AMD: Pirate Islands (R* 3** series) Speculation/Rumor Thread

[Razor1]

somethings are really weird about these AMD benchmarks, seems like AF not being used and shadow quality is turned down, both of these effect bandwidth and fillrates

Clueless IMO.
AMD built a card to OC offering value at entusiast level silent and at same price as a reference as 980ti has which is HOT, LOUD and frankly a pity.
Why do you think that suddenly non reference OFFERS HYBRID watercooling to 980ti?
Its due to its HOT and needs more cooling to OC.

Overclocking community is a small part of the enthusiast crowd, many enthusiasts like myself buy cards that as fast as possible but don't overclock them (I used to, but didn't see the need when cards die faster for that extra bit of performance when I'm upgrading frequently, a card ever year or year and half). The crazy overclocking community that go for the 20% and up overclocks are not trying to get their games to run smoother they are doing it for benchmarking (most of them). And they will be using custom coolers if the coolers on the cards aren't good enough. The AMD AIO seems to be pretty solid.

AMD simply is so much better its beyond such posts. (note what I did here)

Not so well said, the automatic feeling of when you see a cooler like the AIO most people are like damn that's some serious cooling, does it perform like it needs that cooling? Epecially with enthusiasts they probably know their cards very well, unlike the general consumers, Its a win for the consumer if they need it, its not a win for AMD when the consumer doesn't need it, its extra cost that goes unused.

Superb deal for me can slam it into the computer having the fastest single core card in the world and smile.

personal prerogative nothing about that

 

[Razor1]

ComputerBase reviewed the R9 390X, it's 26 % (!) faster than R9 290X, but they tested only 1080p:
http://www.computerbase.de/2015-06/amd-radeon-r9-390x-390-380-r7-370-test/3/#abschnitt_leistung
Any 2560×… benchmarks?

There is a few out there, it stretches its legs a quite a bit more at 4k because of the extra memory. Or actually the r290x gets cut actually.

 

[gongo]

ComputerBase reviewed the R9 390X, it's 26 % (!) faster than R9 290X, but they tested only 1080p:
http://www.computerbase.de/2015-06/amd-radeon-r9-390x-390-380-r7-370-test/3/#abschnitt_leistung
Any 2560×… benchmarks?

How this possible?
Was 290X bandwidth limited??

Driver optimisation?
Will AMD release such tweaks to 290 users?

290X was over throttling, even for custom cards?

 

[mczak]

There is a few out there, it stretches its legs a quite a bit more at 4k because of the extra memory. Or actually the r290x gets cut actually.

AMD just cheated with the review driver, that's all. The new driver knows some new optimization tricks (I hope they are legit...), but won't install on old cards.

 

[gongo]

Any proof of their cheating ways??

 

[spworley]

AMD just cheated with the review driver, that's all. The new driver knows some new optimization tricks (I hope they are legit...), but won't install on old cards.

Tech reviewers: please flash the 380x BIOS onto a 280x card for a comparison! The benches would be extremely interesting.

 

[3dilettante]

Except, power draw isn't directly comparable to benchmark numbers, so I don't really see your point.

It's still a power-limited device at stock, perhaps more so since bandwidth has been significantly improved.
Seeing some of the benchmark numbers, I think it works for the best that they didn't lop off one or two speed bins, unless you think the graphs wouldn't be perturbed by a 10-15% drop in clock. AMD has gone ahead and provided data indicating that there is a measurable effect, going in the reverse is enough perhaps to reverse some of the slim wins or ties.

It seems like a worthwhile investment for the benchmark headline product to be given enough of a leg up to get those wins.

 

[mczak]

Any proof of their cheating ways??

Well just read the article. It is mentioned the new driver wouldn't work on the old cards, thus the older driver was used to benchmark the old cards. Maybe you don't want to call that cheating, but nevertheless this creates an artificial performance difference between new and old cards. The good news though for owners of old cards is that you shouldn't have to flash the bios to a new card, just wait for the next official driver to get the benefits (unless the optimizations were too fishy to get included in the driver, but let's hope not...).

