nVidia wins this round - Charlie D.

[notty22]

From CD twitter page :

https://twitter.com/#!/CDemerjian
CDemerjian Charlie Demerjian



HD7970 vs GK104: semiaccurate.com/2012/01/19/nvi… The winner is.......

http://semiaccurate.com/forums/showpost.php?p=149402&postcount=178

Quote:
Originally Posted by Thrillseeker
Is Charlie sure the cards he has seen or heard about are Gk104 and not some other version? (i.e. Gk110 etc)

Yeah, it is the 104, not 110. Well, the 110 is just 2x 104s, so you could say that it is the same chip, but neither of the cards I saw had two GPUs on them.

-Charlie

Which says to me, the GK110 will be 512 mb.

 

[LOL_Wut_Axel]

From CD twitter page :


http://semiaccurate.com/forums/showpost.php?p=149402&postcount=178

Which says to me, the GK110 will be 512 mb.

Don't really see why he's fixated on comparing the HD 7970 to whatever Enthusiast card NVIDIA will come out with. AMD will more than probably release the HD 8970 to counter it once NVIDIA releases it...

 

[notty22]

He's makes a living doing what he does. So he comes across as controversial and it's got him a following of people who both like or dislike him. He probably does not care as long as the web site does well.

 

[Jovec]

What I did expect is 28nm to redefine 40nm price performance but up to this point it really hasn't, imho. But, it's still early with many product sku's from AMD and nVidia to come.

Price reductions via new process nodes are dead. The AMD CEO basically said as much. The new nodes are costly to deploy, have low yields until/if the process matures, and the volume just isn't there to make up for it. Arguably only Intel and smart phone SoC producers are the only ones who are going to do enough volume for the upfront costs to pay off.

Besides, what company has ever willingly passed on savings from moving to a smaller node to the customer instead of pocketing them for itself? Unless competition forced them to.

 

[sontin]

Charlie said, on his forum, that:
-GK110 is actually two GK104;
-Kepler's memory controller is better, but still not by much

I think we can conclude that, if what Charlie said was right, then it GK104 is going to be between a HD 6950 and the GTX 580, most probably around GTX 570/HD 6970 and that nVidia finally turned into a "small die" strategy.

Also, if nVidia wins in perf/mm², I would guess the die size is going to be about 250mm².

His postings are showing that he has no informations and no clue about Kepler. D:

 

[Will Robinson]

It's amazing the turn around in the Green Team's opinion of Charlie and SA....a little NVDA loving and suddenly his articles are getting quoted and fawned over...

Our good friend Lonbjerg must be revolving in his forced retirement home like a Ferris wheel on steroids.

No callouts
-ViRGE

 

[Keysplayr]

It's amazing the turn around in the Green Team's opinion of Charlie and SA....a little NVDA loving and suddenly his articles are getting quoted and fawned over...

Our good friend Lonbjerg must be revolving in his forced retirement home like a Ferris wheel on steroids.

Use your head. Charlie's articles are getting attention because it is LUDICROUS coming from him. Probably his intention to get web hits. It's working isn't it?
Nobody switches like that in a weeks time. Unless something drastic happened.
Believe me. I'd be truly amazed if any Nvidia fan trusted Charlies words beyond the first syllable. I believe Kepler will be fantastic, but good Lord not because Charlie says so. hehe.

P.S. Your Lonbjerg callout could surely use some explaination. Revolving? Please enlighten us less fortunate.

 

[AnandThenMan]

Use your head. Charlie's articles are getting attention because it is LUDICROUS coming from him. Probably his intention to get web hits. It's working isn't it?
Nobody switches like that in a weeks time. Unless something drastic happened.
Believe me. I'd be truly amazed if any Nvidia fan trusted Charlies words beyond the first syllable. I believe Kepler will be fantastic, but good Lord not because Charlie says so. hehe.

Charlie has broken some pretty significant stuff. Fermi hot and late, bumpgate. I don't think it's fair of you to dismiss him outright just because you don't like the message. We'll find out how "ludicrous" his info on Kepler is soon enough, if he's right, will you give him props?

