AMD: Southern Islands (7*** series) Speculation/ Rumour Thread

[Arty]

Doing a rough estimate, I would guess a dumb shrink of GF110 would reduce the die size from 529mm² to be something in the ~315mm² range (memory interface doesn't shrink, TSMC's HKMG processes offer only ~1.8 times the transistor density of 40G without following special layout rules). NVidia needs to put some more tricks in it, a shrunk GF110 isn't going to beat Tahiti in any convincing way. It may be able to match it, if they invest everything of the 28nm process advantage in increased clocks (meaning power consumption would be the same as a 40nm GF110). I think it is not a viable option.

Going off the same tangent, with my calculations I end up with a ~500mm2 die with ~1024CC. Again, this is based on the assumption that this chip is closer to Fermi than Kepler.

And if AMD does another Cayman for 28nm, we might see a slightly more aggressive packing density and they could get 40CUs in less than 400mm2 to respond to Kepler.

 

[boxleitnerb]

Thanks to the amazing drivers, you can actually go from 580 SLI to single 7970 and get a performance boost in DX11 titles

580SLI vs single 7970
BF3 33%+
Skyrim 17%+
F1 2011 -43%
Shogun 2: 18%+
Batman:AA -10%

Skyrim is DX9. And it all depends on the bench scene, obviously. 7970 will almost never beat a 580 SLI unless the latter is hampered by its memory amount. I'm a bit disappointed. 25% across the board faster than the 580 with a few exceptions here and there, but absolutely nothing to put that power to work. What fun it must be to play with CF in 1600p with shitty FXAA/ plain MSAA...not.

 

[rpg.314]

Looking back it's actually amazing how accurate some of the early speculations were (or maybe not considering groups of 4 CUs were a given...). The biggest surprise to me was that AMD didn't scale the amount of ROPs along the memory bandwidth (or at least improve the ROP z rate in some other way).

I guess they didn't need to. Anand's pixel fill rate tests show almost ~50% improvement with the same number of ROPs and per ROP throughput. Which must be mostly due to R/W cache and the 384 bit bus. In real world, the effect will obviously be less due to TMUs competing for the same, but for this iteration it seems fine.

May be next gen they will go back to 48 ROPs tied to mem channels.

 

[Man from Atlantis]

http://www.hardware.fr/articles/848-1/amd-radeon-hd-7970-crossfirex-test-28nm-gcn.html

 

[RobertR1]

The card seems to OC really well and more importantly, OC seems to scale well with performance unlike the 6900 series.

I'll have to keep an eye on how well the retail cards OC on a consistent basis. I would never keep a GPU stock so that's something that could pull me back in. Right now, stock 7970 vs my oc'd 580 is too close to warrant an upgrade but oc vs oc could prove a need esp if we get some quickl driver improvement.

 

[ECH]

There shouldn't, or only marginal (in some cases, PCIe 1.x x16 / 2.x x8 is limiting the performance already on single chip cards by couple %'s)

Well the results are in between PCIe 2.0 vs PCIe 3.0
Vr-Zone
Hardware Secrets
Anandtech

KitGuru actually used 2 different boards: MSI X58A-GD65 (i7 970 @ 4.6ghz) vs Asus Rampage IV Extreme (3960X EE @ 4.6ghz). Although I would have liked to have seen a P67, Sandybridge MB (2600K) to rule out any CPU bottleneck. But IMO, it's provides a wider picture then just using a x79 board that can use PCIe 2.0 via bios change. As you can see the result are more pronounced.


2 things come to mind:
-Are the bios mature enough to suggest that the results are minimal?
-Will Crossfire and/or 7990 make any noticeable use of PCIe 3?

I was still hoping to get a SS of a 32bit Hardware Reserve memory usage using the cards though. Haven't found that yet.


I do have another question. What's with the BF3 and Civ5 results? The difference between a 7970 and 580 are almost negligible. Do amd drivers support multi threaded rendering like nvidia? Or is something else the issue?

 

[CarstenS]

So has Tahiti already the capability to do FMAs in DP with a 1:2 rate or is there just the possibility to scale GCN to that in a future implementation (like the CI top model)?

Otherwise I read your statement as somewhat confirming Tahiti (not only the HD7900 line) is capable of 1:4 only.

