His "most wrong" (among the many things he were right about, though!) was back at the inquirer:
We know how this particular bit of analysis turned out.
Semi Accurate's 4XX views
- Thread starter Cowboy X
- Start date
- Status
Nvidia saying the GTX 480 has a 250 TDP does not supercede the laws of physics. If it walks like a GTX295, and quacks like a HD5970, it's probably a ~290w duck.
Nvidia just made up a number that sounded good.
Silent_Buddha
Legend
Well to be fair it does stay around 250 or less in some games.
Regards,
SB
The card has a "Average observed TDP of 250W" 295W is actually the average observed TDP when anyone other than nVidia benches the card, and I think Groo would've died of auto-erotic asphyxiation if he knew the card would draw north of 320W in some other benchmarks.
If you want to attack Charlies sources, ask about NVIO or something.
Last edited by a moderator:
It seems Nvidia was tweaking TDP numbers till the last second, and the number they came up was more wrong than Charlies (informed?) prediction in actual games and benchmarks
Interesting bit is, no one was closer to actual TDP numbers than Charlie, even Rys (who was referring to NV engineers)
Probably Groo has some moles in the fabs.True, Fermi appeared to be surprisingly robust in the teselation, still there are some sparks of truth in his reasoning. For example, Evergreen tes. stays the same top to bottom (more or less), while Fermi tes. performance is strongly tied to the shaders. So how about heavy shaders game and tes.? How about mid/lower gpu's parts, who have 1/2, 1/4, etc. shaders?
IBM built NV40. At least some of them anyway.
You mean a number that would look like it jived with the PCI-E bus certification.
Fermi tessellation seems to be scaling with number of GPCs, so if GF100 appears not to be hampered by its throughput, chances are that'll stay true for mid-range and lower parts too. But no matter what: A nosedive is something else.
And you're right: in Evergreen everybody tessellates the same (AFAWK). So the bottleneck appears to be less constricting in mid range and lower end parts. But there's also less shaderpower and since every tessellated patch needs some shading (HS, DS) plus the related pixelwork, it'll just shift the bottleneck.
I agree. He is biassed but his info was mostly reliable.
You mean like the part where the GF100 had no tesselator, which was later revised to tesselation will be slow because the shaders are co-opted to do tesselation, which was further revised to OK it does tesselation, but all the derivatives will have no tesselation or suck at tesselation (depending on which article you read)?
Then we can consider the whole GF100 is vaporware, unmanufacturable and will never be available to the general public meme (he's still spinning that one to say he always meant 'nearly unmanufacturable' and by no availability he always meant minimal availability stretched over many months.)
How about the numbers, dates and pricing guesses? I think if you mapped out all the guesses he took regarding release date, he probably at one point or the other must have guessed every week in Q2 and Q3 of this year.
Give him his due though, he clearly broke some facts first. You just had to weed through 10 incorrect guesses to arrive at the right one.
I will say that Charlie clearly has good contacts on the Taiwanese / Chinese side, some solid information from some of those contacts, and a good enough understanding of silicon processes and manufacturing and the current state of the industry to report on and interpret the information he's getting. But he looks at things through such a biased filter, always assuming the worst case for any possible interpretation of limited facts, that you are forced to apply a "Charlie filter" to everything you read from his site. The compound errors that pile up from the worst case assumptions he makes in his extrapolations tend to throw off the conclusion of nearly every analysis he does, but there's always some accurate information and analysis mixed in with the the misses. You just can't blame a site called Semi-accurate though. It's right there in the name. He's telling you it's semi-accurate - in no way is he overselling his own articles, which just makes it funny when people argue that it's all accurate. Even Charlie doesn't believe that.
This may sound like a lot of negative criticism, but I do read SemiAccurate, because it can be interesting (if sometimes infuriating) and does often break news, but I always do so with my Charlie Filter on at 100% strength.
Then we can consider the whole GF100 is vaporware, unmanufacturable and will never be available to the general public meme (he's still spinning that one to say he always meant 'nearly unmanufacturable' and by no availability he always meant minimal availability stretched over many months.)
