Tahiti or Kepler Respin with TSMC 28HPm ?

[24601]

How likely is it AMD and/or NVidia to respin Tahiti or Kepler for 28HPm for access to higher performance at same TDP?

 

[BrightCandle]

AMD have specifically stated nothing this year so presumably they are going to await a process shrink before they release a new card.

I have no idea if Nvidia intends to use the high performance process and produce a bigger better chip bit I doubt they will only release titan this year.

 

[VulgarDisplay]

AMD have specifically stated nothing this year so presumably they are going to await a process shrink before they release a new card.

I have no idea if Nvidia intends to use the high performance process and produce a bigger better chip bit I doubt they will only release titan this year.

I wouldn't be surprised if both companies ride it out this time.

 

[3DVagabond]

I thought Tahiti and Kepler were already using 28nm HP?

 

[24601]

I thought Tahiti and Kepler were already using 28nm HP?

HPm is lower leakage than HP with same or higher clocks.

 

[AtenRa]

HPm is lower leakage than HP with same or higher clocks.

Any links for that ????

 

[24601]

http://www.tsmc.com/english/dedicatedFoundry/technology/28nm.htm

The 28nm high performance (HP) process is the first option to use high-k metal gate process technology. Featuring superior speed and performance, the 28HP process targets CPU, GPU, FPGA, PC, networking, and consumer electronics applications. The 28HP process supports a 45 percent speed improvement over the 40G process at the same leakage/gate.

TSMC also provides high performance for mobile applications (HPM) technology to address the need for applications requiring high speed as well as low leakage power. Such technology can provide better speed than 28HP and similar leakage power as 28LP. With such wide performance/leakage coverage, 28HPM is also ideal for many applications from networking, tablet, to mobile consumer products.

Tegra4 and the new Kraits are HPm, so we know Nvidia can do HPm if they want to.

 

[ShintaiDK]

GPUs dont fit the HPm at all.

 

[24601]

GPUs dont fit the HPm at all.

Considering Fermi core clocks fit in 40LP (shader clocks in 40g) I'm pretty sure Kepler and Tahiti do just fine on HPm.

GPUs don't do 3 ghz therefore don't need 28HP. HPm is good for at least 2 ghz.

 

[Tuna-Fish]

GPUs don't do 3 ghz therefore don't need 28HP. HPm is good for at least 2 ghz.

... Do you have a clue of what you are talking about? Even a little bit? Clock speeds are not something that can be compared cross applications like that. It cannot possibly make any sense ever.

If I design two chips, one where the critical path is twice as long as the other, the same process will run one at 4GHz and other at 2GHz. The 4GHz one doesn't "need" any more from the process than the 2GHz one.

GPUs run at slower clocks because critical paths in them are longer. Their lower clocks do not in any way mean they require less from the process. The reason their critial paths are longer is that the natural, common operation in them is a FP FMA, so that's what they are optimized to efficiently provide. CPUs are optimized to efficiently provide addition and subtraction. Since those are easier operations, clock speeds are higher.

 

[24601]

... Do you have a clue of what you are talking about? Even a little bit? Clock speeds are not something that can be compared cross applications like that. It cannot possibly make any sense ever.

If I design two chips, one where the critical path is twice as long as the other, the same process will run one at 4GHz and other at 2GHz. The 4GHz one doesn't "need" any more from the process than the 2GHz one.

GPUs run at slower clocks because critical paths in them are longer. Their lower clocks do not in any way mean they require less from the process. The reason their critial paths are longer is that the natural, common operation in them is a FP FMA, so that's what they are optimized to efficiently provide. CPUs are optimized to efficiently provide addition and subtraction. Since those are easier operations, clock speeds are higher.

I say again.
Fermi core clock was on 40LP.
Fermi core clock went up pretty damn high.
28HPm (And even 28HPl) go far higher clocks than 40LP for the same exact architecture.

Therefore. Kepler and Tahiti should do fine at 28HPm.

I shouldn't have to spell this out to people.

When I say things in a complicated, specific, or technical manner, people flame me for apparently playing tricks on them.
When I simplify things deliberately so a wider audience can understand me, people accuse me of having no clue about what I'm talking about.

Internets, where logic goes to die.

 

[tviceman]

I'll be disappointed with Nvidia if they don't refresh GK104/106/107. AMD has the better prices and bundles right now and nvidia cannot expect to ride their name all year long. I doubt they'll change the process on which it is manufactured if they do refresh, though.

