Ah, moderation makes an appearance
but that is kinda the point dude. you are defending the (then) ludacris cost of a "gaming" card because it sold well with professional users...
Bit of a reading fail on your part then. What I said was people buy the tool for the job, and sales are sales regardless of the end users intent - it is actually no different to the sales (and inflated pricing) attached to Radeon cards (also marketed as gaming) due to sales to miners, many of whom did nothing gaming related with the cards at all.
Would you be so kind to explain how TSMC would be capable of manufacturing a 4096-shader Fiji on its existing 28nm processes?
Assuming this is a genuine question then...
In theory, it would be fairly easy. Most people
should realize that a large die performance/enthusiast GPU devotes ~50% of its die area to cores and TMU's. The remaining 50% comprises the uncore (memory controllers, memory interfaces, command processor, transcode engine, raster ops etc.)
The green area's are the core, everything else the uncore
The uncore is relatively fixed in size if the memory interfaces (bus width) remain static. Hawaii at 2816 cores is 438mm^2, half of which is cores and texture address units (220mm). If the core count is increased by 45% ( to 4096) then the area devoted to it increases to 319mm^2. Add the 220mm^2 for the uncore and the resultant die area becomes 539mm^2 - or just slightly smaller than GK 110.
That is how TSMC is capable of manufacturing a 4096 shader Fiji. Whether they are the foundry involved depends on when AMD decided to use GloFo's 28nm SHP process for GPUs in addition to Kaveri APUs. One of these two processes will almost certainly be the manufacturing node involved.
Are you claiming that everything they stated back in November last year is correct?
If you'd bothered to read what I wrote it would be obvious that what I was pointing out was that 3DC attributed the name Fiji to the 4096 shader part. I might also point out that many other sources do the same including a well known AMD brown-noser who claims intimate knowledge of AMD's business (
although you'll have to stump up a fee to breach the paywall ). Have AMD swapped the names around? were they in the right order to begin with? Who knows, although I'd note that the other parts in the hierarchy don't seem affected.
So, you think 3dcenter's info is plausible, while Chiphell's is not?
3DC don't release leaks, they gather information and extrapolate from that. Their membership includes a number of industry insiders, coders, architects. Chiphell on the other hand are a conglomeration like any forum based site. The validity of their information depends upon the individuals posting there. Some is legitimate, some is quasi-legitimate (access to samples but results/info massaged for PR spin ***cough**Coolaler**cough***), some is estimation/guesstimation, and some is outright bullshit. Chiphell posts should be taken on a case by case basis- especially from posters with little or no previous track record of providing reliable information.
In this particular instance, we have a poster with no previous record for releasing reliable leaks, quoting a manufacturing process wholly unsuited for large GPUs, using a naming convention at odds with the rest of the tech world, and showing results that would indicate perfect scaling for both vendors which supposes mature drivers for both AMD and Nvidia months out from launch....all this plus a single source having access to not just one unreleased top-tier card, nor two, nor three, but
four - access that includes both AMD and Nvidia.
I also find it difficult to accept that this guy benchmarked four unreleased cards (along with comparisons with many released cards) across 20 games, yet can't provide any shred of photographic evidence, no standard benchmark validations (Heaven, 3DMark), nor power figures for AMD's top part, nor any single game numbers. All a bit convenient.