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

[-Slacker-]

Will the jump to 28nm mean less power consumption and temps? My knowledge on gpus sin't all that good.

most likely only for cards at similar speed brackets, eg. 6870 to 7770 or something like that.

 

[Borealis7]

power consumption doesn't necessarily go down. because 28nm allows less voltage (and in the end less power consumption) the amount of transistors is raised until a certain TDP is reached.
cards will still need 6+6 or 8+6 power connectors and will still have 150W or even 300W TDP but they'll be stronger than previous cards with the same TDP.
what might go down is idle power which we've seen drop significantly with the 5800 series due to power gating and process shrinking.

 

[AtenRa]

AMD is reportedly already preparing its next series of Radeon HD 7000-series GPUs, and planning to start mass production of the new chips in May this year...
http://www.bit-tech.net/news/hardware/2011/04/15/radeon-hd-7000-due-june-or-july/1

My guess on the specs

Going by the 100% jump in performance from HD4870(55nm) to HD5870(40nm) i would say the new HD7970(28nm) will be around 100% Faster then HD6970(40nm) and slightly faster than two of them in crossfire

Released befor the Nvidia GTX680

Stream Processors 3048
Texture Units 128
ROPs 64
Core Clock 1020MHz
Memory Clock 1.7GHz (6.8GHz effective) GDDR5
Memory Bus Width 256-bit
Frame Buffer 2GB
FP64 1/4
Transistor Count 3.2B
Manufacturing Process TSMC 28nm
1x6,1x8 pin power, TDP 250W
Price Point $369US

How did you come to the 128 Texture Units and 3048 SPs ???

AMD Cayman has 1536 SPs and 96 Tex Units (24 SIMDs x 4 Tex), if they will double that then 7970 will have 3072 SPs and 192 Tex Units.

 

[RussianSensation]

100% jump from 55nm to 40nm? Don't remember that....

HD5870 (40nm) can be about 80-100% faster than an HD4870 (55nm) in the modern games actually.

Mafia 2
Civ 5

Either way the boost on 28nm should be substantial after the 'refresh performance' of GTX5xx and HD69xx lines.

 

[Lonyo]

HD5870 (40nm) can be about 80-100% faster than an HD4870 (55nm) in the modern games actually.

Mafia 2
Civ 5

Either way the boost on 28nm should be substantial after the 'refresh performance' of GTX5xx and HD69xx lines.

It's also a lot bigger (256mm^2 vs 334mm^2).
Not hard to make something 100% faster when you make it bigger as well, and on a smaller process.
I don't see AMD making a potentially 510mm^2 die for their top end chip to double the speed of the HD6970 (389mm^2).

If you look at the improvement from the HD4870 to the HD6870, that's a better measure of how much a process can help, since that's 256mm^2 vs 255mm^2, pretty much exactly the same die size, and it's not a doubling in speed.

http://www.anandtech.com/bench/Product/304?vs=290

50%+ improvement across the board, which is still pretty decent, but it doesn't get all that close to 100%.

 

[AtenRa]

Since AMD will discuss the new GPUs in Developer Summit at 13-16th of June, i don't expect anything before that time.

http://developer.amd.com/afds/pages/keynote.aspx

 

[themodernlife]

Make no Mistake Both Companies Will prepare something for the BF3 PC release. Absolutely no question about it. So Sept-October looks very much more likely.

 

[badb0y]

Unless AMD makes something compelling, I am going with nVidia next round. I like to switch things up every generation.

 

[exar333]

My 5870 is good until we have some seriously powerful options on 28nm (NV or AMD). The AMD options looks awfully good.

 

[busydude]

If you look at the improvement from the HD4870 to the HD6870, that's a better measure of how much a process can help, since that's 256mm^2 vs 255mm^2, pretty much exactly the same die size, and it's not a doubling in speed.

This is a valid comparison IMO.. increasing performance/die space makes sense.

 

[Acanthus]

Haha we will not be seeing a 100% performance increase.

40nm to 28nm is a 100% increase in transistor density.

For a chip of the same size, and clockspeed increases as well, it is easy to imagine a 100% increase in performance.

You can be limited in other areas though like memory bandwidth or architectural bottlenecks.

Also May production does not equal a May launch. It usually takes a 3-4 month production ramp before a hard launch.

 

[RussianSensation]

It's also a lot bigger (256mm^2 vs 334mm^2).
Not hard to make something 100% faster when you make it bigger as well, and on a smaller process.
I don't see AMD making a potentially 510mm^2 die for their top end chip to double the speed of the HD6970 (389mm^2).

