Interesting R3x0 information from ati-news.de

[MuFu]

http://www.ati-news.de/HTML/Berichte/Asus/R9800-XT/Asus-Radeon-9800-XT-Seite1.shtml

Der Weg vom Tape-Out zum R360

Der Original R300 wurde im 0.15 Mikron LV Prozess von Taiwan Semiconductor Manufacturing Co. kurz TSMC gefertigt. Der 0.15 Mikron LV Prozess stellt den besten Kompromiss aus Leistung (high Performance) und Verlustleistung dar. In späteren Tests stellte sich heraus, dass durch Erhöhung der Core Spannung in Verbindung mit einer akzeptablen Ausbeute (yields) der Chiptakt weiter erhöht werden konnte.
Durch einen leicht verbesserten Fertigungsprozess, als 0.15 Mikron LVOD bezeichnet, konnte man den Chiptakt um weitere 38 MHz zum Tape-Out mit 287 MHz anheben.

Kurze Zeit später begann ATI mit der Entwicklung des R350. Nach der Valetierung aller möglichen Fertigungstechnologien, inklusive dem neuen 0.13 Mikron Prozess, hat sich ATI dazu entschlossen, den R350 in einem verbesserten 0.15 Mikron LVOD Verfahren fertigen zu lassen.
Die Designer legten beim R350 besonderen Wert auf die Isolierung (vermindern von Leckströmen) und auf verkürzte Signallaufzeiten im Chip. Im Zusammenspiel mit dem 0.15 Mikron HS Prozess konnte ATI den Chiptakt um weitere 55 MHz auf 380 MHz anheben.

Nach 6 Monaten entwickelt sich der 0.15 Mikron Prozess sehr gut und ermöglicht ATI eine sehr gute Ausbeute der Waver. Das bot ATI die Gelegenheit, den Takt weiter zu steigern. Doch man hatte besonders mit Problemen zwischen den Core I/O Pads und der Core Logic zu kämpfen. Beide Teile des Chips arbeiten mit einem unterschiedlichen Takt und nutzen zudem noch andere Transistoren mit etwas unterschiedlichen Antwortzeiten.
Die Ingenieure von ATI konnten das Problem durch die Implementierung der so genannten Delay Lock Loops Struckturen zwischen den beiden Chipteilen im R360 beheben.

Im Zusammenspiel von 0.15 Mikron mit Delay Lock Loops konnte ATI den Takt zur Radeon 9800 Pro deutlich steigern. Alle ATI Radeon Chips durchlaufen zahlreiche Tests, um die volle Zuverlässigkeit sicherzustellen, bevor ATI die Chips zur Produktion freigibt.
Um auch unter extremen Bedingungen den Betrieb der Radeon 9800 XT zu gewährleisten, hat sich ATI für einen konservativen Default-Takt von 412 MHz entschieden, um eine Überhitzung des Grafikchips zu vermeiden.

I think this is the jist of it but feel free to correct me if you have a better understanding of these things (or German, lol)...

R300 initially taped out using TSMC's 0.15u LV process (the best compromise for thermal performance:clockspeed) but was then switched to 0.15u LVOD which yielded a ~40MHz clockspeed boost and still produced adequate yields [so reports of early samples struggling much above 250MHz were correct, if not accurate in their reasoning]. When developing R350, ATi settled on an improved 0.15u LVOD process and concentrated on tweaking the design to decrease leakage currents and shorten signal paths. This contributed towards a further 55MHz increase in the qualified speed. A further 6 months on and the process had improved even further. Yields were good enough for ATi to consider another substantial clockspeed hike. This time the main area that needed attention was the I/O interface, which is of course asymmetrical and involves transistors with different timing parameters.

Now I don't quite understand the last bit, but it seems DLLs are implemented at the interface to remove clock skew and reduce I/O sync delays between the core logic and TL connections. Whether this means they were using analog PLL technology before R360, I don't know. What little EE knowledge I had has all but gone now.

Some of that is new info, I'm sure. Apologies if it isn't.

MuFu.

 

[xGL]

It certainly is new info for me, though I didn't understand half of what I read

 

[demalion]

There was something mentioned by ATI about the 0.15 micron process limitation being in the I/O interfacing, due to physical/electrical characteristics of the packaging/proccess size (AFAIU). The new info here seems to be specifics of what they've done to address that in the R360, and a presumably solid confirmation of the mechanism of clock speed evolution for the R3xx cores in general.

 

[digitalwanderer]

So for us thickies ATi got better at making the R3xx chip as they went along, right?

 

[MrGaribaldi]

So for us thickies ATi got better at making the R3xx chip as they went along, right?

Yes... At least they learned to use the improvements that TSMC made to their .15 process...

The way from the Tape Out to the R360 That original R300 was manufactured briefly in 0.15 the Micron LV process by Taiwan Semiconductor Manufacturing CO TSMC. The 0,15 Micron LV process represents the best compromise from achievement (high performance) and energy dissipation. In later tests it turned out that by increase the core tension in connection with an acceptable yield (yields) could be further increased the chip clock. By an easily improved manufacturing process, when Micron LVOD designates 0,15, one could raise the chip clock around further 38 MHz to the Tape Out with 287 MHz.

Short time began ATI later with the development of the R350. After the Valetierung of all possible production technologies, including new 0,13 the Micron process, ATI decided to it, the R350 in improved 0,15 a Micron LVOD procedure manufactures to leave. The designers attached particular importance with the R350 to the isolation (to decrease of leakage currents) and on shortened signal running times in the chip. In interaction with that ATI the chip clock could raise 0,15 Micron HS process around further 55 MHz to 380 MHz.