How about the numbers, dates and pricing guesses? I think if you mapped out all the guesses he took regarding release date, he probably at one point or the other must have guessed every week in Q2 and Q3 of this year.
Give him his due though, he clearly broke some facts first. You just had to weed through 10 incorrect guesses to arrive at the right one.
I will say that Charlie clearly has good contacts on the Taiwanese / Chinese side, some solid information from some of those contacts, and a good enough understanding of silicon processes and manufacturing and the current state of the industry to report on and interpret the information he's getting. But he looks at things through such a biased filter, always assuming the worst case for any possible interpretation of limited facts, that you are forced to apply a "Charlie filter" to everything you read from his site. The compound errors that pile up from the worst case assumptions he makes in his extrapolations tend to throw off the conclusion of nearly every analysis he does, but there's always some accurate information and analysis mixed in with the the misses. You just can't blame a site called Semi-accurate though. It's right there in the name. He's telling you it's semi-accurate - in no way is he overselling his own articles, which just makes it funny when people argue that it's all accurate. Even Charlie doesn't believe that.
This may sound like a lot of negative criticism, but I do read SemiAccurate, because it can be interesting (if sometimes infuriating) and does often break news, but I always do so with my Charlie Filter on at 100% strength.
Silent_Buddha
Legend
Which is all not necessarily untrue either. GF100 doesn't have a dedicated tesselation unit. It does use the shaders. And it might actually be slower when doing tesselation with shader work (Dirt 2 Dx9 to Dx11 performance hit on GF100 is extreme). And we won't know about derivatives until they are released.
Which may or may not be true depending on when you think about them. Was GF100 unmanufacturable in 2009, sure appears to be. Was it unmanufacturable in Jan. - Feb. of 2010? Pretty much, although I suppose they could have released a trickle. And we still have yet to see just how manufacturable it is. It's quite possible that GF100 A3 will be limited to what wafers have already been used for it, and it was just selling those to get something back rather than proceeding with mass production. That might or might not be waiting until B1.
In other words we still have no clue whether GF100 A3 is manufacturable with good enough returns to post a positive revenue stream. Considering they still didn't have enough for anything resembling a respectable hard launch on the 26th of March, this remains in doubt, with many speculating it won't be until at least B1 that GF100 is manufacturable, if by manufacturable you mean being able to turn a profit with the product being manufactured.
Considering, Nvidia and manufacturing partners were also wrong on that, I don't think you can blame Charlie for passing on the information that was relayed to him by people within those organizations.
And compared to everyone else releasing GF100 rumors. Well, it appears he was far more accurate than most if not all of them.
Regards,
SB
Last edited by a moderator:
DavidGraham
Veteran
I think Nvidia experience with DX11 games is still in infancy , ATi itself kept optimizing for Dirt 2 till the last driver , may be Nvidia needs to do the same.
In Metro 2033 , the performance hit of enabling tessellation on HD5870 was 15% , on GTX 480 it was 7.5% , however for DOF , it was 38% for HD5870 and 35% for GTX 480 , that should count for something right ? (Using the same seetings , GTX480 was 23% faster than HD5870 ) :
http://www.hardocp.com/article/2010/04/05/metro_2033_directx_11_gameplay_performance_iq/5
Hey, silly season is over for GF100.
Nah it takes a while for FUD to evolve into fact
I'll be surprised if anyone knows the original source for this bit of hilarity.RE: Fabs for GPUs
IBM is very expensive and notice that NV has stayed away from them ever since NV40. UMC is generally a follower and behind Intel, IBM and GF and usually TSMC on most process nodes. I don't believe that UMC has ~40nm in production, but I could be wrong. I don't really follow UMC that carefully since they tend not to be on the cutting edge.
TSMC is the 800lb gorilla of the fab business - IP and tool vendors target TSMC by default, etc.
More importantly, you cannot just change foundries at the drop of a hat. Your physical design is tied to the characteristics of the foundry, their design rules, etc. etc. Also, if you use 3rd party IP, you'll need to port that over (note that GF is hiring up tons of people to do that). Starting your physical design over will cost you a lot of time...at which point, you're product isn't that competitive.
