[bit-tech] Radeon HD 7000-series rumoured for May production

[cusideabelincoln]

Except that Cayman is about 10% faster per transistor than Cypress/Barts is. So you can't estimate HD7970's performance increase over HD6970 based on how much faster HD6870 was over HD4870.

Now you are backtracking and misconstruing my argument? You said, "The proper comparison is to look at HD4870 to the HD5870 since HD4870 is a high end card and HD6870 is a mid-range card." I'm simply stating it's just as valid to compare the 4870 to the 6870 as it is to the 5870. You get different viewpoints doing both methods. One gives you an idea of the jump in performance using the same die size (important), and the other gives you an idea of the jump in performance when implementing new features...

Also, the PCIe 3.0 controller will likely take up more space than the current 2.1 controller. Then again AMD added dual graphics engines into Cayman. That surely took up extra space. Will they add more graphics engines? Even then we are making it too simplistic since AMD can increase transistor density by re-jiggling the GPU design to make it more efficient:

Cayman packed 23% more transistors into 16% more die size compared to Cypress, while Tessellation performance increased 1.5~3x! Simply by going to 28nm, AMD would be able increase SPs, TMUs and ROPs on their next chip. That doesn't even account for any other tricks they may have up their sleeve

And going from the 4870 to the 5870/6870 AMD had to dedicate more die to the memory controller and they did other things as well besides just increase the SP count.

(recall AMD doing a re-spin of the HD4870 on the same 55nm process and they netted another 13% increase in clock speed in the HD4890). I think you guys are underestimating the jump to 28nm. It will be FAR more impressive than the 15% performance increase of HD5870 --> HD6970.

I don't think we are. I personally don't think AMD is going to make a die size as big as Cayman. As such, it's going to be relatively harder for them to reach this magical 100+% increase. You think comparing the 4870 to the 5870 is valid, but despite going down a process node they still had to significantly increase the die size. Adding features is going to cost relatively more die size than not adding those features when you jump from Cayman to NI. And Cayman is probably the die size limit for AMD, at least IMO.

Do you get what we're saying now? We get what you're saying: Efficiency. AMD will improve it. No doubt, I think they will too. But I also think they'll use this efficiency to get the die size down. That has been AMD's operating philosophy the last few generations, and it has worked for them. Now if they change their gameplan then everything is thrown out the window and you'll be right.

 

[Concillian]

If AMD is shooting for a very quick release, then it will keep die size on the small side and go with a relatively "safe" design that it can scale down and push out quick. Then they'll work on the follow on to answer whatever nVidia pushes out.

There's no reason for the initial release to be large if they suspect they will beat nVidia to market. I know I am waiting on 28nm and I'm sure many others are too. Some don't care about branding, since the first out of the gate are pretty much guaranteed to be significantly faster than the current offerings at a given price point.

They're in the game to make money, and if you want to make money, you want a small, die with modest power requirements that has performance equal or better than a 6870, but in the $150 range.

I don't expect huge jumps because AMD is so focused on the business aspects. We will likely never get another GTX8800 / HD4850 just because they are more aware of what the market will bear in terms of price. There are SO MANY card models out there at tiny performance / price intervals. They are hyper-attuned to how much people will pay for performance, and they will adjust cost to demand quickly (like 5850 did) rather than keep prices that will sell larger volumes (like 4850 did).

The world has changed. Not necessarily for the better for consumers, but in a way that will keep AMD from having to worry about being swallowed by it's own debt.

All of these factors, which are admittedly speculation, but it has me believing that the first release will be something like a "7750/ 7770" range card that has performance like a 6850 but on a die size closer to a 5770 (perhaps with a bit more die space to fit the pads for a 256bit memory bus). This gen does have the disadvantage that they can't move from DDR3 to DDR5 to get the extra memory bandwidth. The gains in the low to midrange cards from 4xxx --> 5xxx were as much from extra memory bandwidth as from extra transistors. There will be a smaller boost in bandwidth this gen. GDDR5 is still GDDR5. Some memory controller revision might make better GDDR5 possible, but expect only 10-20% type bandwidth improvements.

 

[RussianSensation]

The gains in the low to midrange cards from 4xxx --> 5xxx were as much from extra memory bandwidth as from extra transistors. There will be a smaller boost in bandwidth this gen. GDDR5 is still GDDR5. Some memory controller revision might make better GDDR5 possible, but expect only 10-20% type bandwidth improvements.

