AMD "Tonga" GPU Arrives This August

[btarunr]

In a bid to counter NVIDIA's bestselling GeForce GTX 760, AMD is preparing a new 28 nm GPU, codenamed "Tonga," as detailed in our older article on the chip. At the time of its writing, we had two theories on what "Tonga" could be, one held that it could be a counter to the GM107, and the other, that's is a step above "Curacao," in a bid to counter the GTX 760. We're now learning that AMD could launch the first graphics cards based on this chip, some time in August. The chip will replace the ailing "Tahiti Pro" silicon, from which is carved out the Radeon R9 280. While the R9 280 offers performance competitive to the GTX 760, it loses out big time on power consumption and heat. The cheaper R9 270X, on the other hand, offers lower performance, and similar power levels. "Tonga" could offer nearly as much performance, while featuring a new combination of components, that help AMD lower not just power draw, but also overall costs.

The 28 nm "Tonga" silicon is expected to feature 2,048 GCN2 stream processors, 128 TMUs, 32 ROPs, and a 256-bit wide GDDR5 memory interface. You'll notice that its memory interface is narrower than that of the R9 280, but performance is made up for with a higher stream processor count, and probably higher clock speeds, too. The card could come in memory capacities of 2 GB, with some manufacturers innovating 4 GB variants. There's no word on what the company could end up naming the first cards running this chip.



View at TechPowerUp Main Site

 

[EzioAs]

If this is to be a mid-range card, please let the VRAM be more than 2GB. With 4K monitors starting to get cheaper and next-gen consoles (also upcoming games) leveraging more VRAM, next-gen GPUs should have more VRAM than previous ones.

 

[FrustratedGarrett]

If the chip does indeed pack 2048 GCN2 stream processors, then it's most likely faster than the 280X and could potentially be competitive with 290/290X card while consuming much less power on load.

 

[arbiter]

If the chip does indeed pack 2048 GCN2 stream processors, then it's most likely faster than the 280X and could potentially be competitive with 290/290X card while consuming much less power on load.

eh that narrower bus could counter act all that, they say higher count will make up for it but that is not always the case. Back when nvidia released 8800 cards, I had 8800gts 320mb card and when it died didn't want to watch for rma crap so got 8800gt. Despite same thing as 8800gt had more shaders then it had a norrower bus then gts did, 256bit vs 320bit. 112 shaders vs 96, when i played TF2 on 8800gt i noticed fps tanked a ton on the gt where never had problem before on gts. So lowering memory path in favor of more shaders can impact end fps. that was only 64bit difference this is 128bit so. No this not trying to start an nvidia vs amd spat, just my experience when they narrow memory path in favor of more shaders.

If the chip does indeed pack 2048 GCN2 stream processors, then it's most likely faster than the 280X and could potentially be competitive with 290/290X card while consuming much less power on load.

280X has 2048 as well but has 384bit bus so it will have an advantage, though since don't know clock speed of the ram its hard to say how much. If its 1500mhz then would put bandwidth around just under 200GB/s which is a bit short of 280x which is around 288GB/s. AMD likely will push clock higher then 1500mhz probably 1650-1750 maybe (all quad DDR so have to times it by 4.) to help it our less they don't want to make 280x unsellable. Even at 1750 only puts it at 224GB/sec. No matter how you cut it still gonna effect avg fps.

 

[HalfAHertz]

Will it be based on GCN 1.1 like the 260x and 290/290x? If so it will include goodies like revised shaders and TrueAudio. Maybe it will feature the revised memory controller of the 290 that supposedly was 50% smaller.

 

[Prima.Vera]

So no new generation this year?

 

[CookieMonsta]

This thing is identical to a bog standard 7970/280x in almost every metric except memory bandwidth and power consumption.
The loss of memory bandwidth (maybe not actual memory capacity as the 256bit bus will allow 4gb ala Hawaii) will deeply impact high resolution and heavy MSAA/SSAA performance. I expect this SKU to be extremely competitive with Tahiti at 1080p and under but it will bleed performance really fast at 1440p and beyond. That being said, if it comes equipped with 7Ghz or 7.5Ghz VRAM, the performance would be quite close to a stock 7970 of similar clockspeed. Something tells me they will equip this with 6Ghz modules or even 5Ghz ones if they recycle the Hawaii Memory controller design.
If it is priced at around the $250-$300 mark, its a real winner for the mainstream.

