nVidia G9x - informacie (povodne aj spekulacie)

fobos

wendigo5 napísal:Ypsilonova os ma by´t akoze pocet tranzistorov?

Ano

1 miliarda tranzistorov sa uz dlhsie potulovala internetom ...

ale este by bolo zaujimave vediet, ci myslia performance alebo high end verziu G92 (alebo co vobec myslia), nakolko bolo G92 1000M zmenenych na - G?? ????M

wendigo5 napísal:nevyzera to na malu spotrebu

ty myslis pravdepodobne "malu" spotrebu G92 - ako performance cipu - vobec sa zatial nevie kolko bude mat tranzistorov, ale veril by som skor chyrom, ze okolo ~ 500 miliónov.

to na obrazku by bol potom high end cip pozostavajuca buD z jedneho jadra, fyzicky = 2xG92 alebo GX2 riesenie s dvoma cipmy.

G92 uz aj podla nazvu nie je high end, ale vyssi mainstream/performance cip
-------------------------

vsetko su to samozrejme spekulacie

keno10

jedna otazocka k comu mi to bude ked je HDD POMALY JAK SVINA

(oproti ostatku) loading zastihne kazdeho

, najlepsie by bolo keby sa uz dokazal PC napojit na hlavu a rezervovat si cast vykonu....

u ati(amd) nevem co si mam mysliet pretoze tie ich ovladace sa stale vylepsuju, ale asi anjdolezitejsie je aby neskrachovali

pokial prebehne este vsetko potrebne bez problemov - tak G92 uvidime v novembri ...

v PCinlife diskusnom fóre sa objavili celkom zaujimave informacie ohlade G92:

G92 (teoreticky by to mohla byt "GT" varianta)
740MHz jadro
64 SPs
256Bit zbernica, GDDR3 @ 1800MHZ (realne) 512MB
Spotreba medzi 7900GS a 7900GTX
9700 3DMark06 @ QX6700
cena medzi: 249 - 299 USD

G98 (

? pravdepodobne nie, nakolko ma byt G98 low entry cip)
800MHz jadro
32SPs
256Bit zbernica, GDDR3 @ 1600MHZ (realne) 512MB
vyssia spotreba ako 8600GTS
7400 3DMark06 @ QX6700
cena medzi: 169 - 199 USD
-------------------------------------

Dalsie spekulacie pochadzaju od PHK, ktory je zdrojom takmer vsetkych leaknutych informacii o grafickych cipoch v nedavnej minulosti

- G92 ma mat 6 clustrov @ 1.6GHz SC (aritmeticky vykon by tak zodpovedal 4 clustrom @ 2.4GHz)
-10 000+ v 3DMark06
- high end verzia ma mat 2xG92 jadra v jednom package (co nedava velmi zmysel, ale mozne to je )
- high end verzia ma mat teda aritmeticky vykon 1TFlops/s
- 800MHz takt pre TMUs/ROPs (jadro)
- G92"GX2" ma podporovatl "hybrid-SLI" - vDaka comu ma byt v 2D prevadzke zapnute len 1 jadro
- ma existovat dokument (PDF) v ktorom sa spominaju menovane data

補充一點別的東西: 兩顆G92該怎麼連接的問題,其實要回頭檢視G8x的TCP、以及CELL computing Board。

從CELL computinb Board的結構上、RSX對XDR存取的效率可以知道,線性對應定址、Cache coherency protocol做得好的話,一個GPU可以非常有效率地對另一個device的local memory作讀寫,所以其實我們大可把多數GPU用線性定址mapping串接起來,所以要是G92有512bit,那兩個G92就可以256bit對外、256bit 互接就可以了;而G92的單晶片產品只要封裝上不實作完整的512bit,就可以達到產品區隔。

當然實體層不見得是這樣,可是觀念應該很接近了:既然上層結構(TCP)已經做了切割,那麼要做的其實只是把TPC--ROP中間這段 corssbar跨晶片連結起來就可以。那麼前端FIFO workload的分配做好之後,所有的動作其實就都與Driver沒有關係,那連結效率就只剩底層I/O overhead,以及分離成兩塊的記憶體造成的一些零碎問題,但是這都可以解決(極端地來說對每塊ROP而言記憶體都是切塊分散的)。

用BR03對這部分也有一個明顯的好處:GPU BIOS其實在BR03上,對Host而言BR03後全部視為同一個GPU,除了能跨越BR03的NVIDIA之外。這對上述的實做法其實也是有利,只是沒搞錯的話目前Q uadro Plex靠的也是BR03,這顆晶片適合這樣像多層星形連結的方式嗎?這倒是還有疑問。

http://www.aiplus.idv.tw/blog/index ... blogId=2

Add other things: how connected the two G92 problem, we have to look back G8x TCP, and the Cell computing Board.

Cell computinb Board from the structure, access to the RSX right XDR efficiency know, linear counterparts addressability, good Cache coherency protocol, a GPU can be very efficient to another device to read and write the local memory, so we can use the most GPU linear addressability mapping rods , so if a slave G92, G92 will be two external pipelines, pipelines can access each other; The G92 single-chip products, as long as the package is not a complete slave, we can meet product segmentation.

