DT: Next-generation 28nm GPUs Could Be 45 Percent Faster

[dmoney1980]

Sources claim TSMC's manufacturing is that good

Taiwan Semiconductor Manufacturing Company (TSMC) is the world's largest semiconductor foundry, and as such is constantly under pressure from its customers and competitors. The company recently announced that its newest 28nm process has entered mass production. Smaller process geometries typically mean that a faster, smaller, and cheaper processor will be available.

Two of TSMC's biggest customers are AMD and NVIDIA. Both have taped out GPU designs using TSMC's 28nm High Performance (28HP) process. AMD's next-generation GPU series is codenamed Southern Islands. The Tahiti GPU is supposed to launch early in December, while NVIDIA's Kepler GPU will launch in February of 2012.

28HP is the first process from TSMC to use High-k Metal Gate (HKMG) technology, as opposed to the typical silicon oxynitride (SiON) found in 40nm GPUs. HKMG uses a material with a high dielectric constant instead of the traditional silicon dioxide gate dielectric. This allows for a substantial reduction in gate leakage, thus lowering overall power consumption and allowing for higher clock speeds.

According to sources within TSMC, the 28HP HKMG process is doing very well. So well, in fact, that it supports up to a 45 percent speed improvement over the firm's own 40G process used to make the last two generations of video cards. This speed improvement is based on the same leakage per gate, but the GPU firms may choose to favor lower power consumption over a pure speed boost.

Our AMD contacts declined to respond to these assertions and directed our attention to a presentation made in June at the AMD Fusion Developer Summit by Eric Demers, the Chief Technology Officer of AMD Graphics.

The flip side of the manufacturing process is the architecture and design, and AMD has already previewed the basis for its future graphics architectures, known as Graphics Core Next (GCN). The basic design of GCN will form the foundation for the next few generations of AMD graphics processors.

The fundamental unit of AMD’s previous designs has been the Streaming Processor, utilizing a Very Long Instruction Word (VLIW) architecture designed to take advantage of Instruction Level Parallelism (ILP). That will be supplanted by the Compute Unit, comprised of multiple 16-wide vector Single Instruction, Multiple Data (SIMD) units designed for Thread Level Parallelism (TLP).

L1 and L2 caches will support the CUs, while the GPU will have access to the main system memory. Support for C++ programming has been added, making it easier to program for the GPU and CPU within the same application.

While all of these changes will benefit compute applications, it is not yet clear what impact this will have on current games and those already in the pipeline. The architectural changes may end up helping or hindering performance. Nevertheless, the entire 28nm graphics lineup will support resolutions of up to 16000 x 16000 pixels.

Of course, yields have always been a big problem with introducing a new node. TSMC had significant challenges with its 40nm process, leading to shortages of the Radeon HD 5800 series. HKMG processes typically use atomic layer deposition for high-k materials, which has been a challenge even for Intel.

Ultimately, we know that the next-generation of GPUs will be "significantly" faster, but the final clock speeds are still being determined, so final performance numbers will have to wait until the official launch. The current word on the street is December 6.

source: http://www.dailytech.com/Nextgeneration+28nm+GPUs+Could+Be+45+Percent+Faster/article23158.htm

 

[Red Storm]

Only 45%? Could be?

:|

 

[TakeNoPrisoners]

45% seems a little low. From 4870->5870 was about 100%

 

[Skurge]

45% seems a little low. From 4870->5870 was about 100%

The 45% could just be clock speeds.

45% higher clocks + Double the transistors in the same die area could mean a 7970 being near 6990 performance at the 200W level. If that makes sense.

 

[boxleitnerb]

Hopefully. 45% is nothing for a new generation.
I'm afraid there will be a saturation in performance. If you look at the past, new generations brought an increase of 60-80% in average, but power consumption increased too. So we never had a doubling of perf/w except in special scenarios for each new generation. Now, they are at 250W already, sometimes even exceeding that in real world testing (no furmark). Do they plan to make 300-400w cards now?

 

[Barfo]

Cool story bro.

 

[atticus14]

yeah i think this only takes into the most basic outcomes...how much more they can pack into the same area and whatever increased speeds they can push out.

the +45% is essentially free with the shrink, the other % can be made up with increased die size and new architecture

This doesnt mean companies will take advantage of it though, die shrinks can also mean small improvements for free but greater profits for companies. (same chip, smaller die, more chips per wafer = sell it for old price for as long as possible and get way more profit.)