No, with Tahiti AMD chose to implement the 1:4 option, saving transistors for other units and keeping the die small.

 

[silent_guy]

I wonder if the revamped memory pipeline and paging support lends itself to improved crossfire performance.

Yes, that, and PCIe3?

 

[3dilettante]

If there's a crossfire bench where they compared PCIe2 vs PCIe3, it might illuminate things, possibly.
There wasn't much difference when a PCIe2 x16 slot was split to two x8s, so it didn't seem like the bandwidth was really that much of a constraint.

It didn't seem like the bus was a significant factor until 3-4 cards got involved, but it is possible that with 28nm two high-end cards might notice a difference.

 

[mczak]

I guess they didn't need to. Anand's pixel fill rate tests show almost ~50% improvement with the same number of ROPs and per ROP throughput.

Please. Don't use that pure memory bandwidth test for anything concerning ROPs (though I think AMD only needs to double memory bandwidth again to finally hit the ROP limit there ).
That said, 32 ROPs might make sense from another angle, to prevent a situation like with GF110 - the chip may have 32 ROPs but really the primary purpose of half of them seems to be to serve as gates to the memory channels (ok there should be benefits for higher early-z rejection rate)... It will never output more than 16 pixels per clock anyway (it can't rasterize more, and even if it could it can't output more data per clock from the SMs neither). I bet Tahiti is still 2x16 pixels rasterization rate so at least more color ROPs would be a waste.
Maybe next gen can rasterize 4x16 pixels .
From that angle actually Pitcairn looks to be very decent.

 

[silent_guy]

If there's a crossfire bench where they compared PCIe2 vs PCIe3, it might illuminate things, possibly.
There wasn't much difference when a PCIe2 x16 slot was split to two x8s, so it didn't seem like the bandwidth was really that much of a constraint.

It didn't seem like the bus was a significant factor until 3-4 cards got involved, but it is possible that with 28nm two high-end cards might notice a difference.

Is the the shared virtual memory space and/or on-demand texture loading something that needs to be explicitly coded for? Or could a driver do this behind the scenes. AFAIK, the decision to put a texture in GPU memory or on system memory is already made by the driver, so to add more granularity could probably done in the driver as well.

 

[fellix]

I bet Tahiti is still 2x16 pixels rasterization rate so at least more color ROPs would be a waste.
Maybe next gen can rasterize 4x16 pixels .

Definitely, with that much bandwidth, alpha blend-rate now are at its peak in Tahiti, with the exception of HDR formats, where the limit is clearly in the ROPs. Still no depth/stencil benchmarks, though.

 

[caveman-jim]

But couldn´t nvidia redesign GTX 580 to fit in 250 mm2 with a increase in frecuencies and already match HD 7970?. If so imagine the 350 mm2 with the better Keppler architecture.

That seems unpossible to me.

~520mm2 to 250mm2 going from 40nm to 28nm? That would be some ...serious... optimizations, quite the feat of engineering. I would expect a pure die shrink to be around 350mm2, and then shrink more if there were layout optimizations to be had, or features to drop.

Anyway, back to Tahiti.

Speeed! Speeeeeeeeeeeeeeeed!

At 2560 and higher resolutions.

 

[ECH]

Hey, Wait A MINUTE...

Are all reviewers using STOCK GTX 580? I read this post and started to wonder what's going on.

Edit:
I found a reply to the previous post here. So a 7970 at stock clock still beats an overclocked GTX 580.

 

[Ailuros]

You want to delay product that is ready and the fastest thing to be available for the sake of a little performance in a title or two?

If it's only a title or two then they're doing a whole damn lot of damage to the average performance increase compared to its predecessor.

If that were the case we'd never release any new hardware. It not like performance isn't constantly improving on all hardware, it's just the rate of change will be higher with such a substantial change.

Since "delay" here is obviously a matter of perspective as the product in question did not under any circustamce meet it's originally promised timeframe, we could either be talking about a further delay or when foundry yields would had been better and drivers more mature and spend pages over nothing. I'll make it more simple for the case in point: AMD managed with its own strategy to significantly reduce GPU prices in the past years. If the competition should follow pace now and increase the prices too I'm not in the position to tell you what any company should or shouldn't do but at least I know which company to give credit to for it. And it obviously won't be positive this time.

