Nvidia GT300 core: Speculation

Status
Not open for further replies.
I believe there are 4 rows with 8 SPs each.
The rows appear to be in pairs, which in the earlier diagram indicated SIMDs (2 rows of 4). So two 32-wide SIMDs?

The 4 dots are probably 4 DP units.
The B3D GT200 diagram has little green blobs for these. Seems they've become little blue blobs.

The lack of MI unit is good.

The B3D GT200 diagram incorrectly describes the MAD+MUL units - MUL is part of the MI.
From this GF100 diagram I infer that a cluster now contains SP-MAD and DP-MAD ALUs.

So if single-precision ALUs only consist of MAD this makes the compilation/score-boarding/operand-fetching much simpler. Though the separate DP unit still mucks things up somewhat.

Jawed
 
Yeah, the GT200 diagram isn't a terribly useful reference if you want to guess at SM config for the new chip. I should really fix it, too, to match what the G80 and GT200 articles talked about, rather than what NV put forward.

Blue dots aren't DP units.
 
If blue ain't DP units, then did they go cpu like? 2:1 sp:dp ratio. If so, then it looks like they are reusing sp alu's to do do dp as well.

There seem to be 16 blocks, each block having 32 sp mad's. That's consistent with what we have heard so far. 384 bit bus is new though.
 
The four blue dots are the 4 threads that the ARM core supports which is added per cluster, obviously. Together with a 16-wide vector unit with a predication bit for each lane (that's two dots per lane), of course :p

*SCNR*

Seriously, if Rys is already doing diagrams of GF100 (while those for HD5k are still not out yet?), I'll definitely wait for the GF100 before deciding where to sink my money.
 
Blue dots are SFUs.

If that's the case then it can't be one dot per lane like in the GT200 diagram. Nvidia's diagrams had each SFU unit as being 4 lanes wide. In which case the SP:SFU ratio would have risen to 2:1.

But if this thing really "only" has 512 ALUs, what sort of clocks would it need to be competitive with HD 5870, and more importantly the X2?
 
But if this thing really "only" has 512 ALUs, what sort of clocks would it need to be competitive with HD 5870, and more importantly the X2?

Performance figures provided along with the specifications talk about 70% of a 5870CF setup. So 20% over a 5870.
Hence the notion of a X2, it is required for performance leadership.
 
Performance figures provided along with the specifications talk about 70% of a 5870CF setup. So 20% over a 5870.
Hence the notion of a X2, it is required for performance leadership.

There are performance figures out there too? If so, it's not going to be nearly as bad for them as last generation in terms of perf/mm. Assuming, of course, they can actually release the thing.
 
Performance figures provided along with the specifications talk about 70% of a 5870CF setup. So 20% over a 5870.
Hence the notion of a X2, it is required for performance leadership.
Actually that would be 40% over a 5870, if we assume perfect crossfire scaling.

Personal note: that level of performance over a 5870 seems unlikely to me.
 
Performance figures provided along with the specifications talk about 70% of a 5870CF setup. So 20% over a 5870.
Hence the notion of a X2, it is required for performance leadership.

Actually that sounds pretty lackluster, and this would be the figures from Nvidia right? so in reality it wont even be that good as of course PR cherry picks it's performance data.
 
Last edited by a moderator:
Don't mix up G100, a G9x with G100, a codename pre-dating G200.

Call it NV70 to be safe.

Since so far you've been so damn wrong on more than one accounts, it would be wiser to hold back a bit at this stage and let things evolve. You'd be very surprised to be completely wrong on that one too.
 
If that's the case then it can't be one dot per lane like in the GT200 diagram. Nvidia's diagrams had each SFU unit as being 4 lanes wide. In which case the SP:SFU ratio would have risen to 2:1.
Different SFUs.

But if this thing really "only" has 512 ALUs, what sort of clocks would it need to be competitive with HD 5870, and more importantly the X2?
GT200 had 240 SPs and was faster than 800 SPs from AMD.
GF100 will have 512 (different) SPs, why wouldn't it be faster than 1600 from AMD even if we forget for a minute everything that might have changed in the architecture?
 
SPs nor FLOPs nor any other singled out component define by themselves the performance of any GPU. If it would be that way I'll build you a nice hypothetical scenario with a next generation GeForce that has not 512 but 1024 SPs and still gets slaughtered when it comes to 8x MSAA performance. Is the hypothetical example good enough?
 
So if GirlFriend 100 has a 384bit bus with let's say 1200Mhz GDDR5, then it would be offered a total memory bandwidth that would only be 50% larger than 512bit X 1200Mhz GDDR3 (ie GTX 285)!

AMD went for a 33% increase, so I guess Nvidia could have a better chance to let the theoretical performance doubling (if we are talking about a doubling here) of their new chips, to actually shine!

==================

Do you think that Nvidia will be able to provide something new, regarding their dual gpu card? I mean they don't possess a HT link license or something? What about Lucid? (Here i go with Lucid again! :p)
 
I wonder if BoyFriend 100 agrees....ok this is silly :D

Any body has any idea what prompted the code name change? From GT300->G300->GF100?

I'm not just "any" body mind you, but "the" body...(not reallly). When I said several times in the past that I had heard that after G80 something with a G200 alike codename will follow and then a G100 which was to be the DX11 part it was beaten down each and every time. I know it doesn't make sense but NV apparantely always had reserved the "100" for the X11 architecture.
 
I know it doesn't make sense but NV apparantely always had reserved the "100" for the X11 architecture.
Why? It does actually.
GT200 = GT architecure, v2.0 (presuming that G80 was GT v1.0)
GT21x = GT architecure, v2.1
GF100 = GF architecure, v1.0
With F=Fermi, T=Tesla and G=... Graphics? GeForce?
 
Status
Not open for further replies.
Back
Top