 

[Termie]

45% seems a little low. From 4870->5870 was about 100%

Actually, it wasn't even close. Despite the common perception that most generations are a 100% jump, they are usually closer to 50%:

HD3870 -> HD4870: http://www.anandtech.com/show/2556/13. 60 percent.

HD4870 -> HD5870: http://www.anandtech.com/show/2841/17. 50 percent.

This is depite the fact that the 5870 had twice the theoretical computational power of the 4870 (in shaders).

The new chip would be just about on target at 45%. Here's a thread on the topic: http://forums.anandtech.com/showthread.php?t=2200297

 

[superccs]

<-- has new GPU boner.

 

[TakeNoPrisoners]

Actually, it wasn't even close. Despite the common perception that most generations are a 100% jump, they are usually closer to 50%:

HD3870 -> HD4870: http://www.anandtech.com/show/2841/17. 60 percent.

HD4870 -> HD5870: http://www.anandtech.com/show/2841/17. 50 percent.

This is depite the fact that the 5870 had twice the theoretical computational power of the 4870 (in shaders).

The new chip would be just about on target at 45%. Here's a thread on the topic: http://forums.anandtech.com/showthread.php?t=2200297

There are quite a few benchmarks that show 2x performance from 4870->5870.
That would make it 100%

 

[Termie]

There are quite a few benchmarks that show 2x performance from 4870->5870.
That would make it 100&#37;

I'd say there are very few, really: http://www.anandtech.com/bench/Product/176?vs=162

The truth is that you only start seeing 100% gains between generations when something separate from the computational power acts as a bottleneck, for instance VRAM. The original 4870-512MB, for instance, is quite crippled in modern games. We're seeing that to some extent now with the 5870 1GB, which may be slightly crippled vs. a 6950 2GB at high resolutions with AA, despite having very similar theoretical power.

Another example of a crippling bottleneck would be the inability to use a new DX version - again, the 4870/4890 (and the GeForce 200 series) are held back in many ways, both in quality and performance, by being limited to DX10 in BF3.

I speak from experience - I have upgraded graphics cards 4 times in the past 7 years. The first two upgrades were between one generation and the next, and I got 100% increases. The last was two generations, and I got about the same increase. While it used to be true that you could gain approximately 100% in actual frames per second (not theoretical power) between generations, I have not seen that hold true since the 7900 to 8800 series on the NVidia side, and the x800 to x1900 on the AMD side (and that was actually two generations, really, since x1800 was in between). AMD hasn't offered 100% gains between generations for quite some time, ever since they abandoned the monolithic approach after failing pretty badly with the x2900xt (and to some extent the x1800).

Look, if we consider the 5800 to 6900 and 480 to 580 generational jumps (they each took about a year, after all), the increases were barely 20%. When the 7000 eventually arrives, it will be 2.5 years after the 5870, and we're now being promised a 45% gain.

Yes, the times they are a' changin'.

 

[happy medium]

In later driver revisions the 5870 was about 100&#37; faster than a 4870x2.

 

[boxleitnerb]

Yeah, that's another thing. Architectures mature. Performance at launch is not final, so to speak.

 

[Termie]

In later driver revisions the 5870 was about 100&#37; faster than a 4870x2.

I think you mean 4870, not 4870x2. The 5870 can be beaten by a 4870x2.

Yeah, that's another thing. Architectures mature. Performance at launch is not final, so to speak.

That's what we think, but if you look at performance in new games, it's not entirely true, until a huge bottleneck is reached.

For instance, I see that here with the 4890 absolutely failing: http://www.anandtech.com/bench/Product/176?vs=162

But it didn't fail here: http://www.techspot.com/review/379-crysis-2-performance/page5.html. In fact, the 5870 is only 35% faster.

 

[SickBeast]

45&#37; faster is quite good for a next generation GPU these days.

A GPU that's 30% faster than a GTX 580 will be nice, especially if they price it well.

 

[Acanthus]

The article implies 45&#37; more clockspeed. It talks about the transistors being 45% faster, not the final product.

This, coupled with doubling the number of transistors, should make these monster GPUs for performance.

