Tonga Versus Hawaii, Then and Now

[ultima_trev]

I find it interesting how much Tonga's performance has stagnated since its release akin to Hawaii which has gotten much stronger. Could it be that despite the GCN 1.2 ISA, Tonga didn't need to be optimized as much Hawaii as it's really just a slightly modified Tahiti?

7970 GHz VS 290
http://www.anandtech.com/bench/product/1031?vs=1068

290 VS 285
http://www.anandtech.com/bench/product/1333?vs=1068

390 VS 380X
http://www.anandtech.com/bench/product/1592?vs=1594

In the first comparison, the R9 290 doesn't pull that far away from the 7970 GHz. But in the second and third figures, the difference becomes far more pronounced...

 

[Silverforce11]

Because AT uses a reference 290. 947 core clock, but it actually throttles. I had lots for mining back in 2013.

390 @ 1050mhz with faster vram is easily 10-15% faster than ref 290.

 

[ShintaiDK]

380X is also clocked higher than 285.

The fact is Tonga owners are the big losers in the AMD pond. Nothing new.

 

[JDG1980]

I find it interesting how much Tonga's performance has stagnated since its release akin to Hawaii which has gotten much stronger. Could it be that despite the GCN 1.2 ISA, Tonga didn't need to be optimized as much Hawaii as it's really just a slightly modified Tahiti?

Tonga is, by far, the weirdest graphics chip AMD released on 28nm. It didn't represent any real improvement over its predecesor (Tahiti); not in absolute performance, nor in perf/watt, nor in perf/mm^2. This is despite the fact that it packed in about 16% more transistors in roughly the same die space as Tahiti.

No one has ever explained why AMD released this thing. My own speculation (and this is completely unconfirmed) is that Tonga was originally supposed to be AMD's first 20nm chip, and when that process turned out to be unsuitable, it was hastily back-ported to the 28nm mobile process. (Tonga and Fiji are on the TSMC 28nm mobile process, not the high-performance process that the rest of AMD's 28nm GPUs use.) I guess they figured it was at least cheaper for AIBs to build than Tahiti (due to being able to use a 256-bit memory bus without losing performance), even if it didn't have any other significant advantages. Or maybe they had the contract with Apple for the Retina iMac ahead of time and had to deliver something (Tahiti wouldn't work because of its 384-bit bus, which I think is incompatible with MXM).

 

[ShintaiDK]

Tonga is, by far, the weirdest graphics chip AMD released on 28nm. It didn't represent any real improvement over its predecesor (Tahiti); not in absolute performance, nor in perf/watt, nor in perf/mm^2. This is despite the fact that it packed in about 16% more transistors in roughly the same die space as Tahiti.

Performance/watt looks better:

 

[Piroko]

It's rather unconvenient that the games Anandtech tests are as old as they are. There are quite a few recent titles where Tonga benches better than Tahiti (Hitman) or even soundly beats anything in its class (The Division). In other titles it benches slightly slower (Tomb Raider) but still punches well in its price class.

Comparing it to Hawaii is a bit meh, they were never intended to be in the same price range. Also, it never got to show us how it would perform with 384 Bit and more ROPs. That amount of raw bandwith and ROP power might be the main reason why Hawaii performs so well in current games.
That might shift in favor of Tonga, as the tendency for newer game engines has always been to push for more and more shader load.

As for the hate by Shintai, that's a reliable sign that the card is a sleeper :sneaky:

 

[Adored]

I bought a 285 on the cheap to last me until Polaris/Pascal, been annoying to see it fail in the DX12 stuff tbh, especially in Ashes. Doesn't feel like a loss because it was so cheap but still a bit disconcerting.

 

[JDG1980]

Performance/watt looks better:

Compared to 280X on modern games, yes. But that's not the most efficient implementation of Tahiti - the memory controller is being pushed beyond its design intention. The HD 7950 is 16%-20% more efficient than the R9 285, or at least was at the time of its release.

Since they had to use a different process from the old chips anyway, it's surprising they didn't put Tonga on GloFo's 28SHP to fill their WSA quota. 28SHP was a better process than TSMC's, at least if the cat cores are any indication (GloFo Puma+ had higher clocks than TSMC Jaguar in the same power envelope, even though there were no architectural changes).

As an aside, a large part of AMD's perf/watt woes on 28nm is due to the fact that they're still using 2012 designs and pushing them much harder than was ever originally intended. GCN was made to run optimally at core speeds of 800-900 MHz, and the memory controllers were designed for a RAM clock of 1250 MHz. That worked fine against the first generation of Kepler, but to meet Maxwell in raw performance, efficiency had to go out the window. Every desktop graphics card AMD sells now is, in effect, factory overclocked.