I am not saying that HD7970 will be 100% faster than HD6970, but there is a strong possibility that AMD left out some innovations and improvements on 40nm which they will likely incorporate into their 28nm GPU designs. There could still be some "hidden" improvements to their newest architecture. Keep in mind Kepler is going to be a Fermi II architectural design. NV has outlined that Kepler will be significantly faster than Fermi. So AMD can't just release GPUs with a mere 20% performance increase or they won't stand a chance against a new architecture from NV.

This is a valid comparison IMO.. increasing performance/die space makes sense.

But you shouldn't be comparing a high end card of the previous generation (i.e., HD4870 was 55nm) to a mid-range card of this generation (i.e., HD6870 on 40nm). 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. 55nm to 40nm is also only half a node decrease in size. Going from 40nm to 28nm is a full node. So the performance difference compared to the HD5870 should be substantial at 2560x1600 (for lower resolutions, there is practically no reason for anything faster than an HD6970 imo).

Also, HD6870 has 2x the ROPs, 16 more TMUs, and 320 more SPs compared to the 4870, while being clocked 20% faster (and AMD achieved that on just half a node!). Remember, those ROPs are 2-4x slower than they are in the HD6970 architecture, depending on the scenario. Take a 28nm HD6970 chip, double its ROPs and shaders, increase clock speeds to 1000mhz and that thing will smoke a 40nm 6970.

 

[RussianSensation]

Sometimes we forget how quickly videocard performance moves:

HD4890 (55nm) - April 1, 2009
SPs = 800
GFLOPS = 1360
Pixel Fill-rate = 13.6 GTexels/sec
Texture Fill-rate = 34.0 GTexels/sec

HD6970 (40nm) - December 15, 2010
SPs = 1536
GFLOPS = 2703
Pixel Fill-rate = 28.2 GTexels/sec
Texture Fill-rate = 84.5 GTexels/sec



The performance difference between the HD5870 and the HD4890 in some games is staggering; and the 5870 was released around September 23, 2009.

 

[Lonyo]

I am not saying that HD7970 will be 100% faster than HD6970, but there is a strong possibility that AMD left out some innovations and improvements on 40nm which they will likely incorporate into their 28nm GPU designs. There could still be some "hidden" improvements to their newest architecture. Keep in mind Kepler is going to be a Fermi II architectural design. NV has outlined that Kepler will be significantly faster than Fermi. So AMD can't just release GPUs with a mere 20% performance increase or they won't stand a chance against a new architecture from NV.



But you shouldn't be comparing a high end card of the previous generation (i.e., HD4870 was 55nm) to a mid-range card of this generation (i.e., HD6870 on 40nm). 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. 55nm to 40nm is also only half a node decrease in size. Going from 40nm to 28nm is a full node. So the performance difference compared to the HD5870 should be substantial at 2560x1600 (for lower resolutions, there is practically no reason for anything faster than an HD6970 imo).

Also, HD6870 has 2x the ROPs, 16 more TMUs, and 320 more SPs compared to the 4870, while being clocked 20% faster (and AMD achieved that on just half a node!). Remember, those ROPs are 2-4x slower than they are in the HD6970 architecture, depending on the scenario. Take a 28nm HD6970 chip, double its ROPs and shaders, increase clock speeds to 1000mhz and that thing will smoke a 40nm 6970.

"Full" nodes:
65 -> 45 -> 32 -> 22
"Half" nodes:
55 -> 40 -> 28 -> 20

55 to 40 is not a half to full node step, it's a full jump, from a half to a half, skipping the "full" node in between, just like 40 to 28.

Also, comparing high end to high end doesn't have any relevance if you are also changing the die size, like I said initially.
If I make a 100mm^2 55nm die, then I can make a 40nm product which is at least 5x as fast if I make it on a 250+mm^2 die. (Assuming a linear improvement with die size and twice as many transistors per mm^2)
Doesn't mean that when I step down to 28nm I can get another 5x improvement because then I'd be making a die which is 625mm^2.

Just like saying "they got a 100% improvement from 55nm to 40nm" doesn't make sense since a significant amount of the improvement likely came from simply making a bigger die.
If they had made a 33% bigger die on 55nm, a 389mm^2 product, then it would have been faster than the 256mm^2 HD4870. Not a hard thing to do when you are working with more space.