 

[ShintaiDK]

I'll be disappointed with Nvidia if they don't refresh GK104/106/107. AMD has the better prices and bundles right now and nvidia cannot expect to ride their name all year long. I doubt they'll change the process on which it is manufactured if they do refresh, though.

The 28nm refresh for both companies will most likely come in Q4, if it comes. Until then its game bundles, Titan and renaming.

But dont expect anything great until 20nm. But that sounds like a 2015 event.

 

[3DVagabond]

I'll be disappointed with Nvidia if they don't refresh GK104/106/107. AMD has the better prices and bundles right now and nvidia cannot expect to ride their name all year long. I doubt they'll change the process on which it is manufactured if they do refresh, though.

Considering the 670/680 have smaller GPU's, less VRAM, and simpler PCB's nVidia could easily drop prices if needed. They are obviously happy with the way they are selling.

 

[Tuna-Fish]

I say again.
Fermi core clock was on 40LP.
Fermi core clock went up pretty damn high.

Fermi shader clock could go pretty damn high because it had shorter critical path. The whole fast clock/slow clock idea was that things that were too complicated were pushed to the slow clock, and the fast clock just ran the basic alus. Even then, Fermi couldn't forward the result of a previous computation to the next one in two cycles -- it took something like 18 cycles iirc. This allowed individual pipeline stages to be shorter, which allowed higher clocks. The clock difference between Fermi and Kepler/Tahiti is architectural. For Kepler to reach the same clocks as Fermi, it needs a *lot* faster transistors (and especially interconnects).

28HPm (And even 28HPl) go far higher clocks than 40LP for the same exact architecture.

True. But, Kepler/Tahiti are not the same architecture. Both have considerably longer critical paths, mainly because they are much better at forwarding results. This means that their clocks are necessarily slower on a process-independent basis. They need those fast transistors just to approach the same speeds the past gen ran at.

Therefore. Kepler and Tahiti should do fine at 28HPm.

I do not know about how good HPm is -- I haven't yet talked to a chip engineer with experience on it, and TSMC publications are only marginally reliable. However, that was not what my outburst was about.

It was about:

GPUs don't do 3 ghz therefore don't need 28HP. HPm is good for at least 2 ghz.

Which really still makes no sense, even with your explanation. GPUs require fast transistors and interconnects just as much as anything else that is made on that process, and saying "GPUs don't do 3 ghz therefore don't need 28HP" is a wtfbbq meaningless statement, that makes me question all your knowledge about semiconductors.

 

[24601]

Fermi shader clock could go pretty damn high because it had shorter critical path. The whole fast clock/slow clock idea was that things that were too complicated were pushed to the slow clock, and the fast clock just ran the basic alus. Even then, Fermi couldn't forward the result of a previous computation to the next one in two cycles -- it took something like 18 cycles iirc. This allowed individual pipeline stages to be shorter, which allowed higher clocks. The clock difference between Fermi and Kepler/Tahiti is architectural. For Kepler to reach the same clocks as Fermi, it needs a *lot* faster transistors (and especially interconnects).


True. But, Kepler/Tahiti are not the same architecture. Both have considerably longer critical paths, mainly because they are much better at forwarding results. This means that their clocks are necessarily slower on a process-independent basis. They need those fast transistors just to approach the same speeds the past gen ran at.

I do not know about how good HPm is -- I haven't yet talked to a chip engineer with experience on it, and TSMC publications are only marginally reliable. However, that was not what my outburst was about.

It was about:


Which really still makes no sense, even with your explanation. GPUs require fast transistors and interconnects just as much as anything else that is made on that process, and saying "GPUs don't do 3 ghz therefore don't need 28HP" is a wtfbbq meaningless statement, that makes me question all your knowledge about semiconductors.

http://www.arm.com/about/newsroom/a...s-for-tsmc-40nm-and-28nm-process-variants.php

At TSMC 40nm LP (low power), ARM’s existing POP offering for the Cortex-A5 and Cortex-A9 processors is being augmented with the new Cortex-A7 POP. In addition, working in concert with TSMC, ARM will offer new POP variants supporting the latest high-speed options for TSMC 40nm LP, so those process options can take full advantage of the POP implementation benefits. ARM’s POPs for TSMC 40nm LP for Cortex-A5 (1.0 GHz) and Cortex-A9 (1.4 GHz) are shipping in production chips by ARM partners in such applications as smart-TV, set-top box, mobile computing and smart phones.