 

[Keysplayr]

Charlie has broken some pretty significant stuff. Fermi hot and late, bumpgate. I don't think it's fair of you to dismiss him outright just because you don't like the message. We'll find out how "ludicrous" his info on Kepler is soon enough, if he's right, will you give him props?

Don't like the message? I love the message. I'm an Nvidia fan, why wouldn't I love that message?
You're getting the wrong message. This isn't about Charlie being right or wrong. It's about his motive. Is he telling the truth? Hard to believe since his history clearly shows a malignant, obsessive-compulsive hatred of Nvidia. I mean EVERY article.
You don't find any of this odd beyond belief? I find that odd.

 

[AnandThenMan]

You don't find any of this odd beyond belief? I find that odd.

I actually don't. Reason being, Charlie has been shown to hold a grudge (understatement to say the least) so it is possible AMD did something to him that he can't let go. Maybe he has decided to "switch camps" so to speak. Possible.

What is certain is, there is no going back for him, if he is trolling all of us with this article I don't see how people can trust him again. I believe he is serious with this article.

 

[Keysplayr]

I actually don't. Reason being, Charlie has been shown to hold a grudge (understatement to say the least) so it is possible AMD did something to him that he can't let go. Maybe he has decided to "switch camps" so to speak. Possible.

What is certain is, there is no going back for him, if he is trolling all of us with this article I don't see how people can trust him again. I believe he is serious with this article.

Wonders never cease. Charlie has a cult following. Do you think they will abandon
him if they discover he is trolling?

 

[AnandThenMan]

Like any cult leader, the hard core followers will support their master even if it means the cyanide pill. D: But the rest, yea he'll lose a lot of respect from them.

 

[Cookie Monster]

Maybe kepler (Gk104) will be NVs version of the RV770?

 

[blackened23]

Maybe kepler (Gk104) will be NVs version of the RV770?

The only thing that can mean is that NV managed to make the die smaller than Tahiti. Now that doesn't necessarily mean that the performance will be higher In fact, a 256 bit memory bus and a smaller die probably means lesser performance but better effiiciency with higher profitability.

This is pretty shocking news if true, because traditionally NV has *never* gone for small die designs. The ramifications of a small die is higher profits and better efficiency. Something NV has always failed at in the past. Maybe they had a breakthrough recently?

Who knows though.

 

[AnandThenMan]

Are all GCN based GPUs going to be using a 384bit bus?

 

[Cookie Monster]

The only thing that can mean is that NV managed to make the die smaller than Tahiti. Now that doesn't necessarily mean that the performance will be higher In fact, a 256 bit memory bus and a smaller die probably means lesser performance but better effiiciency with higher profitability.

This is pretty shocking news if true, because traditionally NV has *never* gone for small die designs. The ramifications of a small die is higher profits and better efficiency. Something NV has always failed at in the past. Maybe they had a breakthrough recently?

Who knows though.

Actually they have in the form of G71 and G92. Instead of going to a 32 pipeline refresher, they opted for a refresh with fewer transistors, optimizations and hence resulted in a smaller sized chip. I still remember the stock cooling of the 7900GT variant. It was absolutely shocking (absolute bare minimum) but guessing that the margins on these cards were huge.

Also, what do you mean by better efficiency due to having a smaller die size? The layout and transistor density for a given chip on the top of my head would only affects how many you could produce per wafer. When looking at from a performance/power point of view, it doesn't effect it as much, if not at all.

If memory serves me right, AMD usually has the more dense layout as they are able to pack more transistors with a given space than nVIDIA for the past couple of generations.

 

[Cookie Monster]

Are all GCN based GPUs going to be using a 384bit bus?

I think only the high end will be using GCN based chips (VLIW5/4 for the rest of the lineup - correct me if I am wrong) so your probably right unless they will release an SKU based on severely cut down versions of the tahiti chip.

 

[blastingcap]

I have had some challenges with AMD's drivers in Eyefinity in 2011, but those paled in comparison to the nightmare I had tonight where I spent two hours fiddling with drivers to tune performance before giving up and rolling back to 11.6. At many points during the driver fiddling, I wound up with 2 heavily color-corrupted screens, with only my third screen being at ALL legible, and that's because it was being driven by my motherboard's IGP.