 

[RCoon]

So no new generation this year?

With TSMC and GloFo sucking balls at the new smaller wafer production, we're stuck on 28nm for a while.

 

[Brusfantomet]

This thing is identical to a bog standard 7970/280x in almost every metric except memory bandwidth and power consumption.
The loss of memory bandwidth (maybe not actual memory capacity as the 256bit bus will allow 4gb ala Hawaii) will deeply impact high resolution and heavy MSAA/SSAA performance. I expect this SKU to be extremely competitive with Tahiti at 1080p and under but it will bleed performance really fast at 1440p and beyond. That being said, if it comes equipped with 7Ghz or 7.5Ghz VRAM, the performance would be quite close to a stock 7970 of similar clockspeed. Something tells me they will equip this with 6Ghz modules or even 5Ghz ones if they recycle the Hawaii Memory controller design.
If it is priced at around the $250-$300 mark, its a real winner for the mainstream.

Remember that the 760 is a 2 GiB at 256 bit/6 GHz part as well, so ass you say its going to be good for 1080p. but i guess we could simulate this easy, just down clock the memorybuss of a 280/7950 or compare the 7870 XT (non Ghz edition) as its a Thati with 256 bit bus to a 7950

 

[alwayssts]

Will it be based on GCN 1.1 like the 260x and 290/290x? If so it will include goodies like revised shaders and TrueAudio. Maybe it will feature the revised memory controller of the 290 that supposedly was 50% smaller.

I certainly agree odds are it will contain not only trueaudio, but also adaptive sync (as Richard Huddy has said without saying in multiple interviews) and a greater pixel clock ability. As for the memory controller, while using something similar to Pitcairn/Hawaii in 384-bit/5ghz-5.5ghz guise is one option, It would appear they went the opposite direction with 256-bit/7ghz memory...something closer to what's in Bonaire. That would suggest a card with a shorter design (<10.5 inches), aimed at 1080p, and/or perhaps a die size that couldn't sustain 384-bit/5500...essentially a replacement for the market Pitcairn used to occupy at launch, Tahiti occupied recently (around $200-350), and similarish to rv790->Barts later-series refresh parts of yore. While 7ghz would appear to be less efficient than wider and slower (when die space allows), 7ghz now is only 1.5-1.55v (as opposed to most-everything being 1.6v around the 7000 series launch). It also allows a pretty nifty setup in mobile; The same chips can be binned to 1.35v (5.5ghz from Hynix, same thing used in the ps4), and such a setup could allow a sku with power and yield friendly core clock of ~800mhz.

It's important to note that while Tahiti had it's merits for design (during it's heyday), a 384-bit/6ghz (1.6v) setup could essentially support 2048sp up to around 1350mhz. While that made sense for early silicon, now-a-days thing have evolved, with both GK100/Hawaii being higher up the foodchain and replacing Tahiti for the larger buffer/higher rez and also power envelope solutions. Clockspeeds on 28nm have also settled lower than on initial products as yield has increased...or at least that's how amd has been playing it.

Extrapolating, and assuming 2048sp is correct, I assume we get something around 1000-1050/7ghz. While on straight compute unit/bw efficiency it should have a clock similar to Hawaii, if you rather use a metric of what the average shader/fillrate usage is (ie they want to clock it pretending they are usually only using around 1880 units, essentially the practical difference between gk104 and Tahiti) it would be closer to 1050mhz or so.

I think the grand take-away will be exactly what this article puts forward: It will be a part between 760 and 770. Probably overclock to a similarish level to a stock 770. Hopefully it will be priced between the two, if not closer to the former, as well as being shorter and using under 225w (which using 2GB would imply...4GB may hamper clocks within that envelope)...essentially bringing Tahiti to Pitcairn's market (over time). IOW, exactly what others have postulated: this is (or will be, depending on your pov) the new standard for 1080p.

