The AMD Execution Thread [2007 - 2017]

[jimbo75]

There are clues all over the place, read the recent Anandtech article on stuttering and you can see how AMD is guilty of leaving performance on the table. It was the first iteration of GCN as well, and drivers have already given a large boost. AMD were probably also paying a driver penalty for being first to market for so long as well. Don't get me wrong though - the 7970 should never have existed with the clocks it came out at. They just didn't try hard enough or got arrogant due to Nvidia's previous ineptitude.

Imagine AMD actually fixes all these self-inflicted disadvantages one day, say with the 8000 series. They'll have their vastly better dev relations on top now. If it ever all clicks for AMD, Nvidia will be on their knees. They always seem to do one thing right and another two things wrong though.

 

[Raqia]

The cards have been out for more than a year! And even before release the drivers would have been in development for who knows how long. Granted, the first 6 months were still a mess. How that happened is a curiosity. They were breaking VLIW architectures in the drivers at the time too.

I imagine they rewrote the GCN driver before console dev. kits shipped out, or will do so before retail units ship as this will be an important milestone for the architecture's install base; they'll also be getting much more dev. feedback once console development kicks into gear. Rewriting a memory manager to be faster than before but also stable is extremely delicate work and a year or more is a reasonable amount of time for all the necessary validating and reiterating.

ATI has always been more aggressive than nVidia with process nodes and hardware; it seems like they try to beat nVidia to the punch with hardware that's fabbed at a lower node or a generation ahead in architecture and catch up with software only later. Rory Read has stated that they will be postponing node transitions in the future, and I imagine that he'll spend the R&D money on driver development instead to keep competitive.

 

[thatdude90210]

AMD.
If everything had been correct with their lineup and especially 7970 performance, then GTX Titan and GTX "680" wouldn't have existed in that form. And now because of it, NVIDIA thanks G O D and sells a 250$ card for 440$...

Forget what the 7970 SHOULD be, but just what it IS at the moment. It's roughly the same performance of the GTX680, plus two great games, yet the 680 is still around $70 to 100 more (just glancing at newegg). The 680 is obviously price there because buyers are still buying it. If people are willing to continue to pay that price, that's the consumer's fault because there is something out there with at least equal (if not greater) value for $70-100 less.

 

[Wynix]

Yeah, I remember that nVidia had said something about AMD's stack being less than they had expected. It's clear the BOM for the 7970 was higher than that of the 680 with a more complex PCB for the wider memory bus, more GDDR5, a larger die, and more expensive cooling solutions. The compute performance of the 680 also screams mid range card.

I see this alot but i just don't buy it, keplernomics being the main reason, additionally the GTX 680 was in short supply at launch(this would not have been a problem if it was truely a mid range card), lastly the Titan is way too compute heavy compared the the GTX 600 series, the only reason that explains the disparity is that Nvidia had planned this as the Workstation GPU from the start.

 

[UniversalTruth]

GTX 680 was in short supply at launch

Or because TSMC had problems with their manufacturing capacity. Maybe...

this would not have been a problem if it was truely a mid range card

Really weird logic.

 

[swaaye]

Look at it this way:

RV770 = 256 mm²
GK104 = 294 mm²
GF114 = 360 mm²
Tahiti = 365 mm²
R600 = 420 mm²
GK110 = 551 mm²
GT200 = 576 mm²

GK104, NVIDIA's "small die strategy". Except their edition of ATI's strategy doesn't include cut throat pricing because GK104 is a spitfire that can hang with AMD's finest.

 

[Kaotik]

Or because TSMC had problems with their manufacturing capacity. Maybe...

Yes, clearly, especially considering how the bigger chips from AMD were in short supply too - oh wait - they weren't.

 

[UniversalTruth]

The bigger chip from AMD is always in much lower demand, so it's possible even with tight supply, there would be some availability.

 

[jimbo75]

That's just not true. Up until the 3rd quarter of last year they had pretty comparable market share. It's rarely worse that 60-40 either way.

edit - sorry my bad, didn't notice the bigger part. In that case yes you're probably right that Nvidia has quite a bit higher demand for the pricier chips. However I recall AMD saying that supply at TSMC wasn't too bad but they were monitoring it closely in case it became a problem. That was in Q1-Q2 last year, my guess was they were selling most of what they were making.

 

[jimbo75]

This might help them a bit, at least in terms of perception - http://hexus.net/tech/news/cpu/54057-rumour-amd-unleash-centurion-fx-5ghz/

A 5GHz octo-core must be a threatening CPU on raw performance at least. Not sure I want to know the TDP on it.

 

[eastmen]

This might help them a bit, at least in terms of perception - http://hexus.net/tech/news/cpu/54057-rumour-amd-unleash-centurion-fx-5ghz/

A 5GHz octo-core must be a threatening CPU on raw performance at least. Not sure I want to know the TDP on it.

it wont be fast enough.


They need to focus on improving the core or they are as good as done in the traditional cpu market.

