[ combuster @ 27.03.2010. 10:15 ] @
http://www.anandtech.com/video/showdoc.aspx?i=3783

Citat:

Fermi/GF100/GTX400 has been a long time coming. For an industry where the major players are usually in lockstep with each other in some fashion, this latest cycle has been distinctly out of sync. In September of 2009 we saw AMD launch their Radeon 5000 series, and spend the next 6 months as the performance leader for the discrete graphics market for virtually every price point. NVIDIA on the other hand missed the latest cycle by a mile, and when AMD was delivering 5000 series cards NVIDIA could only talk about the High Performance Computing applications of their next GPU, Fermi. It would be the next big thing, both figuratively and literally.

To wrap things up, let’s start with the obvious: NVIDIA has reclaimed their crown – they have the fastest single-GPU card. The GTX 480 is between 10 and 15% faster than the Radeon 5870 depending on the resolution, giving it a comfortable lead over AMD’s best single-GPU card.

AMD’s 2GB Radeon 5870, which will be launching soon. We know the 1GB 5870 is RAM-limited at times, and while it’s unlikely more RAM on its own will be enough to make up the performance difference, we can’t fully rule that out until we have the benchmarks we need. If the GTX 480 doesn’t continue to end up being the fastest single-GPU card out there, we’ll be surprised.

Fermi’s compute-heavy and tessellation-heavy design continues to interest us but home users won’t find an advantage to that design today. This is a card that bets on the future and we don’t have our crystal ball. With some good consumer-oriented GPGPU programs and developers taking up variable tessellation NVIDIA could get a lot out of this card, or if that fails to happen they could get less than they hoped for. All we can do is sit and watch – it’s much too early to place our bets.


IMHO zahebali su stvar. Bigtime. Prvo sto im je yield jednocifren po nekim navodima. Cak ni full spec ne mogu da proizvedu u ovom trenutku pa okidaju core-ove. Greje se i srce ko blesav sto i nije cudo za celih 500mm^2 povrsine die-a. Toliki gpu verovatno i prouzrokuje tako nizak yield u 40nm proizvodnom procesu.

Po meni poslovace sa gubitkom i to zescim. Ne verujem da ce ih biti dovoljno na trzistu, makar ne u narednih 4-5 meseci to je sigurno. Cak je i GTX470 skuplji od 5850 a ne nudi toliko vece performanse a greje se duplo a srce za 50% vise (mozda i jace). Koliko se karta greje mozete pretpostaviti i po isecenom pcb-u sa zadnje strane i hladnjaku koji se vidi sa prednje (o 5 copper pipe-ova da ne pricam, zalman fora).

Jedino ako ih teselacija izvuce pa ce se mozda i primetiti boljitak u nekoj nvidia optimized igri ali ni to nece skoro...

Vraca se losa karma za 3dfx ;-) U ono sto sam uveren je da ce tu malu dostupnu kolicinu razgrabiti nv fan-ovi 3pujuci se da imaju supercomputing zivotinju u kucistu. Na tu kartu i igra nvidia, pravili su hype-a oko ovoga i van Stevinih proporcija.

E Tesla, sta doceka...


[ Ivan Dimkovic @ 27.03.2010. 12:24 ] @
Anand je imao jako dobar intervju sa ATI-jem koji je izuzetno kvalitetan i govori o radu na stvaranjy ATI Cypress cipa:

http://www.anandtech.com/video/showdoc.aspx?i=3740

I tu se moze videti geneza NVidia-inog FAIL-a sa Fermijem... strahovito pogresne procene o TSMC 40nm procesu ce ih kostati uspeha u ovoj generaciji, a vrlo verovatno i sledecoj.

Nije to nista novo, i ATI je imao slican Epic Fail momenat kada je NVidia izbacila G80 pre nekoliko godina. Trebalo im je preko 2 godine da se skockaju i vrate nazad kao konkurentni.

--

Sve u svemu, Fermi arhitektura za gaming je osudjena na propast. NVidia ce vrlo brzo morati da dampuje cene i da gube silne pare po svakoj prodatoj kartici. Ostaje im HPC trziste koje ce sigurno kupovati Fermi kartice bez problema, samo je pitanje koliko je to HPC trziste i da li moze da pokrije gubitke na mainstream trzistu.

Jedna stvar je sigurna - u NVidia-i sigurno tesko rade na post-Fermi generaciji.
[ combuster @ 27.03.2010. 12:29 ] @
http://www.semiaccurate.com/20...migtx480-broken-and-unfixable/

Ovo sam citao pre par nedelja, mnogo lepo opisuje u kakvim su g*vnima...

Citat:

Hot, slow, late and unmanufacturable


Sve je lepo covek rekao u jednoj recenici tj u samom naslovu. Ne bih rekao da je slow ali ajde sad.
[ Ivan Dimkovic @ 27.03.2010. 12:34 ] @
Yep.