Anecdote and/or TLDR:

xbone is to Pitcairn (7870/270x) as PS4 is to....?
(Think about that for a minute, and do some 30->60fps and 720p->1080p conversions).

Make sense?

 

[Casecutter]

We know Tonga is 28nm, but then is there anymore regarding it being made by GloFlo? If that's the case AMD didn't switch (risk) going with them without an improvement in the cost per chip.

Is there any truth that this is Pro version, and that there could be an XT as 2432sp part, but is holding it back (GloFlo having teething pain on yields)? It's certain this will be much more than the 760, while perhaps right on the back of the 770, and I'm inclined to think a MSRP of $240. Considering quite a number of 280's have been $200 here lately it holds some merit. Then latter an XT with 2432sp for $280-300, they could drop Tahiti altogether, and their margins would be excellent. Then if there is a 295 (full XTX Hawaii) with HBM could they bring it in at $600, and drop 290/290X to slot in at $350/500, while AMD has <$200 all tied up.

Giving the situation (TSCM 20nm miss-step) AMD is in a good place bringing GloFlo in and cultivating their production. If this is what's happening AMD might see the chance to perhaps not take a huge bite of market share, but provide a true alaround 1080p card for under $250, and have them see really good margins on those sales.

...being shorter and using under 225w (which using 2GB would imply...4GB may hamper clocks within that envelope)... the new standard for 1080p.

I agree with everything but 225W! The 7950/280's are 200W parts, a 760 is 170W I couldn't imagine 225W making the grade. I would hope that efficiency is a big part of the mix and if it isn’t in the new architecture (GCN2.0 or whatever) it's no-starter. Especially if from GloFlo as that would be a terrible reflection on them, it would be hard for GloFlo to come out of the gate making that for AMD if that is what Tonga delivers.

I think on this they’ll want to be at minimum closer to 170W. We won’t see GK107 type efficiency, but for a true 1080p full-Enthusiast quality experience I think in perf/w it will be fine if they can emulate what OC’d GTX750Ti have provided.

 

[FrustratedGarrett]

I think the rumor about the Tonga GPU packing 2048 ALUs is wrong! The Tonga GPU is very likely smaller than Tahiti with less ALUs and a narrower memory controller, measuring ~250 mm^2, and it will likely consume ~100 Watts under load while performing in the same league as Tahitis. It will be based on GCN 2.0 and I doubt the guys at AMD have been sitting around doing nothing for two years and a half with the vanilla GCN micro architecture.

 

[sweet]

AMD cards have more VRAM > more power, and also wider bus lead to more memory controller > more heat. If you ask them for a power efficient card, you will have a less VRAM, less bus card. And that is NOT the card I expect from AMD.

 

[HalfAHertz]

it shouldn't be as energy efficient as the 760, it must be more energy efficient or it would be considered as a failure. Because a few months after it the 760's successor based on Maxwell will be coming out. And it will have to match that or bite the dust.

 

[Xzibit]

it shouldn't be as energy efficient as the 760, it must be more energy efficient or it would be considered as a failure. Because a few months after it the 760's successor based on Maxwell will be coming out. And it will have to match that or bite the dust.

Energy efficiency trumps performance and price now ?

 

[Peter1986C]

One can only have performance when being efficient. Remember Netburst?

 

[LemmingOverlord]

Well, the new silicon doesn't mean they discovered fire. They are just making options to align their product with their rival Nvidia.

In order to keep the cards efficient (ie: performing without shedding too much heat), AMD can do a few things

1) A 256-bit memory bus (actually 2x128-bit) = 2GB of GDDR5 RAM. That's the same as the GTX 760 and I don't think it'll make much of a difference in terms of a lower framebuffer than the R9 280. AMD can stick in 1.5GHz GDDR5 instead of 1.4GHz, making up for the loss of bandwidth with increased throughput. This also makes the card cheaper.