No one is going to want to spend $800 bucks on a amd cpu which will use a ton of energy when they can simply buy a $300 intel part and get better performance with less energy / heat production

 

[swaaye]

Maybe it will come with a phase change cooling setup like the Kryotech K6 systems of the olden times.

 

[3dilettante]

The promise is 5 GHz on air, although it's not clear if that is turbo. The rumor is ambiguous on whether that speed is an out-of-the-box speed or a guaranteed OC.

Since CPU core speed is only one part of the picture, and things like the memory controller, uncore, and thermal policy probably won't not scale as well, the benchmarks would probably scale less than the optimistic +25% best-case.
Even with the best case, AMD's perf/GHz has lagged so much in so many reviews that even increasing it by that much doesn't make it a winner in a lot of tests. Notably less embarrassing versus (stock, and cheaper (edit: and cooler, more performance consistent)) non-Haswell i7s, sure.

 

[jimbo75]

It's about 15% faster at 5GHz compared to 4/4.2 Turbo, including some single-threaded benchmarks. If it can make 20% in multithread, and assuming 5GHz is all cores and not the best turbo it might be ok.

$795 is clearly a joke but I guess they might feel that it's a collector's item - first 5GHz stock cpu with 8 cores definitely sounds good. Anyway, with luck this will be the last one of the Bulldozer fiasco cpu's and Steamroller will be the beginning of something more like what is required.

Kinda quiet on that front though AMD claims it's still on track for this year...

 

[Exophase]

5GHz for all cores on air seems woefully unrealistic. There are already motherboards that can't properly handle the 4GHz stock speed (for FX-8350) under heavy load without throttling...

 

[itsmydamnation]

5GHz for all cores on air seems woefully unrealistic. There are already motherboards that can't properly handle the 4GHz stock speed (for FX-8350) under heavy load without throttling...

and I can undervolt @ 4.6ghz on vishera with the stock cooler. I appear to have a really good chip but it shows that 5.0ghz on highend air will be possible.


Also piledriver performance basically scales with core clock, memory above 1666 provides almost no benefit and increasing the unncore/L3 clock provides almost no benefit.

If AMD can sell it why not, good on them.

 

[sheepdogexpress]

Why? What will prevent AMD from matching Maxwell?

I think the primary thing holding back AMD will be the share holders.

AMD is in a transition period right now and it has to pick which type of company it want to become. It can't afford to spend money on R and D on all the markets it is currently involved in and something is going to take a cut. Rory Read has already said he plans on making AMD a profitable company by reducing costs while making the type of revenue they currently produce.

With AMD getting into this new server/arm thing and HSA, its either going to be the CPU division that going to be taking a cut or the GPU division. And I think its going to be the GPU division.

Both Nvidia and AMD struggle to make money in this division. AMD has barely been profitable compared to how much it spent into aquiring ATI. In the long run, the deal has costed them more than it has made them.

The desktop graphic market is shrinking so much, that it only leaves room for one player and AMD is primarily a CPU company.

Nvidia doesn't have a choice and has to spend most of its R and D on graphic development. They stop advancing graphics and they are done. They can't be Intel, because they don't have that type of lead in branding or performance.

As a result, Nvidia's midrange catching up to AMD top end should not be a surprise this generation. Considering how much more Nvidia spends, it was likely to happen.

The thing that going to work in Nvidia's favor next generation is it seems they seemed to have bisected the consumer graphic division and the professional division.

If they can continue to make 300mm2 parts that primary purpose is gaming and make 500mm2 parts that primary purpose is the professional market, AMD is going to have a tough time competing with a 360mm2 part that does both.

For a flagship, 360mm2 squared is too small to be a successful card at both when your facing a 550mm2 monster on one end, whose primary purpose is computing tasks and a 300mm2 card at the other end whose entire die is dedicated to gaming.

If AMD didn't pack in any games this generation, they would have been toast as far as marketshare goes.

 

[Kaotik]

Both Nvidia and AMD struggle to make money in this division. AMD has barely been profitable compared to how much it spent into aquiring ATI. In the long run, the deal has costed them more than it has made them.

At least it has been profitable part of the time, when exactly was the CPU division profitable last time?

 

[jimbo75]

They are both in a similar boat. ATI has cost AMD $billions but without the purchase both would be heading out the window by now. Tegra is well on it's way to costing Nvidia $billions as well. Neither company had any real alternative except the merger which was thrown out due to JHH's demands. That company would probably have been competing with Qualcomm by now.

Both companies need to do better than what they are doing, but it's not in graphics it's in CPU/SoC. The assumption should be that Nvidia will concentrate more on perf/watt which could mean lesser performance improvements than we're used to seeing at the high end. Historically Nvidia's gpu's get more uncompetitive compared to AMD's as they come down the price points, die sizes and TDP's - they just don't scale down as well.

 

[UniversalTruth]

ATI has cost AMD $billions but without the purchase both would be heading out the window by now

Truth might be a little bit different and actually ATi could be the major reason why they are in so poor condition now.
I suspect ATi could have been doing much better as an independent corporation. AMD killed them, both of them.