If you compare HD4870: 115 GB/sec vs. HD5770: 77 GB/sec or even better HD4890: 125 GB/sec vs. HD5850: 128 GB/sec bandwith, it becomes very pretty clear that almost all of the performance gains from HD4xxx to HD5xxx generation were not related to memory bandwidth improvements.

There is also no indication that current top cards are memory bandwidth limited. HD6970 has more than enough memory bandwidth. Certainly when I run my GPU at HD6950's memory bandwidth or HD6970's memory bandwidth, the performance difference is very negligible. So I don't think it's critical to worry about memory bandwidth yet (they can always put faster GDDR5 chips to get another 10-15% memory bandwidth increase).

Also, HD2900XT series had 512-bit of memory bandwidth. So AMD doesn't necessarily need to stick to 256-bit memory bus.

But I also think they'll use this efficiency to get the die size down. That has been AMD's operating philosophy the last few generations, and it has worked for them.

That's probably a reasonable assumption, but not guaranteed based on history.

X1800XTX = 264 mm^2
X1950XTX = 315 mm^2
HD2900 XT = 420 mm^2

There is actually a very clear trend that AMD GPUs have grown in die size at a rate of ~15-30% per generation in the last 2-3 years. While their die sizes are smaller compared to NV's chips latest, you can't say that AMD's philosophy has been to decrease the die size:

More recently
HD3870 = 192 mm^2
HD4870 = 260 mm^2 (+35%)
HD5870 = 334 mm^2 (+28%)
HD6970 = 389 mm^2 (+16%)

It sounds reasonable that AMD may try to shrink the die size since they are moving to a 28nm process. At the same time, since they have been successful in the last 3+ years when they have increased the die size, they could try for a 28nm 400 mm^2 monster chip. We can't say with confidence which way they will go. They may make a 260mm^2 chip and launch it at $299 like they did with HD4870 or they can make a 400mm^2 chip and launch it at $399-$499, etc.

" I'm simply stating it's just as valid to compare the 4870 to the 6870 as it is to the 5870. You get different viewpoints doing both methods. One gives you an idea of the jump in performance using the same die size (important), and the other gives you an idea of the jump in performance when implementing new features...

Yes, I understand your point. But HD6870 is not the same architecture as HD6970 is. So you can't assume that because HD6870 improved performance by 50% from HD4870, that HD7970 will "only" improve performance from HD6970 by 50%. HD7970 is likely going to be an extension of 6970 series. However, since AMD packed 22% more transistors into 16% more die space while improving tessellation performance 1.5-3x, the transistor density at 28nm may allow HD7970 series to double the number of SPs over HD6970, and improve tessellation performance further with a minimal die size increase. I think we will see a more than 50% performance improvement.

 

[Silverforce11]

It would be quite ballsy if AMD actually went into a new node with a huge performance winning GPU. Unlikely given their past strategy. However, nobody saw the 375W TDP 6990 coming.

 

[RussianSensation]

It would be quite ballsy if AMD actually went into a new node with a huge performance winning GPU.

9700 Pro put the lights out on both GeForce 4600 and GeForce 5800 series.

Look at the GTX580 vs. 480. It's only about 15-18% faster. NV has been waiting for 28nm to unleash the 'beast' with Kepler. AMD's GPU is going to have to be at least 50% faster than HD6970 or it will be left in the dust. Also we haven't had a sizeable performance increase between generations since September 2009 (HD5870). It's coming (unless NV and AMD think another 15% performance boost is sufficient for console ports :sneaky.

 

[Grooveriding]

Also we haven't had a sizeable performance increase between generations since September 2009 (HD5870). It's coming (unless NV and AMD think another 15% performance boost is sufficient for console ports :sneaky.

With the next gen 28nm cards there will likely be no reason to have anything more than the flagship single-gpu from either company for gaming at 1920x1200 and below.

With the state of pc-gaming these days, there is nothing you can't run at that resolution maxed out, on a dual-gpu setup of a couple 6970s, 580s or a 6990.

 

[Dark Shroud]

With the state of pc-gaming these days, there is nothing you can't run at that resolution maxed out, on a dual-gpu setup of a couple 6970s, 580s or a 6990.