Now fine, they could increase the die size going from 40nm -> 28nm, but that's got nothing to do with how much they are getting from the process, that's just what they are getting from having a bigger die, which is a meaningless piece of information if you can't or don't increase the die size.
Since AMD are already close to the 400mm^2 mark, unless they want to "do an NV" and go for a 500mm^2+ die, they aren't going to get performance improvements from a larger die. A sensible guess would be that they would aim for a similar size die to this generation, meaning a die size vs die size comparison makes a lot more sense than a high end vs high end comparison where die size has changed by 33%.

 

[Grooveriding]

Just out of curiosity. I read something awhile back about the technology getting close to its limits in terms of how small it could get. I believe it was something along the lines of 1Xnm would be the smallest.

My question is that seems to only be 2 years off. So what then ?

 

[bryanW1995]

I think the 7970 will be as fast as 2 6950's in crossfire and have 2816 sp's. I believe it will include some cpu/gpu like features kinda like the fermi chips and with better tessalators. Core clocks around 850.

I'm suprised that nobody has called out the OP on the "5870 is more than twice as fast as 4870" comment. 4870x2 was ~ 10% faster than 5870 at launch, and in non-dx11 games is still faster to this day. 5870 was actually somewhat of a disappointment at launch imho, it was just quickly overshadowed by things like "cardboardgate" and 300w single gpu cards so it ended up being highly successful. If this new theoretical 7970 really has double the sp's, rop's, etc etc, is twice as fast as 6970, and runs at 1020 mhz then it will likely cost a lot more than $369 at launch.

Just out of curiosity. I read something awhile back about the technology getting close to its limits in terms of how small it could get. I believe it was something along the lines of 1Xnm would be the smallest.

My question is that seems to only be 2 years off. So what then ?

Don't underestimate TSMC's ability to disappoint.

 

[AtenRa]

65nm to 55nm = Half Node
55nm to 40nm = Half Node
40nm to 32nm = Half Node
40nm to 28nm = Full Node

With a Full node we will have double the transistor count at the same die size of the previous node. That means we can double the HD6970 and keep the same die size.

HD5870 has double the transistor count and larger die size of the HD4890 because we had a half node shrink from 55nm to 40nm.

 

[tigersty1e]

it depends on what extra they pack into the die. there were some features left out of the 69XX series because of space reasons.

but we won't see a near ~100% improvement in performance. looking back at 4870 to 5870 jump is not right because 4870 had amd's first ddr5 mem controller. 5870 jumped mem speeds from 900 to 1200. 33% improvement plus other timing enhancements. amd is sort of at the limit with how much speed they can squeeze out of the controller.

the other question is will amd have this card ready for pcie 3.0. yes, the card will probably not need the extra bandwidth, but nvidia's next gen cards will most likely come with pcie 3.0. the masses won't want a card with "old" technology next year when amd and nvidia both have their cards out.

triple monitor support through dvi only?
more control over poertune?

 

[3DVagabond]

Temperature has nothing to do with process. Having a smaller process, bigger process, more, fewer transistors, higher, lower clockspeeds don't matter when it comes to temperatures. The heatsink and fan is the main determining factor.

In terms of power, it means less power use at a set level of performance.
That doesn't mean the top end GPU uses less power, it means it uses an amount of power which could be higher, lower or equal to the previous top GPU, but it runs faster.

28nm = more transistors in a given die space, more performance at a given price point.

When asking if it will offer lower power and temps, the answer is yes. Take a gtx-580 and shrink it to 28nm, all else being equal, it will use less power, and therefor, be cooler. As you say, if you were to make a chip the same size as a 580 on 28nm, then it will offer more performance for the same price, assuming same yields and nothing else changed. It will likely not use less power, nor produce less heat, simply due to the shrink.

Unless AMD makes something compelling, I am going with nVidia next round. I like to switch things up every generation.

We're talking video cards, not women.

 

[busydude]

Don't underestimate TSMC's ability to disappoint.

Is GF competing with TSMC at 28nm this time? There were reports that AMD are going to go with either TSMC or GF in the future..

 

[ViviTheMage]

I think the 7970 will be as fast as 2 6950's in crossfire and have 2816 sp's. I believe it will include some cpu/gpu like features kinda like the fermi chips and with better tessalators. Core clocks around 850.

If this is truth, I will be upgrading my xfire'd 6950's!!

 

[bryanW1995]

65nm to 55nm = Half Node
55nm to 40nm = Half Node
40nm to 32nm = Half Node
40nm to 28nm = Full Node

With a Full node we will have double the transistor count at the same die size of the previous node. That means we can double the HD6970 and keep the same die size.

HD5870 has double the transistor count and larger die size of the HD4890 because we had a half node shrink from 55nm to 40nm.