 

[Head1985]

Tonga have some good IPC gains in some games.Same clock same memory bandwidth
http://www.hardware.fr/articles/945-24/tonga-vs-tahiti-round-2.html

 

[Piroko]

As an aside, a large part of AMD's perf/watt woes on 28nm is due to the fact that they're still using 2012 designs and pushing them much harder than was ever originally intended..

I disagree. GCN is known to undervolt well, their woes are mostly because AMD isn't binning very precise.

 

[KompuKare]

I disagree. GCN is known to undervolt well, their woes are mostly because AMD isn't binning very precise.

Binning and not bringing the advances they made on their APUs to their GPUs. Dynamic boosting and more power gating like on Kaveri and Carrizo would probably do wonders for their power consumption figures. It seems that APU gets to use up resources from the GPU division*, but the opposite does not happen. Maybe there really is a pride thing at AMD where CPU gets more resource but GPU does not despite GPU bringing in actual money.

*Always suspected that a lot of the GCN architecture ws designed with long-term APU goals in mind, and that in the short-term (until the last year's console ports) this design was actually hurting Radeons in terms of die sizes and power consumption.

 

[tential]

There were very few reasons to buy tonga or Fiji really so you should have known it wouldn't be the most cared for platform

 

[Silverforce11]

The 380 and 380X performs fine in modern games, and it's Tonga.

The only "neglect" was from Gears of Wars devs, and they fixed it pretty quick.

The extra transistors over Tahiti is for updated UVD blocks, more ACEs (Tonga Firepros and Virtualized workloads), updated tessellation engine and memory compression. Basically AMD designed it for Apple, Firepros and testing tech like compression, so they can get away with a 256 bit bus for a cheap SKU.

AMD's biggest mistake so far isn't even Fiji being unbalanced, it still is the reference 290/X. Imagine being the guy in those meetings: "What if I told you, for +$10 on build costs, we can make these GPUs run 10% faster, use less power and run very cool & quiet..." and have it shot down as a bad idea! lol

 

[grimpr]

I
As for the hate by Shintai, that's a reliable sign that the card is a sleeper :sneaky:

Great idea! we have a new metric and benchmark here, his eternal and infernal hatred for everything AMD, the more hate and venom, the better the product and buy! :thumbsup:

 

[grimpr]

Basically AMD designed it for Apple, Firepros and testing tech like compression, so they can get away with a 256 bit bus for a cheap SKU.

Nailed it, Tonga was really designed for Apple. :thumbsup:

 

[Silverforce11]

I bought a 285 on the cheap to last me until Polaris/Pascal, been annoying to see it fail in the DX12 stuff tbh, especially in Ashes. Doesn't feel like a loss because it was so cheap but still a bit disconcerting.

Why do you think the 285 is failing in DX12 Ashes? It's (380 = 285 with a new name) onpar with the 7970Ghz which is the full Tahiti with more SP, more bandwidth etc.

The 380X is at the same performance level, because Tonga's improved Tessellator is not in play in Ashes. In effect, the Tonga chip is giving similar performance with lower power usage and a smaller bus.

The result is the same at Computerbase.de

http://www.computerbase.de/2016-02/...afikkartenbenchmarks_von_full_hd_bis_ultra_hd

The 380 is matching the 280X (7970Ghz full chip!), despite being a cut-down Tonga.

 

[LTC8K6]

285 was crippled with only 2gb of vram, though...

Had it been released as a full Tonga with 4gb, it would have been a different story...

 

[Adored]

Guru3D's results were the same as mine during the beta 2.

I couldn't even run it above 15fps on anything above lowest settings and that was only 1080p. It's clearly a Tonga issue as the 380X tanks below the 370 as well.

It actually seems better in DX11 now (just tried) so maybe they fixed something late on.

 

[poofyhairguy]

Tonga has the most to benefit from an async future if that ever comes to pass, as it has as many ACEs as Hawaii (and the PS4). Until then it doesn't have a clear advantage over Tahiti.

 

[kawi6rr]

Great idea! we have a new metric and benchmark here, his eternal and infernal hatred for everything AMD, the more hate and venom, the better the product and buy! :thumbsup:

LOL, sounds good! It must be a great buy if its hated so much my that individual

 

[nenforcer]

I bought 2 x HIS Radeon R9 285 used in October 2014 for $300 shipped as my very first AMD graphics cards and for the price I've been impressed with the price/power/performance. I'm only gaming at 1080P and Star Wars:Battlefront (DX11) and Battlefield 4 (Mantle) (both DICE games) are mainly what I've played on them.

It's a little disappointing the driver performance hasn't improved along with Fiji as both cards are GCN 1.2 I believe but I believe they will be useful for quite a while since both consoles are GCN with the same number of ACE as somebody above said. The 2GB frame buffer will be limiting if I ever tried to go above 1080P but since they consoles can't do that I should be fine.