2) Reduce GPU overhead = this a bit more arbitrary and involves a lot of software optimization, not only hardware optimization... this will keep both idle and active power consumption lower than the current Tahiti Pro.

3) Better control over GPU power states = also through hardware and software optimization, AMD will need to really get this right if they want to cut idle power, which is usually higher than Nvidia's. Power/Clocking up/down efficiently is critical.

4) Reduce Double Precision FP performance = The Tahiti XT GPU implementation has a massive advantage over its siblings in Double Precision FP performance, and even a greater one over the GTX 760. If AMD can trade-off Double Precision FP performance for power savings...

All in all, and staying with the 28nm node, there isn't much more that AMD can do except adjust to the competition's own products in the same shelf.

 

[Aquinus]

Well, the new silicon doesn't mean they discovered fire. They are just making options to align their product with their rival Nvidia.

In order to keep the cards efficient (ie: performing without shedding too much heat), AMD can do a few things

1) A 256-bit memory bus (actually 2x128-bit) = 2GB of GDDR5 RAM. That's the same as the GTX 760 and I don't think it'll make much of a difference in terms of a lower framebuffer than the R9 280. AMD can stick in 1.5GHz GDDR5 instead of 1.4GHz, making up for the loss of bandwidth with increased throughput. This also makes the card cheaper.

2) Reduce GPU overhead = this a bit more arbitrary and involves a lot of software optimization, not only hardware optimization... this will keep both idle and active power consumption lower than the current Tahiti Pro.

3) Better control over GPU power states = also through hardware and software optimization, AMD will need to really get this right if they want to cut idle power, which is usually higher than Nvidia's. Power/Clocking up/down efficiently is critical.

4) Reduce Double Precision FP performance = The Tahiti XT GPU implementation has a massive advantage over its siblings in Double Precision FP performance, and even a greater one over the GTX 760. If AMD can trade-off Double Precision FP performance for power savings...

All in all, and staying with the 28nm node, there isn't much more that AMD can do except adjust to the competition's own products in the same shelf.

Nvidia didn't reduce DP performance for power consumption. They did it to expand the regular SP capacity which is what games typically use. DP gets used in gpgpu a lot more than rendering.

 

[HalfAHertz]

Well, the new silicon doesn't mean they discovered fire. They are just making options to align their product with their rival Nvidia.

In order to keep the cards efficient (ie: performing without shedding too much heat), AMD can do a few things

1) A 256-bit memory bus (actually 2x128-bit) = 2GB of GDDR5 RAM. That's the same as the GTX 760 and I don't think it'll make much of a difference in terms of a lower framebuffer than the R9 280. AMD can stick in 1.5GHz GDDR5 instead of 1.4GHz, making up for the loss of bandwidth with increased throughput. This also makes the card cheaper.

2) Reduce GPU overhead = this a bit more arbitrary and involves a lot of software optimization, not only hardware optimization... this will keep both idle and active power consumption lower than the current Tahiti Pro.

3) Better control over GPU power states = also through hardware and software optimization, AMD will need to really get this right if they want to cut idle power, which is usually higher than Nvidia's. Power/Clocking up/down efficiently is critical.

4) Reduce Double Precision FP performance = The Tahiti XT GPU implementation has a massive advantage over its siblings in Double Precision FP performance, and even a greater one over the GTX 760. If AMD can trade-off Double Precision FP performance for power savings...

All in all, and staying with the 28nm node, there isn't much more that AMD can do except adjust to the competition's own products in the same shelf.

Number 4 isa great point. That's one of the reasons the 680/770 has such a good performance/power ratio.

 

[LemmingOverlord]

Nvidia didn't reduce DP performance for power consumption. They did it to expand the regular SP capacity which is what games typically use. DP gets used in gpgpu a lot more than rendering.

Agreed. AMD has that choice to make. Of course they can opt to have the best of both worlds, I think. They can kill DP performance, boost SP performance (although I doubt they'll to do that), and bring down that massive 250W TDP from the 280 to a greener... I dunno ~200W?