90, 65, 45, 32 are full nodes. At least in recent memory, you've always seen cpus on these nodes. 80, 55, 40, and 28 are 1/2 nodes. So you're correct in everything but the 55-40nm jump b/c that is actually a full node. Apparently the new normal for gpus is to keep to 1/2 nodes. ATI was at 80 and 55 while NV was at 90 and 65, respectively, but nvidia finally caught up (7 mos late) at 40 nm. Now that they've bitten the bullet they should theoretically have made up the process gap, though this will in large part depend upon tsmc's and, later, GF's ability to deliver.


doh, ninja'd by lonyo!

 

[AtenRa]

90, 65, 45, 32 are full nodes. At least in recent memory, you've always seen cpus on these nodes. 80, 55, 40, and 28 are 1/2 nodes. Apparently the new normal for gpus is to keep to 1/2 nodes. ATI was at 80 and 55 while NV was at 90 and 65, respectively, but nvidia finally caught up (7 mos late) at 40 nm. Now that they've bitten the bullet they should theoretically have made up the process gap, though this will in large part depend upon tsmc's and, later, GF's ability to deliver.

Im talking from one process to the next, from 55nm to 40nm is a half node and from 40nm to 28nm is a Full Node.

A full node shrink will allow you to ~double the transistor count and ~keep the same die size in relationship to the previous process.

GT200 at 65nm = 1.4B transistors and 575mm2
GT200B at 55nm = 1.4B transistors and 470mm2 (Half Node from 65nm)
GF100 at 40nm = 3.2B transistors and 530mm2 (Full Node from 65nm) double the transistor count same die size from 65nm. (40nm is a little bit more than full node from 65nm)

 

[RussianSensation]

"Half" nodes:
55 -> 40 -> 28 -> 20

"The 32-nm process is a cost-down version of its 40-nm technology, while 28-nm is considered by TSMC as a ''full-node'' offering." - EETimes

A sensible guess would be that they would aim for a similar size die to this generation, meaning a die size vs die size comparison makes a lot more sense than a high end vs high end comparison where die size has changed by 33%.

It does if you are comparing apple-to-apples GPU architectures. But you are not, since HD6870 is a Barts architecture which is more or less the same achitectural design as Cypress (meaning VLIW-5), while HD6970 is a new architecture with VLIW-4 and 2x faster ROPs. Therefore, we would need a 256mm^2 HD6970 to compare to the HD4870 using your logic. But such card is not available. Even if AMD retains the same die size, by going to 28nm, they will be able to add more ROPs, TMUs and SPs. I don't understand why you keep coming back to AMD having to increase die size. If you think HD7000 series is a mild-refresh performance wise, how do you expect them to compete with Kepler exactly?

 

[bryanW1995]

Im talking from one process to the next, from 55nm to 40nm is a half node and from 40nm to 28nm is a Full Node.

A full node shrink will allow you to ~double the transistor count and ~keep the same die size in relationship to the previous process.

GT200 at 65nm = 1.4B transistors and 575mm2
GT200B at 55nm = 1.4B transistors and 470mm2 (Half Node from 65nm)
GF100 at 40nm = 3.2B transistors and 530mm2 (Full Node from 65nm) double the transistor count same die size from 65nm. (40nm is a little bit more than full node from 65nm)

You are flat out wrong. 55 to 45 is a half node, 55 to 40 is a full node. 65 to 40 is 1.5 nodes. GPU's skipped 45 and 32 b/c nvidia wanted to catch up to AMD's 1/2 node advantage, so now they're all on a half node - half node upgrade schedule. The last time anybody pulled a 1/2 node upgrade was when nvidia went from 65 to 55nm, with for example the gtx 260 192 to gtx 260 core 216.

"The 32-nm process is a cost-down version of its 40-nm technology, while 28-nm is considered by TSMC as a ''full-node'' offering." - EETimes

You are absolutely correct, but a few people have been confused by this. For TSMC, the "new normal" is to always go in 1/2 nodes. If ATI hadn't pushed this, we would have gone from 90-65-45-32 because their only competition, nvidia, was already on that schedule. When nvidia saw the advantage that DAAMIT was able to enjoy by using 80/55/40/28/etc, they pushed extremely hard to get on that schedule as well. Now that TSMC has skipped 45nm and 32nm, their new normal architecture is 55/40/28/20/etc so for them 28nm is a full node offering.

Kinda depressing that the article is 2 1/2 years old. TSMC is really on the ball lately, huh?

Maybe IDC will see this and link us to his "Pirates" theory on process nodes, that was a good read.