All PC gamers are supposed to be using a triple monitor set-up with 3D enabled.

/s

 

[Bateluer]

Man, I'm still rocking the 4870.

 

[RussianSensation]

Man, I'm still rocking the 4870.

Still a respectable card. All things considered, the GTX280 didn't age well despite its $600+ price tag at launch. $300 HD4870 proved to be a much wiser choice. The 4870 is about 2x slower than the HD6970. I still think you can hold out until 28nm GPUs if you waited this long.

 

[cusideabelincoln]

Yes, I understand your point. But HD6870 is not the same architecture as HD6970 is. So you can't assume that because HD6870 improved performance by 50% from HD4870, that HD7970 will "only" improve performance from HD6970 by 50%. HD7970 is likely going to be an extension of 6970 series. However, since AMD packed 22% more transistors into 16% more die space while improving tessellation performance 1.5-3x, the transistor density at 28nm may allow HD7970 series to double the number of SPs over HD6970, and improve tessellation performance further with a minimal die size increase. I think we will see a more than 50% performance improvement.

I don't really know why you keep saying 6870 is not the same arch as the 6970. It doesn't really have anything to do with what I was saying and doesn't do anything to your argument. If anything you should be making your argument illustrating the architectural differences between the 6870 and 4870. So I'll just counter myself and do the argument for you...

Because it's actually more complicated to compare the 4870 to the 5870/6870; part of the die size is dedicated to the tessellator. So there is simply more than performance to consider; there is also a feature implementation that the 4870 can't do. You can't really compare performance if one card can't run at all, like the 4870 can't run in DX11. There is no apples to apples comparison when you do that, but it does give you an idea what happens when you add more features.

I guess we just need to look at the performance from the 4870 to the 5870/6870, subtract the die size of the tessellator, and get a rough estimation of what we can expect when we move to a new node with a new chip. If we look at the 5870, I think it's safe to say it will still be bigger. The 6870 would certainly be a bit smaller. I don't think the difference will be enough to change the overall conclusions we're trying to draw here.

I never said the jump from Cayman would only be 50%. I expect it to be higher than that. But I do not expect 100% or more jump. So, somewhere in between is where I'd expect the performance to end up.

If Nvidia is going to create a bigger die, either good luck to them (Fermi...) or AMD could be forced to catchup (homerun).

Also I don't think trying to come up with a ratio of transistor increase to die size increase really means anything. It's frivolous data that has been done before. 3870 to 4870: +43% transistor count, +36% die size. HD4870 to 5870: +124% transistor count, +28% die size. Obviously the transistor density is going to increase with this node shrink. It probably won't be anything we haven't seen before, though.

 

[AtenRa]

Both 48xx and 58xx have one Tessellation unit. The Evergreen design (HD58xx) is actually a double 48xx with a pair of 10 SIMDs linked together with a cross bar (plus 6 display controllers for Eyefinity). They are both using the 5-Way VLIW design SPs and they both have FP64 support.

Barts architecture (68xx) doesn’t support FP64 and it uses the HD57xx memory controller and a tweaked Tessellation unit with dual dispatch processors for better ILP.

Cayman (HD69xx) is a multicore design, with dual graphics engines (Dual Tess units etc) and a new 4-Way VLIW SP architecture.

HD79xx will have to be compared against HD58xx and HD69xx and not HD68xx.

 

[cusideabelincoln]

Both 48xx and 58xx have one Tessellation unit. The Evergreen design (HD58xx) is actually a double 48xx with a pair of 10 SIMDs linked together with a cross bar (plus 6 display controllers for Eyefinity). They are both using the 5-Way VLIW design SPs and they both have FP64 support.

Barts architecture (68xx) doesn’t support FP64 and it uses the HD57xx memory controller and a tweaked Tessellation unit with dual dispatch processors for better ILP.

Cayman (HD69xx) is a multicore design, with dual graphics engines (Dual Tess units etc) and a new 4-Way VLIW SP architecture.

HD79xx will have to be compared against HD58xx and HD69xx and not HD68xx.

Those things don't take away the intrigue, theory, thought-process, and contrast/compare by looking at R770/R790 in relation to Barts, because of the fact they have the same die size. Those things AMD did to Cypress is how they got the Barts die size down to the same as R770. Which is the entire point that is trying to be made. The point is not how the 7900 will compare to Barts.