 

[eastmen]

Anandtech in the preview mentioned the ram being clocked 20% faster. So its not out of the question that its faster esp if u can maintain higher core clocks

 

[3dilettante]

I'm curious about the air-cooled Fury product and what has to be trimmed for that SKU, and if it will go for the 95C limit like the stock cards of the prior high-end.

Bandwidth measurements after warm-up might indicate if the thermal refresh settings kick in with a shared thermal interface to something that doesn't maintain 50C temps.

 

[mczak]

Anandtech in the preview mentioned the ram being clocked 20% faster. So its not out of the question that its faster esp if u can maintain higher core clocks

Oh yes, I'm certainly not denying that there will be performance differences due to higher clocks. It is though impossible to tell how much of the improvement is due to that and how much due to the driver differences. In particular the 380 definitely offers more of a performance increase than is possible due to clocks, the 390 series is more difficult to judge due to the large memory frequency increase.
Note that this strategy of review drivers only working on new cards isn't really anything new, it is quite common to include some optimizations for new cards only at first when launchning new cards. It is just slightly more odd due to the difference between old and new being really completely artificial at a driver level.

 

[Lightman]

Anandtech confirms Fury has 1/16th FP64 compared to 1/8th in R9 290X/390X.

This time it might be a hardware limitation and not just pure software cap.

 

[gamervivek]

computerbase are using 290 cards that throttle massively(800 something), no wonder they are getting those results.

 

[nnunn]

Still only 1MB L2?

From http://www.pcper.com/reviews/Graphics-Cards/AMD-Exposes-Fiji-World-HBM-Enthusiast/Fiji-GPU

"The chip includes 2 MB of L2 cache, which is double of the previous Hawaii. This goes back to the memory subsystem and 4 GB of memory. A larger L2 cache is extremely important for consistently accessed data for the compute units. It also helps tremendously in GPGPU applications.

local/shared ?

 

[AlNom]

From http://www.pcper.com/reviews/Graphics-Cards/AMD-Exposes-Fiji-World-HBM-Enthusiast/Fiji-GPU

Thanks.

The HBM memory controller apparently takes up less space than a traditional 320 bit GDDR5 unit,

Typo or is that how AMD compared it?

 

[3dilettante]

From http://www.pcper.com/reviews/Graphics-Cards/AMD-Exposes-Fiji-World-HBM-Enthusiast/Fiji-GPU



local/shared ?

It's consistent with the usual arrangement. Hawaii has 16 32-bit channels, Fiji has 32 128-bit channels.
It doesn't seem like it has departed from the norm, so that would be a shared cache with cache slices linked to their channel.

I'm not sure why Anandtech doesn't have typical board power numbers for the 290 series.
http://www.legitreviews.com/amd-radeon-r9-295x2-8gb-video-card-review-at-4k-ultra-hd_138950

The new cards use slightly more.

 

[Lightman]

Another performance metric for Fury X, this time it's DX12 draw call rate:
https://twitter.com/killyourfm/status/611579128941129728?s=09

 

[3dilettante]

It's curious how the showpiece Oxide engine tests showed a reversed position. Possibly whatever that engine was doing hit a sweet spot for Nvidia's command processing method, where the benchmark does not.

The numbers for both seem to be a bit on the low side, but mostly consistent with some of the 4-core tests done earlier.

This seems suggestive of how different Fury's command front end is or isn't from the older chips.

 

[Kaotik]

It's curious how the showpiece Oxide engine tests showed a reversed position. Possibly whatever that engine was doing hit a sweet spot for Nvidia's command processing method, where the benchmark does not.

Wait what? There's Oxide benches somewhere on Fiji?

 

[SimBy]

http://www.anandtech.com/show/8962/the-directx-12-performance-preview-amd-nvidia-star-swarm/3

I'm guessing he is talking about this? That was a long time ago though.