AMD Radeon R9 Fury X Reviews

[Rys]

@Rurouni said:

Didn't know that 3xx series included Fiji-GPUs …

Isn't Fiji only GCN 1.2, so Tonga level? If the current Fiji paired with GDDR5, it would probably perform worse than Fury X. I don't know how AMD could sell that. Maybe if the current 390X have Fury X core with current 390X price, it might do good sales wise, but is it possible? Do HBM enable AMD to pack a lot more CUs? Do that amout of CUs (Fury X) needs the bandwidth provided by HBM to be effective or not?
My point is that the Fiji core isn't exactly new but AMD needs something big to spark peoples interest. Fiji core with GDDR5 can't do that, at least not vs what Nvidia is offering.
What I actually want is for their 3xx lineup to at least use GCN 1.2 exlusively (they can keep the CU count), which sadly AMD didn't do. I don't think Fiji makes sense when paired with GDDR5 (again, it depends on whether those many CUs can perform well with the smaller bandwidth).

To sum it up, my question is, can you realistically make Fury X with GDDR5? If it's possible, would the amount of bandwidth bottleneck the CU utilization? How much decrease in performance are we going to get if Fury only have 384 GB/s bandwidth? Btw, can you underclock Fury memory to around 384GB/s and see how much performance decrease it would get? I know that the utilization efficiency is probably different between GDDR5 and HBM, but I'm curious.

 

[Rys]

@CarstenS said:

I interpreted Chalnoth's posting that he proposed a Fiji-style, massive GPU with conventional GDDR5-memory (which Fiji does not support according to an AMD rep) as a possibly better investment of engineering resources.

 

[Rys]

@Chalnoth said:

I interpreted Chalnoth's posting that he proposed a Fiji-style, massive GPU with conventional GDDR5-memory (which Fiji does not support according to an AMD rep) as a possibly better investment of engineering resources.

Sort of, yes. In any event it's clear that AMD has some work to do to catch up to nVidia in terms of performance per watt. I don't know how much they could have worked on that particular issue with this card and still made a similar release date. I don't know if they even have enough resources to catch up to nVidia any longer. But I suspect that doing HBM this year instead of next year may have been a strategic error on their part.

Not a huge one, mind. The Fury X isn't a bad card. It definitely has a clear advantage in terms of noise at load (provided you don't have a defective cooler).

Very curious to see what the next generation of HBM cards will do.

 

[Rys]

@fellix said:

Very curious to see what the next generation of HBM cards will do.

The more exciting part of that will be the new process node, not so much the HBM itself. The small form-factor PCB the HBM allows is still amusing, though.

 

[Rys]

@Chalnoth said:

The more exciting part of that will be the new process node, not so much the HBM itself. The small form-factor PCB the HBM allows is still amusing, though.

That will probably allow HBM cards to become significantly cheaper than current designs once the economies of scale ramps up.

 

[Rys]

@Rys said:

I don't think the PCB and non-chip components are the dominant cost of those designs, so I wouldn't expect prices to fall too far even at scale.

 

[Rys]

@Alessio1989 said:

I want to see two Fury X in a single adapter in linked node. 8GB should be enough then D:

 

[Rys]

@Ryan Smith said:

Didn't know that 3xx series included Fiji-GPUs …

I'm being a bit pedantic here, but by all reasonable measures you can count the Fury cards as part of the 300 series. It's just like the Titan and Titan Black were part of NV's 700 series, and the Titan X the 900 series; they're cards based on the same generation of chips as the other members of the family (or as close as you can get in AMD's case).

 

[Rys]

@Alexko said:

Being pedantic here, but by all reasonable measures you can count the Fury cards as part of the 300 series. It's just like the Titan and Titan Black were part of NV's 700 series, and the Titan X the 900 series; they're cards based on the same generation of chips as the other members of the family (or as close as you can get in AMD's case).

Speaking of which, I wish NVIDIA and AMD would stop with the Titan/Fury silliness. The seemingly random suffixes (Black, X, etc.) make it difficult to remember in which series a particular card belongs. And it's only going to get worse with future series, as vendors will have to get more creative with suffixes.

 

[Rys]

@CarstenS said:

I'm being a bit pedantic here, but by all reasonable measures you can count the Fury cards as part of the 300 series. It's just like the Titan and Titan Black were part of NV's 700 series, and the Titan X the 900 series; they're cards based on the same generation of chips as the other members of the family (or as close as you can get in AMD's case).

While I would agree from a purely shader-core POV that Fiji is not very much different (if at all) from Tonga, there's certain little things that set them apart in a curious way. And for that matter, the whole 200/300 series is quite a heterogeneous mix of things.

What sets them apart - and I would consider this more importantly than official CU-level is of course the completely revamped memory controller (do we have confirmation about memory controllers tied to RBEs/SIMD btw?) in Fiji and additionally, the Fiji card(s) have a separate installation path entry in the driver (ati2mtag_R7000) as opposed to seemingly similar Tonga (ati2mtag_Tonga) which has some differences - Powergating (of some kind) seems to be disabled in Tonga, but not in Fiji.

Also, Fiji has a more advanced UVD than Tonga (which we know) and Tonga apparently can drive Analog displays, which Fiji and Hawaii can't.