Of course, physical layer is not necessarily so, but the concept should be very close: Since the upper structure (TCP) has been done cutting, in fact, do so only to the TPC - this corssbar ROP middle-chip can be linked together. Then the front FIFO workload distribution in place, all the movements are in fact on the no relationship with Driver, the only link to the efficiency of the underlying I / O overhead, and separated into two memory caused some problems piecemeal, but this can be resolved (extreme right, and each block is credited ROP Recalling all of the cuts spread).

BR03 with this part also an obvious advantages: GPU BIOS BR03, in fact, right after all BR03 Host, the same as a GPU, in addition to leap from BR03 NVIDIA. This is in fact the above is beneficial, but it is not mistaken by the Quadro Plex is BR03, this chip for links such as multi-star way? This is also remains in doubt.

zial niektore veci prelozi google cudne ...
napr. 256bit ako "pipeline"

Nvidia G92 is GeForce 8700 GTS

VR-Zone has learned more about the upcoming Nvidia G92 where the official marketing name is most likely to be GeForce 8700 GTS. As revealed earlier, G92 is 65nm based and has 256-bit memory bus width. The reference 8700 GTS card has 8-layer PCB and comes with 512MB GDDR3 1ns memories so the memory clock of G92 is estimated to be between 900Mhz to 1GHz. Core clock is unknown yet.

hmmm :hmm:, zvlastne, keDze G92 ma nahradit G80@GeForce 8800GTS

bude asi existovat viac variant ...

fobos

Ja osobne si myslim ze to bude slabsie ako 8800GTS 640MB lebo:

- ma rychlejsie pamate ale uzsiu zbernicu, bandwidth je nizsi
- ma menej pamete (sice tych 128MB je zanadbatelnych ale ma)
- 64Sp vs 96Sp, ano G92 ich bude mat rychlejsie, ale takt este nie je oficialny
- a dat taketo 2 cipy na jedno PCB staci na novy High-end nVidie

potom by to znacenie aj sedelo. Inak tie info su uz prekrutene neskutocne :bored:

podla mna bude "performance" verzia G92 vykonnejsia ako 8800GTS

- priepustnost a sirka zbernice nie je vsetko, uzsia zbernica sa da lahko vykompenzovat GDDR 4 pamatami

- menej pamate - nie je tiez az take podstatne kriterium a 512MB Vram bude este urcite "performance" karte stacit

- ako si napisal - vyssie taktovane, vDaka 65nm vyrobnemu procesu (pokial sa potvrdi teza "6 clusterov" - rovnako vela, teda 96SPs) = celkovo vyssi aritmeticky vykon ako 8800GTS - a to je velka vyhoda v novsich hrych, nakolko je trend zvysovania pomeru ALU:TEX (pokial nenastali zmeny v shader core)

- texelfillrate skoro rovnaka ako v pripade GeForce 8800GTS, pokial sa nic nezmenilo a G92 ma 4 clustre @ 800MHz, so 6 clustrami @ 720MHz je texelfillrate len mierne nizsia ako v pripade GeForce 8800GTX

fobos napísal:Inak tie info su uz prekrutene neskutocne :bored:

ano, ale to je momentalne ciel nVidie ...

videl som aj informacie ktore uvadza ju vykon vyssi ako 8800Ultra a samozrejme "klasicku" 384b zbernicu ... :o
inak to by sami pacilo viac...

VR-Zone napísal:VR-Zone learned that G92 reference card is 9-inch long, same length as the GeForce 8800 GTS but it is single slot. The PCB version is P393. The 1ns DDR3 memories are arranged in 16Mx32x8PCS array. The card has outputs like HDTV, HDCP, DVI, but no display port yet. Expect some interesting PCB and cooler designs from the card makers at launch.

VR-Zone napísal:VR-Zone got to know that G98 is positioned as a low end GPU and is meant to replace the current G86 (8400 GS) series. G98 supports PCIe 2.0, 64-bit memory bus and will have 256MB of 2.5ns GDDR2 memories. Memory clock is at 800MHz. Performance is expected to be in between G84 and G86 and is meant to compete against ATi RV620, which is 64-bit too.

G92 might be a high mainstream part

Now we can say that we are confused with G9x series. One thing is certain, Nvidia plans to launch the new chips, codenamed G92 and G98, in November, but now it looks that G92 is indeed a high mainstream part.

Our sources now believe that G92 is the 65 nanometre version that is set to replace 8600GTS and that it can easily be branded as the 8700 card. This card should be faster than GTS but slower than 8800 GTS cards and it should cost between 200 and $250 at launch.

If this is true this means that there won't be a Geforce 8800GTX / Ultra replacement in this year.

fobos

2 Crux: Abazovic nevie co pise link (len 3 dni nazad) . Jeden den natrepe aj o nemoznom a na druhy den sa zisti ze vsetko je uplne inak :roll:

wendigo5

2 fobos...Ja by som fudzillu prirovnal k Novemu casu...Takze pre mna doverihodnost= 0.

nV prezentacia, zial v cinstine:

8 cluster, VP2 -> teda G90 ? alebo G80 s integrovanym NVIO? , okrem toho sa hovori o nezavislom taktovani SC ...

slide, kde sa hovori, ze GeForce 8. podporuje D3D 10.1

Zdroj