I am definitely looking forward to a highly competitive GPU situation in the near future. Even AMD fans should want such a situation. More competition is one of the ways consumers receive lower prices!

If NVDA's part is price/perf competitive and has better multi-monitor support, and if Kepler can drive 3 monitors on one GPU without any need for a FLEX-like card, I'm sold. Also, I want more than 1 but less than 3GB of VRAM. 3GB is just overkill for my personal needs, but 1GB isn't enough for certain games at certain resolutions. 1.5 to 2GB would be a great compromise.

 

[AtenRa]

I think only the high end will be using GCN based chips (VLIW5/4 for the rest of the lineup - correct me if I am wrong) so your probably right unless they will release an SKU based on severely cut down versions of the tahiti chip.

HD79xx, HD78xx and HD77xx will use the GCN architecture.

As for the 384-bit, i will expect only the 79xx series to have it.

 

[Cookie Monster]

HD79xx, HD78xx and HD77xx will use the GCN architecture.

As for the 384-bit, i will expect only the 79xx series to have it.

So that is good news since I thought 78xx/77xx were based on 28nm Cayman chips and not a new chip based on GCN architecture.

 

[SirPauly]

Price reductions via new process nodes are dead. The AMD CEO basically said as much. The new nodes are costly to deploy, have low yields until/if the process matures, and the volume just isn't there to make up for it. Arguably only Intel and smart phone SoC producers are the only ones who are going to do enough volume for the upfront costs to pay off.

Besides, what company has ever willingly passed on savings from moving to a smaller node to the customer instead of pocketing them for itself? Unless competition forced them to.

I guess so, seeing a percentage MSRP increase greater than an over-all percentage performance increase!

 

[Nemesis 1]

Why is this thread not locked and the other thread about AMD 7950 was locked. Why are AMD/ATI threads locked so easily wereas the NV articles can stay open so much longer . Not that I care since I use NV products . But it seems getting a AMD/ATI topic locked in this forum is way to often and way to easy.

 

[Olikan]

Maybe kepler (Gk104) will be NVs version of the RV770?

meh, i was thinking alot in how nvidia would beat a 7970 with 256bits...

diferential gddr5
something better than crossbar, ring bus

....i was getting hyped, then i got an arrow in the knee

 

[BenSkywalker]

I don't understand why the typical Charlie loyalists aren't stating how this must be the truth, he's always right etc, etc. Doesn't make a lot of sense to me. For his general comments, while the man is still a complete BS artist that makes things up as he goes, at the very least this time around it looks like he took the time to add.

When the GTX 285 hit, if it was ~35% faster then the 2900XT and nothing had launched between(the 8xxx/9xxx generation) for nV, how good of a part do people think it would have been? The term 'catasrophic failure' comes to mind. That is a full process node drop. That is what the 7970 is compared to the GTX 580.

Charlie still pulls complete crap out of thin air and posts it as fact, sadly he even tries to stick to his utterly delusional stances even after the fact(Tegra line posting nine figure profits, the same line that was a complete failure and failed to get a single product placement- Charlie still calling it a failure). That he is capable of deducing that something went shockingly wrong with the 7970 simply indicates he knows what a full node drop should be. It appears that AMD took their half node design and changed pretty much nothing and brought it out with only what tweaks it needed to be built on the new process. This would speed up their time to market considerably, and it would also result in them likely ending up with parts that should OC to a rather extreme degree, but also leave them in a position when a part designed for the process actually hits they are going to need to refresh pretty much immediately.

 

[Riek]

meh, i was thinking alot in how nvidia would beat a 7970 with 256bits...

diferential gddr5
something better than crossbar, ring bus

....i was getting hyped, then i got an arrow in the knee

Well a 7970 with bandwidth of the 6970 performs on average 13% less. So even a 7970 with the highest possible memory on 256bit bus would perform the same..

However overclocking would results in next to nothing scaling wise. (unlike what we see on the 7970 with a 384bit bus.).

So my guess is still that Nvidia pushed the bins so that they have excellent performance on a smaller die. At the cost of yields and overclocking capability.


At least if the GK104 performs like Charlie suggested, i'm also wondering why AMD defined the 7970 and 7950 as they did. Nothing prohibited them to set the bar 20% higher on both models without a TDP increase.