So if your design started life targeting TSMC, you simply cannot do an about face and move somewhere else quickly. Sometimes companies will dual track their designs (using two different foundries), but that adds engineering costs and may reduce performance. Overall, it's exceptionally unusual for GPUs.
So for now, TSMC is really the only option. It's probable that AMD will shift to Global Foundries at some point. Fusion is using their high performance SOI process, although I doubt discrete GPUs will use SOI for cost reasons.
DK
IBM is very expensive and notice that NV has stayed away from them ever since NV40. UMC is generally a follower and behind Intel, IBM and GF and usually TSMC on most process nodes. I don't believe that UMC has ~40nm in production, but I could be wrong. I don't really follow UMC that carefully since they tend not to be on the cutting edge.
TSMC is the 800lb gorilla of the fab business - IP and tool vendors target TSMC by default, etc.
More importantly, you cannot just change foundries at the drop of a hat. Your physical design is tied to the characteristics of the foundry, their design rules, etc. etc. Also, if you use 3rd party IP, you'll need to port that over (note that GF is hiring up tons of people to do that). Starting your physical design over will cost you a lot of time...at which point, you're product isn't that competitive.
So if your design started life targeting TSMC, you simply cannot do an about face and move somewhere else quickly. Sometimes companies will dual track their designs (using two different foundries), but that adds engineering costs and may reduce performance. Overall, it's exceptionally unusual for GPUs.
So for now, TSMC is really the only option. It's probable that AMD will shift to Global Foundries at some point. Fusion is using their high performance SOI process, although I doubt discrete GPUs will use SOI for cost reasons.
DK
The question is, how much do these design rules and characterictics change between process changes when using the same fab?
Wasn't it said somewhere that ATI lost ~6 months of work due 32nm process being cancelled?
Do they change so significantly less, than fab to fab, that it's "always" worth it to stay on same fab if you need to design for new process regardless?
aaronspink
Veteran
Depends on how different the processes are. Metal layer changes can have a big impact. Right now the greatest impact of changing processes/fabs is likely on the analog side with the high speed interfaces, esp interfaces that aren't widely used in the industry (aka GDDR5).
But in general when changing processes the first impact you run into is the floorplan. Floorplans are generally pretty tied to the characteristics of a given process. Once you change process/shrink process its generally unlikely that the everything morphs by the same amount so you have to re-shape blocks and move blocks around esp with dynamic and analog circuits thrown into the mix as well as srams.
Then your whole entire P&R back-end has to be redone. Any test chips/structures you've done are also now invalid so you need to do another set of test structure/chips.
Generally within a companies processes the overall changes are less than between processes. There are a lot of gritty details around transistor rules that can change significantly depending on process vendor and generally metal stacks can be significantly different as well. Whereas with the same vendor, its likely that things like metal layers are simply pushed down the stack and remain relatively unchanged except for things like numberings as they add more layers.
Maybe it's finally time to get to the bottom of this...
DX11 tessellation has three steps:
Hull Shader
Fixed function tessellator
Domain Shader
1) It's obvious that GF100 is using the general shaders units for hull and domain shader. Are you saying it's doing the FF tessellator operations on the shaders too? Considering that it's a very specific kind of logic that doesn't map at all to multiple SIMD cores and that it's probably very small in area, that doesn't sound very likely, does it?
2) If you say GF100 doesn't have a dedicated tessellation unit, that's obviously in contrast to RV870. Are you claiming that RV870 is doing hull shader and domain shader in dedicated hardware outside the general shaders or is just the FF tessellator in dedicated hardware? If it's the latter, could you please explain how the GF100 and the RV870 are different (other than GF100 probably has 4 FF tessellators instead of 1 in RV870) ?
I really would like to know, because I'm puzzled by statements like the one you just made. At first sight, they just don't make sense... You obviously have better insight. Please enlighten me.
Thanks.
Last edited by a moderator:
- Status
Similar threads
