting to some.
I get the feeling they placed themselves (too) high on the frequency/power curve, in order to be more competitive on the desktop. It would be interesting to see where it would end up as a product for mobile use.

That is definitely an interesting experiment. The perf/watt should be a bit higher for the mobile parts which will be clocked lower. But I dont expect miracles tbh. Also remember that NV can also do the same and clock lower..and still be at far higher clocks. Another trick AMD seem to have missed is a lower end part to compete with GP108. Polaris 11 and GP107 seem to be natural competitors but a lot of NV's GPU sales are actually from the segment below..i.e. x108 class. Now AMD does have APU's which compete in this segment..but for Intel CPUs..they might have to use a cut down Polaris 11 to compete against GP108.

I'm also interested in seeing how RX470 performs. Its possible that Polaris 10 is ROP limited in some cases so for a RX470 with presumably the same 256 bit memory interface and say 30-32 CU's..the performance could be rather close to RX480.

 

[Love_In_Rio]

Today's AMD can't perform like that.

Then isnt strange people in forums talk about their products like talked about Matrox products in mid 2000s.