I predict 75%-125%+ over the current fastest single GPU.

 

[Ares1214]

Why is everyone complaining, its not like we really even have games that can really test current GPU's. I mean I'm playing BF3 with most of the eye candy on ultra and I'm using a OC'ed 4890, so 45&#37; from the shrink alone and then perhaps another 10-20% with changes to the arch's is not only to be expected, but more than needed at the moment.

 

[Genx87]

I think they are talking about the transistor, not actual designs. Given the same GPU, in theory one could increase transistor density or increase clockspeed by 45%. AMD's design may increase clocks or increase transistors by more than that.

 

[Idontcare]

The article implies 45% more clockspeed. It talks about the transistors being 45% faster, not the final product.

This, coupled with doubling the number of transistors, should make these monster GPUs for performance.

I predict 75%-125%+ over the current fastest single GPU.

That's what I got out of it as well. Purely talking about clockspeeds.

 

[Despoiler]

The context is for TSMC's process. The 45% increase is *clock speed* compared to 40nm not overall performance of the card. We should see top level cards around 1275mhz. Reading comprehension....

 

[Termie]

Why is everyone complaining, its not like we really even have games that can really test current GPU's. I mean I'm playing BF3 with most of the eye candy on ultra and I'm using a OC'ed 4890, so 45&#37; from the shrink alone and then perhaps another 10-20% with changes to the arch's is not only to be expected, but more than needed at the moment.

I'm going to call your bluff on this one. I'll let these charts speak for themselves:

High: http://www.tomshardware.com/reviews/battlefield-3-graphics-performance,3063-8.html
High: http://www.techspot.com/review/458-battlefield-3-performance/page5.html


Ultra: http://www.tomshardware.com/reviews/battlefield-3-graphics-performance,3063-12.html
Ultra: http://www.techspot.com/review/458-battlefield-3-performance/page3.html

While we don't need more GPU power, we've been stagnating since Sept. 2009, so it would be nice to have another generational leap in power.

 

[formulav8]

They specifically mean the Transistors used in GPU's perform 45&#37; faster. Switching speed, ect..

 

[RussianSensation]

"According to sources within TSMC, the 28HP HKMG process is doing very well. So well, in fact, that it supports up to a 45 percent speed improvement over the firm's own 40G process"

^ The sentence is comparing manufacturing processes not GPUs. This reads that 28nm transistors at TSMC can work at 45&#37; faster frequencies as their 40nm predecessors. That doesn't take into account any architectural improvements that AMD will bring.

I would say this time, HD7970 should be faster by a lot more than 45%:

1) Brand new architecture from scratch --> far better DX11 performance, esp. under killer DX11 features such as Tessellation, bokeh DOF. On top of that, scalar architecture makes it far better at geometry computation than the fixed Tessellation engines.

2) Let's say a GTX580 is 15% faster on average than the HD6970. If HD7970 is only 45% faster than HD6970, that would make it only 26% faster than a GTX580 (HD6970 = 100, GTX580 = 115, HD7970 = 145, then 145/115 = 26%).

^ This is highly unlikely because there is 0 chance that Kepler is only 26% faster than a GTX580. AMD can't be naive to think that being just 45% faster for their high-end HD7970 card is sufficient enough to compete with Kepler.

3) 28nm process will allow AMD to add more transistors and increase frequencies. They added 12-15% more performance on top of the 5870 in 6970 by simply doing minor tweaks on the same 40nm process (VLIW-5 --> VLIW-4), added barely any more memory bandwidth and barely added clocks. They were still able to improve Tessellation performance improved significantly. Imagine what can be done on a far more complex architecture that can work on more things simultaneously. Also, I am pretty sure this time they'll add a lot more memory bandwidth.

As rumors point to performance approaching an HD6990, I am guessing HD7970 is 60-80% faster than HD6970 not 45%.

 

[tweakboy]

45 percent faster ? Wtf is that, soo if in a area in BF3 you get 52fps youll get 85fps with the new 6xx series ?

 

[RussianSensation]

45 percent faster ? Wtf is that, soo if in a area in BF3 you get 52fps youll get 85fps with the new 6xx series ?

The title of the article is misleading. It should apply strictly to transistor speed improvement on 28nm vs. 40nm. It doesn't consider any architectural enhancements that Kepler or HD7000 series will bring.