Does the transistors matter?

[Nicked]

I thought the 'other die' was 100M, 80 mem + 20 logic.

Yeah, ahhhh, I misread your post. I thought you meant 250M+ on the main die and 80M on the daughter die.

 

[The GameMaster]

234m-21m-cache(512*1024*8*6=~25m)-sram(98m) = 90m

In fairness, though, we don't know if there'd be any redundant transistors within cores/spus either. It is a nice illustration of the control logic/core count tradeoff in cell though, they've certainly got a lot out of their transistor budget from a execution pov. And of course, the memory structure is kind of fundamental to its performance. Which I guess just kind of illustrates what everyone's saying here!

Yea I was not too sure about that Arstechnica claim on transistor count... 14 million seems a tad much for 256KB of local storage... should of been more like 12.3 million transistors for 256KB of local storage leaving about ~9 million for the rest of each SPE. In which case the numbers becomes 512KB+256KB*7 or about 2304KB of SRAM on the PS3 Cell which would consume a total of 110.6 million transistors + the other 21 million transistors that is not in use due to a SPE being disabled for yield subtracted from the total 234 million transistors (not the 220 million I previously stated, don't know where I got that number... going to back edit that correctly in a moment) for a total of around 102.5 million transistors that are active (not disabled) and not memory related versus XENON's 116 million transistors that are not memory related. As I said before the non-memory related transistor count for logic is about equal as I confirmed for that other poster, and of course there are other things that probably should be removed from the transistor count such as debug logic, the L2 crossbar in XENON, the EIB bus in the Cell, the FlexIO in the Cell, and the XIO memory controller in the Cell and various other smaller things.

Certainly a lot of smaller details...

But as previously stated... transistor usage is more important that transistor count.

In response to One's question...

Compared to regular memory eDRAM is a *LOT* faster and much lower latency than conventional memory (GDDR3 in the XBox360's case), but it is still slower than L2 memory or the local storage on the SPEs. eDRAM is expensive stuff... but L2 SRAM is a *LOT* more expensive than even that. The eDRAM is suited for the type of work it will be used for in XENOS, but in a processor such as XENON or Cell it has too much latency to be useful because of the work that XENON and Cell will have to do.

 

[Titanio]

Xenos logic should be more than 255 Mtransistors. (edram is 80+ Mtransistors)
Cell DD2 is more than 250 Mtransistors.

I was wondering about this..is there an exact figure public?

The SPU counts might also have changed then too, but most of the changes looked to have been on the PPE.

 

[Shifty Geezer]

I think people are dissecting the figures too much, to the point they're utterly meaningless (instead of mostly meaningless). What's the point in taking away the SRAM transistors to measure logic transitors? Logic on it's own is worthless. You need local storage. Fetching every single insruction from RAM will laughably cripple all that logic. 250 million transistors of logic is gonna be the worst CPU ever, and utterly eclipsed by a similar CPU concept but with 70 million logic transistors and the rest SRAM. Neither can the ratio of SRAM to logic be considered meaningful as it's hw that ratio is utilised that matters. If one chip can get more efficiency from 10 MT SRAM feeding 190 MT of logic, vs a chip with 150 MT of SRAM feeding 50 MT of logic, the higher logic wins, and vice versa. It all depends on what the chips do with their transitors (both logic AND local storage) that determines performance. You can't take SRAM out of the equation.

 

[Bill]

Well, one can think of Xenos as 232+8 rops on the daughter die. So 242 (estimate).

Whereas RSX may not have Purevideo and if 16 ROPS is overkill, then erase another ten for 8 ROPS.

So you might end up with 265 vs 242 in reality


But this was my gripe with EDRAM. Total transistors are 500 versus 550. Not bad. But remove the EDRAM, and it becomes 550 versus 420. A much larger difference.

That's why I just dont like the EDRAM. It's just a bad idea. Shame on ATI (I assume it was their idea).

Now, if the EDRAM proves it's use, I will take that back. But only if 360 games look better than PS3 games, as they should given the respective GPU transistor counts. (X360 has more than PS3)

 

[YeuEmMaiMai]

Transister do count but they are useless if they do not perform efficiently. Generally speaking a part with more transisters will have more functionality and to some extent more speed (usually because of high clock speeds). NV30 was an example of poor design while R300 was an example of an efficient design.

 

[Titanio]

Well, one can think of Xenos as 232+8 rops on the daughter die. So 242 (estimate).

Whereas RSX may not have Purevideo and if 16 ROPS is overkill, then erase another ten for 8 ROPS.

So you might end up with 265 vs 242 in reality


But this was my gripe with EDRAM. Total transistors are 500 versus 550. Not bad. But remove the EDRAM, and it becomes 550 versus 420. A much larger difference.

Not sure where you're pulling these figures from :??

Anyway, as already said, transistor counts alone and out of context are a pretty meaningless metric.

 

[Bill]

The true logic count of Xenos is 257m. The daughter die is 105, only 80 is EDRAM.

However, some of that is for AA. Being fixed function I didn't include it.

However all videocards include ROPS, so they rightly go in the count. I just estimate 10m for 8 rops.

Now purevideo is 25 on Nvidia. 10 for 8 "extra" ROPS. Total = -35m.

Of course, one can speculate purevideo was replaced with 8 more pipes and so on. Suit yourself.

 

[Titanio]

The true logic count of Xenos is 257m. The daughter die is 105, only 80 is EDRAM.

However, some of that is for AA. Being fixed function I didn't include it.

However all videocards include ROPS, so they rightly go in the count. I just estimate 10m for 8 rops.

Now purevideo is 25 on Nvidia. 10 for 8 "extra" ROPS. Total = -35m.

Of course, one can speculate purevideo was replaced with 8 more pipes and so on. Suit yourself.

I guess I was confused by this part:

"Total transistors are 500 versus 550. Not bad. But remove the EDRAM, and it becomes 550 versus 420. A much larger difference."

It seemed like some clockspeed figures might have got mixed up in there or something

The only RSX figure we officially have is the "more than 300m" figure from E3. What that's made up of, or what could be considered redundant or not is anyone's guess.

 

[Shifty Geezer]

If it weren't for the fact Bill's on my ignore list I'd question why transitors dedicated to fixed function AA don't count towards a processors transistor count? What about transistors dedicated to fixed addition? Don't they count? And those dedicated to fixed-function texture interpolation? And those dedicated to moving data from registers to cache?

 

[Bill]

I mean total CPU+GPU transistors on X360=500m

PS3=550m.

However, if 80m is EDRAM on X360..

 

[Brimstone]

But this was my gripe with EDRAM. Total transistors are 500 versus 550. Not bad. But remove the EDRAM, and it becomes 550 versus 420. A much larger difference.

That's why I just dont like the EDRAM. It's just a bad idea. Shame on ATI (I assume it was their idea).

Now, if the EDRAM proves it's use, I will take that back. But only if 360 games look better than PS3 games, as they should given the respective GPU transistor counts. (X360 has more than PS3)

How are more non-eDRAM transistors going to provide bandwidth? GPU's are very bandwidth hungry for certain operations.