Za Fermi Fail se znalo vec mesecima. Mada, za korisnike ovo moze biti dobra stvar zato sto ce pre ili kasnije NVidia morati da se takmici sa cenom, sto znaci da ce i ATI morati da obori istu :-)

Uzgred, postoje 100% kompletni GF100 primerci u full-specu, samo sto ih NVidia trenutno binuje za Tesla/Quadro kartice. Kao sto je poznato, yield-ovi su tragicno niski u jednocifrenom %, tako da svaki 100% ispravni silikon zlata vredi (cuj zlata... ta plocica vredi N-tostruko vise od iste kolicine materijala u zlatu ;-). A to znaci da je jedino isplativo valjati ga kao preskupi Tesla/Quadro SKU.

ATI je imao veliku prednost ovde sto su prvi poceli da petljaju sa TSMC 40 nm procesom. Tacno na vreme, reklo bi se - jer su imali vremena da donesu kljucne odluke za Cypress koje su bile "SECI!" (feature!). Sada se jako lepo vidi koliko su te odluke bile dobre. Ima tu jos pikantnih stvari, recimo otkrivanje na vreme da TSMC 40nm proces ima gomilu drugih problema (nepouzdanost veza) itd... NVidia je donela katastrofalnu (za njih) odluku da prepusti ATI-ju da se zahebava sa 40nm procesom.... nisu racunali na to da ce ATI dosta toga i nauciti :)
[ combuster @ 27.03.2010. 12:50 ] @
Citat:

Number one on Nvidia's hit list is yields. If you recall, we said that the yield on the first hot lot of Fermis that came back from TSMC was 7 good chips out of a total of 416 candidates, or a yield of less than 2 percent.

Why are things this bad? The simple answer is that Nvidia didn't do its homework. Just like the failures that lead to the bad bumps, Nvidia simply didn't do test, instead it tried to brute force things that require nuance and methodical forethought. ATI did test. ATI put out a full production run of HD4770 (RV740 silicon) GPUs and used that as a test of TSMC's 40nm process. TSMC looks to have failed, badly, but crucially, ATI learned why and how the parts failed to yield. That learning was fed back in to the Evergreen 5000 series GPUs, and those come off the line at perfectly acceptable yield rates now.

By the time Nvidia found out about the problems, it was far too late to implement them in Fermi GF100. Unless TSMC pulled off a miracle, the design was basically doomed.

Why? GF100 is about 550mm^2 in size, slightly larger than we reported after tapeout. Nvidia ran into severe yield problems with a 100mm^2 chip, a 3 month delay with a 139mm^2 chip, and had to scrap any larger designs due to a complete inability to manufacture them. Without doing the homework ATI did, it is now trying to make a 550mm^2 part.

Rumors abound that Nvidia will only have 5,000 to 8,000 Fermi GF100s, branded GTX480 in the first run of cards. The number SemiAccurate has heard directly is a less specific 'under 10,000'. There will have been about two months of production by the time those launch in late March, and Nvidia bought 9,000 risk wafers late last year. Presumably those will be used for the first run. With 104 die candidates per wafer, 9,000 wafers means 936K chips.


To je ono o cemu pricas, Ljudi su odradili RV740, prsli malo tu, videli gde koci i sad Evergreen samo bije.

Dakle odnos snaga:

ATI je mnogo pre zapoceo sa uspesnom proizvodnjom DX11 GPU-ova, Eyefinity, hdmi i dp (fermi nema dp uopste), imali su dovoljno vremena da proizvedu dovoljan broj kartica, odradili deal-ove sa OEM-ima na vreme i ostavili trziste za nvidia-u taman toliko da zadovolji svoje fanboy-eve.

Ubise ih i ubijace ih u narednih dve godine. A to ne valja. Zato sto ce i oni malo da uspore sa razvojem, zasitice trziste sa lowcost Juniper i high end Cypress chipovima. Cuo sam neke najave sta planiraju za 2011-tu ali to ce po svemu sudeci malo da saceka (ako su pametni a jesu).

Ne verujem da ce doci do pada cena. 5850 jedini ima razloga da se plasi malo jer mu je direktan konkurent GTX470, 5870 nema takmaca a u lowend trzistu ce da zavladaju 5770 i slabiji... A ko ce da pljune onolike pare za GTX480 ja stvarno ne znam :)

[ Ivan Dimkovic @ 27.03.2010. 13:03 ] @
GTX480 ce biti zanimljiv:

a) Fanboy-ima... mada dobro, to se podrazumeva
i
b) CUDA ekipi koji imaju tesne budzete pa Tesla otpada iz igre

Sto i nije neko ogromno trziste, ali ce im bar vratiti neku lovu koju su spalili u predugom R&D-u.