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

A 384 bit 32 CU Tonga could have done significantly more damage than the current 285.

Why did AMD not release it right away?
Well that certainly puts tonga in a better light, but why no full tonga? :(

I think the cryptocurrency craze in late 2013/early 2014 created an "artificial" demand for Tahiti chips. And when the dedicated FPGAs appeared, the cards flooded the second-hand market and the demand curves for retailers fell abruptly.
Final result is AMD and the retailers ended up with truckloads of excess Tahiti cards, occupying the same price range as full Tonga cards.
The 285 appeared probably because it could do 7950 performance with a 256bit bus (color compression),therefore being substantially cheaper cards to produce.
 
Do we know for sure that the die shot presented above is Tonga and not just Tahiti?
TahitiDieShot.jpg


Tahiti.

Update: Oh, ninja'd ... just delete this post :)
 
That rumor doesn't seem believable. No one will buy them at those prices.
Depends on performance of these re-fried cards. Anyway, seeing this rumor is depressing; back in the Geforce 2 days, high-end GPUs cost like 250 dollars - which granted was more money then than it is now, but not 100 friggin percent more...! 500+ for a top of the line board is really rough. At 250, many could justify yearly upgrades if they wanted to, but today you have to have quite a bit more disposable income for that, especially if you want something like a titan.

I suppose this is the way of the future. *sigh*
 
Looks like Tonga has been geared up all the time, as a "performance" SKU in the upcoming R9 300 Series, replacing Tahiti.
 
Do we know for sure that the die shot presented above is Tonga and not just Tahiti?
Well, in the press deck PDF you can see that it does a) look different and b) is labelled Tonga.

http://www.pcgameshardware.de/AMD-R...nga-GPU-256-Bit-Interface-und-32-CUs-1160632/
(first picture right at the top with the alt-text shown)

You must have had a special press deck, not included in ours either
Apparently there were two: One overview and one "architecture" - the last one probably only given to techday attendees?

I've included the architecture-labelled one here:
http://www.pcgameshardware.de/CPU-H...-HSA-drei-Modelle-von-15-bis-35-Watt-1160641/

AMD mkt drone: "boss, I accidentally sent out the deck with Tonga screenshot to this one journalist."
Boss: "let's hope he doesn't notice, please let's hope he doesn't notice."
"Damn..."
Nope, it was sent via PR agency to probably all Carrizo techday attendees. I didn't know until today that other press only got the overview.
 
Depends on performance of these re-fried cards. Anyway, seeing this rumor is depressing; back in the Geforce 2 days, high-end GPUs cost like 250 dollars - which granted was more money then than it is now, but not 100 friggin percent more...! 500+ for a top of the line board is really rough. At 250, many could justify yearly upgrades if they wanted to, but today you have to have quite a bit more disposable income for that, especially if you want something like a titan.

I suppose this is the way of the future. *sigh*
I can't recall a gen since NV30 where the top end card launched at less than $499, GTX 9800 maybe?
 
The CUs look like they've had their layout changed. They don't have the symmetry they had in the Tahiti shot. Possibly, some of the blocks were flipped.
Their orientation was also rotated, although the biggest visual discrepancy besides all that non-PHY area is in the non-CU area.

edit:
Did the scalar cache get replicated for all CUs?
 
The CUs look like they've had their layout changed. They don't have the symmetry they had in the Tahiti shot. Possibly, some of the blocks were flipped.

The GPR arrays are definitely re-organized, compared to what is visible in Tahiti, but at the same it's also different from Kabini's CU implementation.
 
A 384 bit 32 CU Tonga could have done significantly more damage than the current 285.
No it wouldn't. It's got less fillrate than GTX960, only 76%. And GTX960 has 64% of the bandwidth. And they're about equal in gaming performance (despite GTX960 also having way less compute).

http://techreport.com/review/27702/nvidia-geforce-gtx-960-graphics-card-reviewed/4

It needed more ROPs not more bandwidth. The architectural balance is all wrong. Unless it's hiding another 32 ROPs to go along with the extra 128-bit MC, it's basically a complete waste of time. This is why I call Tonga a test chip. It seems like a proof of concept for delta colour compression, since AMD seems to have gone from zero delta colour compression to something that appears to be competitive with NVidia (based purely on synthetics, though) in a single generation.

The real question is whether Fiji will have enough ROPs. Enough being 128, I reckon.
 
No it wouldn't. It's got less fillrate than GTX960, only 76%. And GTX960 has 64% of the bandwidth. And they're about equal in gaming performance (despite GTX960 also having way less compute).

http://techreport.com/review/27702/nvidia-geforce-gtx-960-graphics-card-reviewed/4

Less? I see more. Anyway this old Vantage benchmark is FP16 blend test. Colour writes are faster on GM206 due to the higher core clock -- both GPUs can output 32 fragments.

Also Tonga implements single-cycle FP16 blending and more bandwidth would have helped in some cases.
 
I thought GCN had a distributed register file (256 4-byte registers x 4 interleaved set per lane) from day one. As far as I know, lanes are isolated from each other, as reflected by the ISA design, and all the cross-lane operations are done through the LDS network (some without the need of allocation). With these, I don't see why the register file would be a huge collection of 2048-bit registers in hardware.
Why would you want to address registers per lane? 4 addresses are required per instruction: 3 for reads and 1 for the write.

With a 2048-bit register, there is no concept of "per lane". Remember this is a SIMD. All lanes are acting upon the same instruction. Including the same register ID.
 
Less? I see more. Anyway this old Vantage benchmark is FP16 blend test. Colour writes are faster on GM206 due to the higher core clock -- both GPUs can output 32 fragments.
I was referring to the theoreticals. Sadly TechReport didn't deploy the B3D fillrate tests for this article, so we can't see compressed versus uncompressed.

Also Tonga implements single-cycle FP16 blending and more bandwidth would have helped in some cases.
Can you find a game benchmark where R9 285 is substantially ahead due to raw bandwidth? GTX960 is fundamentally out of its depth according to all fundamentals except fillrate. And it has no trouble competing with R9 285.
 
Can you find a game benchmark where R9 285 is substantially ahead due to raw bandwidth? GTX960 is fundamentally out of its depth according to all fundamentals except fillrate. And it has no trouble competing with R9 285.

Haven't seen any bandwidth scaling tests around for Tonga. But sure, the 27% BW advantage for GM206 is also helping over the competition.
 
Back
Top