 

[Rys]

@3dilettante said:

I won't speculate as to what the impact on the memory standard might have been if AMD's engineering management had a somewhat different focus, but whatever the situation it seems really clear that they didn't make optimal use of the additional bandwidth. It might have been better, for example, if they had waited on implementing HBM until their next architecture and instead gone with a more traditional setup this time around.

Potentially, they may not have had as much freedom this late in the game.
Presentations on 2.5D integration pointed out the wide range of concerns, including figuring out what happens when a functional package involves multiple partners in manufacturing. It's a memory standard, manufacturing problem, and logistics puzzle that AMD can be lauded for hashing enough of those questions to get something to the market, in some form.

The downside is that it may not just be AMD's schedules that get messed up if they decide to hold off and skip HBM1, which does seem to fall a little short of what AMD could have used. But if your future manufacturing means not burning one or more of a raft of partners that have invested their own time and resources in getting something out in a promised time frame, and that product needs a high sale price to make up for the very high costs of initial production, where else can you go?

 

[Rys]

@fellix said:

Fiji could have been strapped with a conventional 512-bit GDDR5 interface, but that would have eaten too much into the tight TDP budget of the already stretched 28nm implementation, since AMD didn't bother to redesign the GCN for more efficiency. They simply had no choice, but to get early into the HBM hype, despite the obvious first-gen limitations.

 

[Rys]

@3dilettante said:

Fiji could have been strapped with a conventional 512-bit GDDR5 interface, but that would have eaten too much into the tight TDP budget of the already stretched 28nm implementation, since AMD didn't bother to redesign the GCN for more efficiency. They simply had no choice, but to get early into the HBM hype, despite the obvious first-gen limitations.

The Anandtech review seemed to extrapolate from some conversations with AMD that HBM saved 20-30W over Hawaii.
Assuming the other tweaks, like the water cooler, improved adaptive voltage, and binning carried over, that would be the power contribution.
What we don't have clear figures for is how much area HBM freed up, so it's possible that some of the CUs would have to be cut, although since we have 20-30W of power to make up, the loss could be somewhat mitigated.

The bandwidth deficit would still be there, but the Techreport's numbers show that while the non-standard 290X model's bandwidth is lower, the mature tech reaches notably better utilization, something like 75% versus Fury's 65%. The achieved bandwidth for the 290X is about 80% of what Fury can do.

10-15% lower power ceiling, possibly a loss of CU count due to area in that range, and 20% less bandwidth.

In terms of overclocking, Fiji's underwhelming overclocking so far might be due to its binning and adaptive voltage eating away at much of the margin overclockers count on, or not working well beyond standard specs.
The officially locked-down memory bus and teething pains associated with it might be partly why Fury is so twitchy at minor overclocks, however.

If the more mature tech could be more readily tweaked with the VRM setup Fury has, it might be able to edge up higher percentage overclocks than Fury's attempts. The non-X version of Fury apparently will drop 8 CUs and gain back 20 or so Watts due to air cooling, so whatever it achieves might be a good proxy for what could have been AMD's top-end product.
Without at least some of the raw bandwidth advantage, there would be situations where the GPU that could have been would probably be less consistent. Conversely, Fury has its own consistency issues, some of which a bigger RAM pool might have helped.

With the premium of HBM and the interposer, and the leeway in pricing AMD has, it might have been a "well, we have to use these interposers somewhere" situation.

 

[Rys]

@Razor1 said:

http://videocardz.com/57177/amd-radeon-r9-fury-launches-tomorrow

press deck leak still 275 watts TDP even with being cut down

 

[Rys]

@3dilettante said:

TDP is something that if you get it wrong you risk damaging or routinely hobbling the card, which is not what is in the slides.
TBP is something marketing can giggle over every time a tech site calls it TDP.

GPUs with far less silicon than Fiji can draw that much power, and DVFS has a lot of leeway, so a modest cut in CU count wouldn't change much.

 

[Rys]

@ToTTenTranz said:

IIRC, the R9 290 has the same announced TDP as the R9 290X, yet the reference cards had a small difference in measured power consumption.

 

[Rys]

@Razor1 said:

well the chip is going to be operating at 75 degrees C.

 

[Rys]

@fellix said:

Actually, thinking of power redistribution, HBM wasn't the first option for AMD. They already did a similar move with Hawaii and Tonga, by slimming down the memory interfacing for lower speeds and power loads, reinvesting the saving back to the rest of the core.

 

[Rys]

@sheepdogexpress said:

In the slide deck, fiji isn't even beating the regular gtx 980 by that much, particularly if AMD official Fury x benchmarks are anything to go by.

At that price range, AMD is going to be going against overclocked gtx 980s. With these cards being typically 10-15% faster than a regular, it just seems like AMD has really fallen behind on performance per mm2 which means margins are much worse than nvidias.

I can't imagine AMD and Fury series in general really breaking even considering the R and D and the supply limitations.

 

[Rys]

@Alexko said:

IIRC, the R9 290 has the same announced TDP as the R9 290X, yet the reference cards had a small difference in measured power consumption.

TBP, but yes.