 

[Silverforce11]

There is an imbalance in Cayman's architecture. It's SP:ROP:Setup ratio is not optimal. 6950 running at 6970 core/mem speed is ~same performance even though its got a lot less shaders. This suggests the architecture can't really keep those shaders running optimally.

Where do you guys think the bottleneck is in Cayman? Perhaps it's just a driver issue? hmm.

If AMD works on this for the 79xx series they could boost efficiency further, on top of doubling transistor density.

Edit: It's actually a very similar problem they had with the 5850 vs 5870.

 

[tviceman]

There is an imbalance in Cayman's architecture. It's SP:ROP:Setup ratio is not optimal. 6950 running at 6970 core/mem speed is ~same performance even though its got a lot less shaders. This suggests the architecture can't really keep those shaders running optimally.

Where do you guys think the bottleneck is in Cayman? Perhaps it's just a driver issue? hmm.

If AMD works on this for the 79xx series they could boost efficiency further, on top of doubling transistor density.

Edit: It's actually a very similar problem they had with the 5850 vs 5870.

It may be an intentional (or unintentional depending on how it's looked at) draw back of designing "on the safe side." Nvidia obviously shoots for the moon with their designs, with their draw back being bigger dies. AMD shoots for efficiency per mm^2, which they may determine what they're delivering is the best ratio for their highest end chip. In other words, they likely know the bottleneck exists but to correct it would add more die space than they deem necessary.

 

[Lepton87]

There is an imbalance in Cayman's architecture. It's SP:ROP:Setup ratio is not optimal. 6950 running at 6970 core/mem speed is ~same performance even though its got a lot less shaders. This suggests the architecture can't really keep those shaders running optimally.

Where do you guys think the bottleneck is in Cayman? Perhaps it's just a driver issue? hmm.

If AMD works on this for the 79xx series they could boost efficiency further, on top of doubling transistor density.

Edit: It's actually a very similar problem they had with the 5850 vs 5870.

I think that Cayman and Cypress are bottlenecked by ROPs units, simply not enough of them. Just compare 5830 to 6870. The same number of everything except for ROPs and huge performance difference.

 

[Silverforce11]

I think you're right. ROP count is identical in 6950 vs 6970 (69xx ROPs are beefier) and in 6870 vs 5850/70.

 

[tigersty1e]

If you compare HD4870: 115 GB/sec vs. HD5770: 77 GB/sec or even better HD4890: 125 GB/sec vs. HD5850: 128 GB/sec bandwith, it becomes very pretty clear that almost all of the performance gains from HD4xxx to HD5xxx generation were not related to memory bandwidth improvements.

There is also no indication that current top cards are memory bandwidth limited. HD6970 has more than enough memory bandwidth. Certainly when I run my GPU at HD6950's memory bandwidth or HD6970's memory bandwidth, the performance difference is very negligible. So I don't think it's critical to worry about memory bandwidth yet (they can always put faster GDDR5 chips to get another 10-15% memory bandwidth increase).

Also, HD2900XT series had 512-bit of memory bandwidth. So AMD doesn't necessarily need to stick to 256-bit memory bus.



That's probably a reasonable assumption, but not guaranteed based on history.

X1800XTX = 264 mm^2
X1950XTX = 315 mm^2
HD2900 XT = 420 mm^2

There is actually a very clear trend that AMD GPUs have grown in die size at a rate of ~15-30% per generation in the last 2-3 years. While their die sizes are smaller compared to NV's chips latest, you can't say that AMD's philosophy has been to decrease the die size:

More recently
HD3870 = 192 mm^2
HD4870 = 260 mm^2 (+35%)
HD5870 = 334 mm^2 (+28%)
HD6970 = 389 mm^2 (+16%)

It sounds reasonable that AMD may try to shrink the die size since they are moving to a 28nm process. At the same time, since they have been successful in the last 3+ years when they have increased the die size, they could try for a 28nm 400 mm^2 monster chip. We can't say with confidence which way they will go. They may make a 260mm^2 chip and launch it at $299 like they did with HD4870 or they can make a 400mm^2 chip and launch it at $399-$499, etc.



