NVIDIA Maxwell aka ΚΑΙΦΕΣ - άνοιξη 2014

[Psycho_Warhead]

Τερατάκι..

[michael92]

Τερατάκι..

 

Με 5-7000$ price tag το λες τερατάρα του πορτοφολιού. Μεγάλο foul της nvidia το στράγγισμα μνήμης από τις gamer κάρτες για να έχουν λόγο ύπαρξης πλέον οι quadro. Όπως και να έχει οι servers/ workstations κλπ έχουν σταθερό budget το οποίο ξοδεύεται όλο, οπότε quadro σε μικρότερη τιμή θα τους επέφερε περίπου τα ίδια κέρδη.

[Ailuros]

http://www.guru3d.com/news-story/nvidia-gm200gm210-gpu-specs-surface.html

 

λιγο νωρις δεν ειναι?

 

Aπλή αντιγραφή όσων έχουμε υποθέσει εδώ και καιρό. Πρόσεξε στα TMUs λέει 256 ή 192. Aφού ο αρχιλαλάκας ο ίδιος γράφει για 24 SMMs είναι δυνατόν να είναι 256; Εκτός αν με πείσει ότι το κάθε cluster έχει 10.66666666666667 ΤΜUs.

 

Και όχι δεν είναι νωρίς. Με τους πυρήνες μέσω Zauba να εντοπίζονται πέρα δώθε εδώ και μήνες είναι πάνω στην ώρα τους.

[manos08]

Εντυπωση μου κανει που εξακολουθεί να εχει 12gb η quadro...<br>

Με την 980 να εχει 4gb, μια gaming εκδοση του gm200 (σαν 780ti) θα εχει 6 τουλάχιστον... Η αντιστοιχη TITAN ποση θα εχει; Ωστε να μην ακυρωνει την quadro;

[Sakiz105]

Εντυπωση μου κανει που εξακολουθεί να εχει 12gb η quadro...<br>

Με την 980 να εχει 4gb, μια gaming εκδοση του gm200 (σαν 780ti) θα εχει 6 τουλάχιστον... Η αντιστοιχη TITAN ποση θα εχει; Ωστε να μην ακυρωνει την quadro;

8

[manos08]

Με 384-bit Memory Bus, αποκλειεται, αναγκαστικα παμε στα 12?

[Sakiz105]

Με 384-bit Memory Bus, αποκλειεται, αναγκαστικα παμε στα 12?

καλα πηρες μετρητης οτι μαλακια ειπα?

 

6 λογικα θα εχει και θα βγαλουν και μια με 12 για περισσοτερο αρμεγμα.

[michael92]

Calling out ένα comment που βρήκα ενδιαφέρον:

 

 

The truth about finFETs is that they have been a DISASTER for the chip industry. The entire industry, years and years ago, threw their weight behind this dramatic shift in chip design, with ZERO proof that it would work out. Instead, they were bewitched by the idea of being able to claim TWO shrinks that, in reality, never happened.

You are confused- understandably so. Well the con runs like this. Change the way in which you describe your process by quoting the SMALLEST dimension of your VERTICAL FinFET transistor (and yes, I know the 'T' in 'FET' stands for transistor). Now when you first use finFETs, do so on your EXISTING process, but claim a 'shrink'. One shrink for 'free'.

But next you reduce the size of the finFET transistor while keeping the same process. And now, as a result, you proclaim to the world a SECOND process 'shrink'. Hence your free two processes that in reality never happened.

So what went wrong with the con. Well the fabs stopped bothering with researching a true shrink (that requires hyper expensive new tools, equipment and materials) and turned their attention to circuits using finFETs. And the finFET, that had been rejected time and time again in the past because of all its 'issues', proved, surprise surprise, to have even more issues now it would have to be etched smaller. The old transistor tech works, and works well, The new finFETs not so much.

It gets WORSE. As finFETs continued to disappoint, just as they had years ago when first tried, the NEED to lower power consumption turned research attention to improving CONVENTIONAL chip design. FD-SOI was invented, and proved to have astonishing potential. BUT the fabs had 'bet the farm' on finFET, and were determined to make this poor technology 'work'.

You see, maturing FD-SOI would give us the equivalent of two TRUE shrinks worth of certain improvement, but it would NOT solve the eventual problem hit when shrinks get just too small for conventional transistors. Fully Vertical transistors will have to arrive at some point if REAL shrinks are to continue- but this necessity does NOT prove that fully vertical transistors will have the requited functionality to make such shrinks worthwhile- we do not know where the finite limits lie.

So we can have LOTS of jam today, cheaply (if we mature FD-SOI and similar), but at the cost of possible jam tomorrow. Or we can focus entirely on jam tomorrow, with the very real possibility that such future jam will be too expensive, taste horrid, or possibly won't exist at all.

FinFET supporters HOPE the industry will once again solve the new problems they face, but have no guarantee this will be the case (and at this time, the much delayed finFET is proving very difficult to make adequate, let alone better). FD-SOI supporters want us to go back to existing chip production, and improve tools, circuit design, materials, and the form of more traditional transistors to gain a lot more performance out of current processes that are both mature and cheap.

Look at what Nvidia has done with the 980 over their other, earlier .28 parts, by using existing fabrication tech more efficiently.

Confusing matter greatly is the split between mains-powered desktop chips and the ultra-mobile low-power parts. It is more than possible that finFET can be made to work adequately for one but not the other. This poses a real dilemma for fabs.

 

http://wccftech.com/tsmcs-16nm-finfet-faces-delays-qualcomm-jumps-ship-samsung/

 

 

 

[Ailuros]

Τι σχέση έχει το ότι μεταφέρει η QCOM μέρος της κατασκευής της στην Samsung με την ενότητα; Εκτός του ότι είναι άρθρο από αρχιβλαμμένα  του wccftech που μα τον Τουτάτη δεν έχουν ιδέα από που κλάνει η κότα. Καλύτερα να βάλουν κάνα φερετζέ να κάνουν τα γιουσουφάκια και να αφήσουν αυτού του είδους τα θέματα σε εκείνους που έχουν μια στοιχειώδη ιδέα.

 

΄Οτι θέλετε από κει και πέρα για Maxwell στην ενότητα της σειράς 900.