 

[Alexko]

If it's only a title or two then they're doing a whole damn lot of damage to the average performance increase compared to its predecessor.

Since "delay" here is obviously a matter of perspective as the product in question did not under any circustamce meet it's originally promised timeframe, we could either be talking about a further delay or when foundry yields would had been better and drivers more mature and spend pages over nothing. I'll make it more simple for the case in point: AMD managed with its own strategy to significantly reduce GPU prices in the past years. If the competition should follow pace now and increase the prices too I'm not in the position to tell you what any company should or shouldn't do but at least I know which company to give credit to for it. And it obviously won't be positive this time.

I don't think general conclusions about AMD's pricing strategy should be drawn from this launch, given that the 7970 is:

• very high end,
• the first 28nm GPU—commercial chip?—ever,
• possibly not available in large quantities,
• the first DX 11.1 GPU,
• essentially without competition.

For all we know, Pitcairn and Cape Verde, or maybe even the 7950 could be priced quite aggressively. And if not, perhaps when Kepler is released. The latter might even trigger a price war, we just don't know.

 

[Ailuros]

I don't think general conclusions about AMD's pricing strategy should be drawn from this launch, given that the 7970 is:

• very high end,
• the first 28nm GPU—commercial chip?—ever,
• possibly not available in large quantities,
• the first DX 11.1 GPU,
• essentially without competition.

Change 28nm with 40nm and DX11.1 with DX11 and tell me why the 5870 had a much lower MSRP under that reasoning. I'll twist it even further, if the 7970 had a <$400 MSRP I should probably had expected by N% lower performance then I see today.

That "very high end" is obviously the joke of the day. Call it very high end marketing and I'll agree immediately.

For all we know, Pitcairn and Cape Verde, or maybe even the 7950 could be priced quite aggressively. And if not, perhaps when Kepler is released. The latter might even trigger a price war, we just don't know.

And if it shouldn't trigger a price war? As much as AMD isn't a charity organisation, NVIDIA is times less one.

 

[Gipsel]

Change 28nm with 40nm and DX11.1 with DX11 and tell me why the 5870 had a much lower MSRP under that reasoning.

I will give you just two reasons:
1. A HD7970 is more expensive to produce than a HD5870 (larger die, higher cost of 28nm wafer, 50% more memory chips and so on).
2. The price situation on the market is different. While the HD5870 was roughly the same amount faster than a GTX285 as the 7970 tops now a GTX580, the GTX285 was way cheaper back then than a GTX580 is today. In fact, AMD priced the HD5870 about 20% higher than the GTX285 (at least that was the price situation here in September 2009). That was (relatively speaking) higher than they are pricing the HD7970 now. And maybe one should consider that the MSRP of a 3GB HD7970 is comparable to or even lower than the current street price of a 3GB GTX580.

 

[Arty]

The price situation on the market is different. While the HD5870 was roughly the same amount faster than a GTX285 as the 7970 tops now a GTX580, the GTX285 was way cheaper back then than a GTX580 is today. In fact, AMD priced the HD5870 about 20% higher than the GTX285 (at least that was the price situation here in September 2009). That was (relatively speaking) higher than they are pricing the HD7970 now. And maybe one should consider that the MSRP of a 3GB HD7970 is comparable to or even lower than the current street price of a 3GB GTX580.

Bingo. And the GTX285 was cheaper back then because of RV770, it was a vicious circle for the IHVs and it came to an end.

 

[Lightman]

To further add to Gipsel post:
- since 2009 we had few percent inflation each year,
- TSMC initial per wafer cost of 40nm was lower than it is for 28nm,
- they were manufacturing constrained for long few months because they priced HD58xx so competitively, a mistake big business rarely repeats twice in the same decade,
- AMD has new CEO who wants to get company back on right financial track and therefore our hero Dave even if tried to convince his colleagues to lower launch price he was on lost position.

I would buy this card 1st day out if it costed $100 less, but it's not and I'm sure AMD will still sell everything they can produce! Then, one day when either AMD launches HD7990 or maybe nVidia shows Kepler the price will fall to justifiable level for me and I will upgrade from my HD6970!

Other than price I really like all the new features, and where I was hoping for big improvements (GPGPU), Thai shines!