Yes, I understand your point. But HD6870 is not the same architecture as HD6970 is. So you can't assume that because HD6870 improved performance by 50% from HD4870, that HD7970 will "only" improve performance from HD6970 by 50%. HD7970 is likely going to be an extension of 6970 series. However, since AMD packed 22% more transistors into 16% more die space while improving tessellation performance 1.5-3x, the transistor density at 28nm may allow HD7970 series to double the number of SPs over HD6970, and improve tessellation performance further with a minimal die size increase. I think we will see a more than 50% performance improvement.

the gains from 4870 to 5870 were largely contributed from memory gains. And why are you comparing 4870 to 5770 and 4890 to 5850? 4890 to 5870 is more appropiate. Take your 5870 and downclock it to 900 or 975 memory and your speed will drop.

Memory at this point is limited by the controller. they can't just put faster chips in to get faster speeds. And 512 bit bus? lol. higher bus is not always better. once we strain the actual memory chips, higher bus will help.

the gains from 4890/4870 to 5870 included the perfect conditions with die shrink and memory increases. This gen, increases will ONLY come from the core. I will be surprised if they increase memory more than 10%.

There's a reason why amd is using 6 gbps memory chips and only clocking them to 5.5 gbps. the controller is at its limit.

 

[RussianSensation]

the gains from 4870 to 5870 were largely contributed from memory gains. And why are you comparing 4870 to 5770 and 4890 to 5850? 4890 to 5870 is more appropiate.

Can you provide support for that? HD5770 has smaller memory bandwidth than HD4870 and they have almost the same performance. HD4890 has almost identical memory bandwidth to the HD5850 but is much slower than the latter. The the majority of the gains from HD4870 to HD5870 cannot possibly have come from memory bandwidth. Most of that came from doubling the number of SPs, TMUs and ROPs. This is why I used HD4870 vs. HD5770 and HD4890 vs. HD5850 comparison to illustrate this point.

Memory at this point is limited by the controller. they can't just put faster chips in to get faster speeds.

Of course they can. Faster GDDR5 memory is going to be available. Also, you are assuming that:

1) Cayman needs more memory bandwidth
2) The memory controller is limited/cannot be improved further.

And 512 bit bus? lol. higher bus is not always better. once we strain the actual memory chips, higher bus will help.

Who said 512-bit bus is necessarily better? I said it's an option which AMD has utilized before, but I didn't say it's more cost effective or "better".

[qoute] This gen, increases will ONLY come from the core. I will be surprised if they increase memory more than 10%. [/quote]

They don't need to increase memory bandwidth a lot, esp. since HD6970 series already has enough of it. They will increase the number of functioning units inside the GPU and possibly its clock speed.

 

[3DVagabond]

There's a reason why amd is using 6 gbps memory chips and only clocking them to 5.5 gbps. the controller is at its limit.

I think it's purely power usage and heat. Possibly binning to give some wiggle room for sub par components. I've seen some custom cooled designs go over 6Gb/s. (Lightning, Direct CUII, Powercolor LCS) So I don't think it's the memory controller not being up to the task.

 

[tigersty1e]

Can you provide support for that? HD5770 has smaller memory bandwidth than HD4870 and they have almost the same performance. HD4890 has almost identical memory bandwidth to the HD5850 but is much slower than the latter. The the majority of the gains from HD4870 to HD5870 cannot possibly have come from memory bandwidth. Most of that came from doubling the number of SPs, TMUs and ROPs. This is why I used HD4870 vs. HD5770 and HD4890 vs. HD5850 comparison to illustrate this point.



Of course they can. Faster GDDR5 memory is going to be available. Also, you are assuming that:

1) Cayman needs more memory bandwidth
2) The memory controller is limited/cannot be improved further.



Who said 512-bit bus is necessarily better? I said it's an option which AMD has utilized before, but I didn't say it's more cost effective or "better".

[qoute] This gen, increases will ONLY come from the core. I will be surprised if they increase memory more than 10%.

They don't need to increase memory bandwidth a lot, esp. since HD6970 series already has enough of it. They will increase the number of functioning units inside the GPU and possibly its clock speed.[/QUOTE]

by any chance are you crazy russian @ ncix?

the jump from 4890 to 5850 is good, but not omgwtfbbq good. (the 5850 and 5870 perform within ~5% with each other, so the jump from 5850 to 5870 is in large part due to the clocks). it is when you take a step back and see the 4890 to 5870 jump, you look at how awesome the card is.

for the next jump to 7 series, expect a 4890 to 5850 type jump and not a 4890 to 5870 jump because memory clocks are already squeezed to the max. as for core clocks, that is give and take. if they increase number of shaders etc. clocks won't increase much.

let's not forget that they will have to pack on a pcie 3.0 controller on the die. as to how much bigger it is is anyone's guess.