After 6 months the 0,15 Micron process develops very well and makes for ATI possible a very good yield of the Waver. That offered the opportunity to ATI to increase the clock further. But one had particularly to fight with problems between the core I/O Pads and the core Logic. Both parts of the chip work with a different clock and use besides still different transistors with somewhat different periods of reply. Engineers of ATI could repair the problem by the implementation of the Delay in such a way specified LOCK loop Struckturen between the two chip parts in the R360.

In the interaction of 0.15 Micron with Delay LOCK loop could increase ATI the clock clearly to the Radeon 9800 pro. All ATI Radeon chip goes through numerous tests, in order to guarantee the full reliability, before ATI releases the chips for production. In order to ensure also under extreme conditions the enterprise of the Radeon 9800 XT, ATI decided for a conservative default clock of 412 MHz, in order to avoid an overheating of the diagram chip.

[EDIT] layout of translation [/EDIT]

 

[MuFu]

I wonder how much of a gamble switching from a general purpose LV process to LVOD really was. Obviously it would have been more expensive, so when you consider that an R300 at 287MHz isn't exactly a bad performer, they really took the bull by the horns when they switched. Must have been more than a few drunken office parties after it qualified at 325MHz.

Do you think people would stilll grumble about the "pathetic" clockspeed increase from 9700 Pro=>9800XT if it had actually been 287MHz=>432MHz? That's a full 50% faster at the same process tier!

Incidentally, I think "OD" stand for over drive. You see even the 9700 Pro had it.

MuFu.

 

[digitalwanderer]

Incidentally, I think "OD" stand for over drive. You see even the 9700 Pro had it.

What? Please elaborate!

 

[MuFu]

Incidentally, I think "OD" stand for over drive. You see even the 9700 Pro had it.

What? Please elaborate!

...and ruin the engineering-related, "humeur de geek"? Never!

MuFu.

 

[Arun]

Must have been more than a few drunken office parties after it qualified at 325MHz.

Can you say "free champage and caviar"?


Uttar

P.S.: WTF is a "humeur de geek"? Isn't it supposed to be "humour of geek?" Or am I just plain stupid? Cause "humeur" = "mood"...

 

[xGL]

I don't understand geek mood/temper either

 

[MuFu]

Must have been more than a few drunken office parties after it qualified at 325MHz.

Can you say "free champage and caviar"?

Heh, well it's ATi so the celebrations probably consisted soley of sireric wearing a party hat, distributing bagels.

P.S.: WTF is a "humeur de geek"? Isn't it supposed to be "humour of geek?" Or am I just plain stupid?

That must be it, yes.

AFAIK, the LV-OD process is designed specifically for higher clockspeeds. The "OD" bit appears to stand for "over drive", but there's very little documentation pertaining to it. Not OVERDRIVEâ„¢, of course.

I assume R360 uses an even further-evolved 0.15u LV-OD process, although it doesn't actually say so explicitly in that article.

MuFu.

 

[Ailuros]

Pardon my french, but your french apparently sucks just as much as mine

 

[MuFu]

Bah, I'm sure we're both perfectly fluent in "Franglais" and that's understood worldwide (although not in Belgium, it would seem).

MuFu.

 

[Arun]

Bah, I'm sure we're both perfectly fluent in "Franglais" and that's understood worldwide (although not in Belgium, it would seem).

MuFu.

Bah, anything english-related in Belgium is a joke. Don't tell me about it. We got 4 hours/week english at school, and our geography teacher who got 1 hour/week advances way faster. Furthermore, in Belgium, you only begin learning english when you're around 14 years old...

I suspect Burkinafaso's national education is better that ours, actually. Oh well.


Uttar

 

[sireric]

OD refers to voltage overdrive, versus a regular LV process, I believe (can't double check right now).

Until you actually get ready to "tape out a chip" changing technology or adjusting the current one isn't particularly bad. If you're going to a faster tech, then your "clock speed" is just going to be faster. You do have to worry that it doesn't get too fast in some cases, but those are reasonably easy to fix (hold time problems).

Party hat and bagels -- How did you know that? Did you see the pictures?

;-)

 

[T2k]

OD refers to voltage overdrive, versus a regular LV process, I believe (can't double check right now).

Until you actually get ready to "tape out a chip" changing technology or adjusting the current one isn't particularly bad. If you're going to a faster tech, then your "clock speed" is just going to be faster. You do have to worry that it doesn't get too fast in some cases, but those are reasonably easy to fix (hold time problems).

Party hat and bagels -- How did you know that? Did you see the pictures?

;-)

I assume ATI done these fixes... How long did it take - in case of R300-350-360?

 

[sireric]

I'm sorry, but I don't understand the question.

If you mean "to switch technology", well, in a stdcell design, the biggest work is on the library side, assuming that the new lib is only getting faster, and things are not drastically different.

 

[MuFu]

Thanks for the clarification Eric. I haven't seen the party pics, regrettably.

MuFu.

 

[nelg]

powered by ATi Hoseron 9800EH

 

[Tridam]

Bah, anything english-related in Belgium is a joke. Don't tell me about it. We got 4 hours/week english at school, and our geography teacher who got 1 hour/week advances way faster. Furthermore, in Belgium, you only begin learning english when you're around 14 years old...

I suspect Burkinafaso's national education is better that ours, actually. Oh well.

In Belgium nearly everyone can begin learning english around 12 years old Unless they choose dutch or german instead of english.

Anyway I fully agree about our poor english learning. I'm sure I can speak a better english than my teacher when I was 16-18 years old I don't think this is specific to Belgium. I think it's also the case in France.

Regarding our global education, I'm sure it's far better than Burkinafaso's one. It's not so bad