videocardzAMD Radeon R9 M295X with Tonga GPU has 32 Compute Units

[csbin]

http://videocardz.com/50737/amd-radeon-r9-m295x-tonga-gpu-32-compute-units

Some fresh leaks have been discovered by one of our most dedicated readers, Cloudfire. What he discovered are the specs of the upcoming Tonga GPU, which as you know is being slowly being prepared by AMD. Tonga is based on the same GCN architecture known from Hawaii or Bonaire. Tonga is actually a Tahiti replacement, but since it’s much more power efficient, AMD has decided to replace Pitcairn/Curacao/Neptune with something fresh.
Tonga has the same amount of Compute Units as Tahiti. Assuming it has the same SP layout as other Volcanic Islands GPUs, then we should expect 2048 unified cores. The number of TMUs and ROPs is debatable at this point, as we don’t know if mobile Tonga will have the same numbers as desktop variant. The best scenario is 128 TMUs and 32 ROPs respectively. This would be the exact same chip as Tahiti in terms of specs, but not in terms of the interface. Tonga has 256-bit wide memory bus, which is more or less a standard for mobile graphics processors.
According to leaked specs, the maximum core clock should not exceed 800 MHz. Of course, that would be the boost clock. AMD Radeon R9 M295X would feature 4GB GDDR5, but it does not mean we won’t be seeing 8GB variants too. This is undoubtedly the biggest advantage of GTX 880M right now, so it would be silly if AMD had no such option being prepared.

 

[tviceman]

A few things.

1) 28nm lives on.

2) Considering that Hawaii is hardly more efficient than Tahiti, there has to be more going on than just chopping Hawaii down in size in order to get the power and thermal constraints to fit within a notebook.

3) I am glad to see both companies innovating despite not being able to move on to a more advanced node yet.

4) The 8gb vram for the gtx880m is silly. Less than 1% of anyone in the market looking for that type of card will want or be willing to pay the premium for 4 more gigs of vram, especially when by the time that much vram is needed that type of card will be fairly slow.

 

[DiogoDX]

Will be interesting if this fit the same TDP of M290X.

 

[3DVagabond]

This chip is actually Pitcairns replacement. At the moment it's slotting into Tahiti's spot in the product stack. This tells me it's going to be ~$300 product, at least at first (Not happy about that!).

As far as it's efficiency, it's very likely to be compute limited compared to Hawaii and Tahiti. I'm guessing (only guessing) that it's going to show a generational improvement over Pitcairn efficiency wise. Likely won't be super impressive without a process shrink to go along with it, but decent none the less.

So much for my Nostradamus impersonation of the day.

 

[parvadomus]

This chip is actually Pitcairns replacement. At the moment it's slotting into Tahiti's spot in the product stack. This tells me it's going to be ~$300 product, at least at first (Not happy about that!).

As far as it's efficiency, it's very likely to be compute limited compared to Hawaii and Tahiti. I'm guessing (only guessing) that it's going to show a generational improvement over Pitcairn efficiency wise. Likely won't be super impressive without a process shrink to go along with it, but decent none the less.

So much for my Nostradamus impersonation of the day.

Tonga is the first Volcanic Islands GPU, I expect a nice efficiency improvement.

 

[NostaSeronx]

I don't see why it isn't using 512-bit GDDR5.

http://content.hwigroup.net/images/articles/290x-hawaii-6.PNG

4 GBs is pretty much telling of a 512-bit interface.

 

[SPBHM]

I don't see why it isn't using 512-bit GDDR5.

http://content.hwigroup.net/images/articles/290x-hawaii-6.PNG

4 GBs is pretty much telling of a 512-bit interface.

512bit with 32ROPs is probably not the most efficient choice, 256bit is probably the way to go, and you can have 4GB with no problem.

 

[NostaSeronx]

512bit with 32ROPs is probably not the most efficient choice, 256bit is probably the way to go, and you can have 4GB with no problem.

Actually, it is an issue with the MXM-B spec. It can only have devices with 256-bit I/Os.

----
Also:
http://beyond3d.com/showpost.php?p=1848298&postcount=97

 

[el etro]

This M2*5x...

The five maybe indicates that we have no VI GPU per now....

 

[TreVader]

This should put AMD in a good place in mobile. They used to dominate mobile GPU


Would like to see something like this on 20nm in the next MacBook Pro

 

[NostaSeronx]

For me comparing Tahiti and Tonga in OpenXL random ray-tracing opencl:

v_add_i32 -> v_add_u32
v_sub_i32 -> v_sub_u32
v_mul_lo_i32 -> v_mul_lo_u32
v_fma_f32 -> v_madmk_f32(x5)

some additional:
s_add_u32
s_not_b32
s_nop

some removals of:
v_lshrrev_b32
s_nop

 

[VulgarDisplay]

A few things.

1) 28nm lives on.

2) Considering that Hawaii is hardly more efficient than Tahiti, there has to be more going on than just chopping Hawaii down in size in order to get the power and thermal constraints to fit within a notebook.

3) I am glad to see both companies innovating despite not being able to move on to a more advanced node yet.

4) The 8gb vram for the gtx880m is silly. Less than 1% of anyone in the market looking for that type of card will want or be willing to pay the premium for 4 more gigs of vram, especially when by the time that much vram is needed that type of card will be fairly slow.

They don't necessarily have to pack this chip as tightly as Hawaii was which would likely keep heat and power in check.

 

[Techhog]

That is a really low clock speed. It might not even beat the GTX 880M.

 

[NTMBK]

My question is how much cache will this have? The oversized cache is the main reason GM107 proved more efficient, as it went out to main memory less often.

 

[tviceman]

They don't necessarily have to pack this chip as tightly as Hawaii was which would likely keep heat and power in check.

Instead of actually significantly improving their architecture, they opt to spread out transistors?

I highly doubt transistor density is going to be tweaked on this chip such that it purposefully makes a dramatic impact on thermals. AMD is already selling their chips at lower margins and in lower volume than nvidia. The last thing they want to do is overshoot die size, equaling less chips per wafer, just so they can drop TDP by 4 watts.

 

[MeldarthX]

or possibly the rumors of gefo's 28 process isn't as leaky as TSM 28 process.....since these are supposed to be coming from there.....

 

[tviceman]

I am going to stick to my theory: AMD made actual architectural improvements to GCN's efficiency.

 

[raghu78]

yeah there is a good possibility that Tonga is manufactured at GloFo 28nm. AMD stated in their Q1 earnings call that they have started shipping GPUs manufactured at GF 28nm. Tonga could be the first AMD GPU at GF 28nm. Since Beema and Mullins are made at GF 28nm and have benefited from process improvements when moving from TSMC to GF, there is room for Tonga to improve power efficiency. Secondly the 256 bit bus means lesser power dissipation compared to Tahiti which sports 384 bit memory controller. Tonga mobile is running at 1375 Mhz GDDR5. Tonga desktop should run at 1750 Mhz GDDR5 similar to GTX 770.

http://www.anandtech.com/show/7974/...hitecture-a10-micro-6700t-performance-preview

"AMD claims a 19% reduction in core leakage/static current for Puma+ compared to Jaguar at 1.2V, and a 38% reduction for the GPU. The drop in leakage directly contributes to a substantially lower power profile for Beema and Mullins."