Some sites are calling Tonga a replacement for Tahiti, but if we're talking about slimming down the card, it's not a replacement, but a new SKU... with a different price point. Any ideas?

 

[tokyoduong]

Agreed. AMD has that choice to make. Of course they can opt to have the best of both worlds, I think. They can kill DP performance, boost SP performance (although I doubt they'll to do that), and bring down that massive 250W TDP from the 280 to a greener... I dunno ~200W?

Some sites are calling Tonga a replacement for Tahiti, but if we're talking about slimming down the card, it's not a replacement, but a new SKU... with a different price point. Any ideas?

I don't think the 280 has 250W TDP

 

[Casecutter]

Agreed. AMD has that choice to make. Of course they can opt to have the best of both worlds, I think. They can kill DP performance, boost SP performance (although I doubt they'll to do that), and bring down that massive 250W TDP from the 280 to a greener... I dunno ~200W?

Some sites are calling Tonga a replacement for Tahiti, but if we're talking about slimming down the card, it's not a replacement, but a new SKU... with a different price point. Any ideas?

I agree but where are you getting 250W for the R9 280 is see a TDP of 200W?
http://www.techpowerup.com/gpudb/2405/radeon-r9-280.html

I think AMD is focused on only the superlative in Gaming Mainstream 1080p. They can knock down a good amount of the DP for more SP / efficiency, all in a smaller package. I'd think a 25% drop in TDP might be realized, they have to just be as proficient vying GK104’s not even looking forward to if Maxwell scales. A very cost effective card for the bulk of gaming.

 

[arbiter]

Well, the new silicon doesn't mean they discovered fire. They are just making options to align their product with their rival Nvidia.

In order to keep the cards efficient (ie: performing without shedding too much heat), AMD can do a few things

1) A 256-bit memory bus (actually 2x128-bit) = 2GB of GDDR5 RAM. That's the same as the GTX 760 and I don't think it'll make much of a difference in terms of a lower framebuffer than the R9 280. AMD can stick in 1.5GHz GDDR5 instead of 1.4GHz, making up for the loss of bandwidth with increased throughput. This also makes the card cheaper.

Um By my calculations even using boosting ram to 1500mhz over 1400 with that bit path drop, the card will suffer a 28% memory bandwidth hit. 1400mhz @ 384bit push's 279GB/s, 1500mhz @ 256bit only hits around 192GB/s. works out to about 28%. Will their be a performance hit even with extra shaders, most def there will be.

 

[LemmingOverlord]

I agree but where are you getting 250W for the R9 280 is see a TDP of 200W?
http://www.techpowerup.com/gpudb/2405/radeon-r9-280.html

I think AMD is focused on only the superlative in Gaming Mainstream 1080p. They can knock down a good amount of the DP for more SP / efficiency, all in a smaller package. I'd think a 25% drop in TDP might be realized, they have to just be as proficient vying GK104’s not even looking forward to if Maxwell scales. A very cost effective card for the bulk of gaming.

Sorry, got it from http://www.techpowerup.com/gpudb/2398/radeon-r9-280x.html and http://en.wikipedia.org/wiki/AMD_Radeon_Rx_200_Series could be I'm misinterpreting?

Still, they need to pull off a cheap power-saving stunt without sacrificing too much performance to trump the GTX 760.

I used the example of using higher-rated GDDR5 even with the bandwidth reduction as it would match what Nvidia has on the GTX 760 ... and all things equal, AMD needs to work on a whole bunch of things to get it down to the GTX 760 power envelope.

 

[Casecutter]

AMD needs to work on a whole bunch of things to get it down to the GTX 760 power envelope.

Well starting at 200W TDP... if Tonga is supplied out of GloFo they have already indicating their 28nm process offers efficiency so if the give 10% just in gate improvement. Then AMD finds 10-15% (Bonaire GCN1.1 provided about that) and figure there’s a lot that they can cut-out from a Tahiti. AMD can juggle out a good amount and still provide superlative 1080p performance, which a 280 already provides in spades, they deliver Perf/W of 20-25% more than a 270X, while at an MSRP of $240 it will get my blood moving!