40LP allows Cortex A9 to get 1.4ghz as of April 16, 2012
(Higher frequency ones were on 40LPG or 40G)

http://www.anandtech.com/show/6787/...re-deep-dive-plus-tegra-4i-phoenix-hands-on/3

Combined with the 2.3GHz max frequency, Tegra 4i’s CPU performance should be a healthy improvement over what we have in Tegra 3 today.

Tegra 4i (Cortex A9) does 2.3 ghz on HPm

1.4 ghz to 2.3 ghz on same fundamental architecture is pretty huge, I would say.

Scaling frequency from Fermi architecture to say Kepler architecture would have to be more than 1.64285714286x less capable for HPm for Kepler to be infeasible.

I highly doubt that.

Also. Good job latching onto my simplification aimed at making sure a wider audience understood what I was saying, and then deliberately not reading my explicit statement about it in the very post you made sure to quote in full.... except for that part.

Your objectivity is questionable.

 

[DiogoDX]

If I remember well Tahiti uses HPL and Kepler HP. So if AMD is going to HP already be an upgrade in the manufacturing process.

 

[24601]

If I remember well Tahiti uses HPL and Kepler HP. So if AMD is going to HP already be an upgrade in the manufacturing process.

Tahiti and GK104 are HP (High Performance HKMG), and Pitcairn/Cape Verde are HPL ( Low power HKMG)

Ah, now I remember why Tuna was completely bullshit.

Tuna you are 100% wrong.

Kepler and Tahiti are 100% doable on HPm. They can even be done on HPl.

/ignore

FUD spreaders go die in a fire. <-= Directed at Tuna and his ilk, rhetorical questions are fun huh bro?

 

[DiogoDX]

Tahiti and GK104 are HP (High Performance HKMG), and Pitcairn/Cape Verde are HPL ( Low power HKMG)

Source please.

Before Tahiti lauch was a rumor than AMD switch to HPL to lauch ahead of Nvidia.

Initially, both of these chips were supposed to be built using the 28nm HKMG high-performance processes, but it seems like TSMC is having some issues with it and that it won't be available until the first quarter of 2012.

As a result, AMD apparently switched to 28nm HPL for Southern Islands, which is also based on the HKMG technology but is tuned for low power.

http://news.softpedia.com/news/AMD-...-Different-28nm-Processes-Report-212552.shtml

 

[tviceman]

The 28nm refresh for both companies will most likely come in Q4, if it comes. Until then its game bundles, Titan and renaming.

But dont expect anything great until 20nm. But that sounds like a 2015 event.

Releasing refreshed parts 6-8 months before 20nm cards hit does not make any sense. If refreshed parts are coming from nvidia, I expect it to start happening in q2, right in between the current products and future 20nm cards.

 

[24601]

Source please.

Before Tahiti lauch was a rumor than AMD switch to HPL to lauch ahead of Nvidia.


http://news.softpedia.com/news/AMD-...-Different-28nm-Processes-Report-212552.shtml

I'm gunna have to find a more believable source for that.

My current sources for that comes from that same approximate rumor.

EDIT:
https://chipworks.secure.force.com/...igationStr=CatalogSearchInc&searchText=tahiti

This report is an SRAM Analysis of the AMD Radeon 7970 (215-0821060) GPU, built on TSMC's 28 nm HP process. This report provides an overview of SRAM technology that includes plan-view bevel, cross-sectional TEM and SEM analysis and targets the minimum SRAM cell. Critical dimensions and extraction of the cell schematic are provided to help understand the SRAM cell.

Chipworks says Tahiti is TSMC 28HP.

 

[DiogoDX]

Agreed.:thumbsup:

Earlier this year, we completed a limited analysis of the high density SRAM on the AMD RadeonTM HD 7970 215-0821060 graphics processor, which was fabricated with TSMC’s HP process. Our TEM analysis confirmed the 215-0821060 transistor structure was identical to that seen in the Altera Stratix V device, as would be expected since both are based on the TSMC 28 nm HP process.

http://www.chipworks.com/blog/technologyblog/2012/12/11/a-review-of-tsmc-28-nm-process-technology/

Strangely this information is difficult to be found in reviews.:hmm:

 

[Tuna-Fish]

Ah, now I remember why Tuna was completely bullshit.

Do you even bother to read my posts? I never once said that Kepler or GCN would not be doable on HPm.

What I took offense with was your completely inaccurate remarks about clock speed.

And no, it's not a simplified for newbies view. It's a simply incorrect view. The only message it can possibly pass to people, whether they are informed or uninformed about the tech is bullshit.