 

[tigersty1e]

I think it's purely power usage and heat. Possibly binning to give some wiggle room for sub par components. I've seen some custom cooled designs go over 6Gb/s. (Lightning, Direct CUII, Powercolor LCS) So I don't think it's the memory controller not being up to the task.

http://www.anandtech.com/show/3987/...renewing-competition-in-the-midrange-market/2

Along with selectively reducing functional blocks from Cypress and removing FP64 support, AMD made one other major change to improve efficiency for Barts: they’re using Redwood’s memory controller. In the past we’ve talked about the inherent complexities of driving GDDR5 at high speeds, but until now we’ve never known just how complex it is. It turns out that Cypress’s memory controller is nearly twice as big as Redwood’s! By reducing their desired memory speeds from 4.8GHz to 4.2GHz, AMD was able to reduce the size of their memory controller by nearly 50%. Admittedly we don’t know just how much space this design choice saved AMD, but from our discussions with them it’s clearly significant. And it also perfectly highlights just how hard it is to drive GDDR5 at 5GHz and beyond, and why both AMD and NVIDIA cited their memory controllers as some of their biggest issues when bringing up Cypress and GF100 respectively.

 

[RussianSensation]

by any chance are you crazy russian @ ncix?

No, I am not actually. Must be another crazy russian.

for the next jump to 7 series, expect a 4890 to 5850 type jump and not a 4890 to 5870 jump

Not sure how they will be able to compete with Kepler then. HD5850 is only about 30-35% faster than the HD4890. Since the GTX580 is about 15% faster than the HD6970, if HD7970 is only 35% faster than the 6970, it will be only 17% faster than the GTX580. Kepler is going to blow away the 580. So AMD has to be aiming for a much larger performance increase to stay in the game imo.

 

[busydude]

Is kepler just a die shrink, or a totally new architecture?

 

[Skurge]

They don't need to increase memory bandwidth a lot, esp. since HD6970 series already has enough of it. They will increase the number of functioning units inside the GPU and possibly its clock speed.

by any chance are you crazy russian @ ncix?

the jump from 4890 to 5850 is good, but not omgwtfbbq good. (the 5850 and 5870 perform within ~5% with each other, so the jump from 5850 to 5870 is in large part due to the clocks). it is when you take a step back and see the 4890 to 5870 jump, you look at how awesome the card is.

for the next jump to 7 series, expect a 4890 to 5850 type jump and not a 4890 to 5870 jump because memory clocks are already squeezed to the max. as for core clocks, that is give and take. if they increase number of shaders etc. clocks won't increase much.

let's not forget that they will have to pack on a pcie 3.0 controller on the die. as to how much bigger it is is anyone's guess.

Just so we are clear here.

6970 = 176GB/s
6950 = 160GB/s
6870 = 134GB/s
5870 = 153GB/s
5850 = 128GB/s

4890 = 124GB/s
4870 = 115GB/s
5770 = 76GB/s

Now look at the 5770 vs the 4870/90. The is a huge gap in bandwidth between the cards, but performance is very similar. Same SPs, very similar clocks. So how do you explain that.

Looking at the 6950, 5870 and 6870. The 6870 is closer to the 5870 in performance than the 5870 is to the 6950, but bandwidth favors the 5870 heavily against the 6870. How do you explain that too?

AMD have been using GDDR5 since 2008 with the 4870, I'm pretty sure they can get their memory controller working with higher speeds when the 7000 series is released. Who knows what kind of speeds the controller in Cayman can handle as it can easily max out the 6ghz of the GDDR5 that the 6970/50 come with.

 

[3DVagabond]

http://www.anandtech.com/show/3987/...renewing-competition-in-the-midrange-market/2

Sorry I wasn't clear. I was referring to Cayman, not Barts. Although Barts does make for the argument that 6900 cards aren't bandwidth limited. Theoretically, we should be able to get >2x 6870 performance with 256 mem bus and 6Gb/s RAM. That alone should be adequate for the next gen parts. Unless we think we're going to see more performance than that from a single GPU with HD7000?