NVIDIA GF100 & Friends speculation

[DavidGraham]

"Pinning" refers to the number of power and/or I/O pins dedicated to the memory controller on the chip/package.

so I guess the reason for lower clocks is the insufficient power pinning ?

could I/O pinning affect clocks too ?

 

[neliz]

so I guess the reason for lower clocks is the insufficient power pinning ?

could I/O pinning affect clocks too ?

I/O could affect it too. basically, what I have heard (and this is from months back and everyone blasted me for it) that the controller is basically the same as in the GT215 and at high frequencies it causes way to much interference, presumably because of leakage.

 

[mczak]

I/O could affect it too. basically, what I have heard (and this is from months back and everyone blasted me for it) that the controller is basically the same as in the GT215 and at high frequencies it causes way to much interference, presumably because of leakage.

What I don't quite understand though why didn't they fix the problems? Even if Fermi were on time, the GT215 would have appeared quite a bit earlier (announcement for first mobile parts was in June 2009, including clocks and power draw, so presumably nvidia had final silicon by then). I doubt they thought that praying or cursing would magically fix it? Or would that still have been insufficient time to fix for GF100?

 

[KimB]

What evidence is there that there is something wrong with the memory controller?

 

[Psycho]

There was a statement from nvidia in one of the reviews (dammit, can't remember which) that the memory controller couldn't run as fast as intended.

 

[neliz]

What evidence is there that there is something wrong with the memory controller?

Nothing for now, besides no card able to reach the max. ratings of the GDDR5 chips they're equiped with, this goes for all 40nm and GDDR5 parts.

someone correct me if I'm wrong, but I haven't seen it yet, I think the GT215 also has a 900Mhz MemClock for GDDR5.

There was a statement from nvidia in one of the reviews (dammit, can't remember which) that the memory controller couldn't run as fast as intended.

Use the Google Luke! That'd be a swell statement to have.

 

[KimB]

The memory clocks also might be lowered somewhat to keep power consumption under control.

 

[Arty]

From the hardwareluxx review:

Looks like a different result than the one shown in Nvidia PR video, no?

 

[Malo]

From the hardwareluxx review:

Looks like a different result than the one shown in Nvidia PR video, no?

Doesn't say which version of Unigine benchmark they tested on. Might not be the Nv edition?

edit: the smoke from the screenshots indicates v2 right? Don't remember smoke from v1.

 

[Mintmaster]

Tool of fool?

'of' or 'or'?

I know English isn't your first language, so FYI it's a saying to call someone a tool when they're being used or strung along by someone else.

 

[neliz]

The memory clocks also might be lowered somewhat to keep power consumption under control.

It shouldn't be that much really, OCing a HD5870 with 50Mhz on the Core and the Memory increases the load consumption by 8W. I wouldn't know why NV's MC would use a lot more than that.

 

[Psycho]

Ahh.. was Anandtech

Given the 384-bit bus, we initially assumed NVIDIA was running in to even greater memory bus issues than AMD ran in to for the 5000 series, but as it turns out that’s not the case. When we asked NVIDIA about working with GDDR5, they told us that their biggest limitation wasn’t the bus like AMD but rather deficiencies in their own I/O controller, which in turn caused them to miss their targeted memory speeds. Unlike AMD who has been using GDDR5 for nearly 2 years, NVIDIA is still relatively new at using GDDR5 (their first product was the GT 240 late last year), so we can’t say we’re completely surprised here. If nothing else, this gives NVIDIA ample room to grow in the future if they can get a 384-bit memory bus up to the same speeds as AMD has gotten their 256-bit bus.

 

[Xenus]

Mem overclock

Wouldn't this kinda shoot a hole in that theory while I haven't seen many sites do an overclock they got around a 25% overclock on the mem.

 

[neliz]

Mem overclock

Wouldn't this kinda shoot a hole in that theory while I haven't seen many sites do an overclock they got around a 25% overclock on the mem.

It would if it would actually come close to what AMD runs it memory at, right?

Also, underclocking memory seems to be disabled?

 

[mczak]

Mem overclock

Wouldn't this kinda shoot a hole in that theory while I haven't seen many sites do an overclock they got around a 25% overclock on the mem.

If it would be consistently. Also note even in the link you quoted it's still below the rated max of the memory. Also seen the GTX470 memory oc on the next page? Memory barely overclocked at all.
Those guys here, http://ht4u.net/reviews/2010/nvidia_geforce_gtx_480/index17.php also only barely broke the 1Ghz memory clock barrier on their GTX480 (they generally seemed to have got a not so good sample, doesn't overclock, uses 320W in Furmark, very loud...).
Maybe there's a fair bit amount of variation, some cards being able to reach a bit higher memory clock than others, but in any case not really anywhere close to what AMD achieves with 5 series cards and the same memory chips.

 

[Dave Baumann]

Mem overclock

Wouldn't this kinda shoot a hole in that theory while I haven't seen many sites do an overclock they got around a 25% overclock on the mem.

Having something OC a single board and run a test showing few visual errors is very different from getting multiple thousands of boards pass production diagnostics.

 

[Xenus]

Obviously but that was the only review when I was looking through them quickly that I saw an overclock on.

 

[rpg.314]

What evidence is there that there is something wrong with the memory controller?

Their memory clocks are ~25% slower than the competition, for an industry standard memory interface, than has been in mass production for ~2 years now.

And yes, you can never have "sufficient" mem bw.

 

[rpg.314]

The memory clocks also might be lowered somewhat to keep power consumption under control.

Doubtful. AFAIK, higher memory clocks would have happily compensated with better perf for lower hot clocks while staying within power budget. IIRC, something like it was done on rv770.

 

[KimB]

Ahh.. was Anandtech

Ah, okay, that makes sense. Well, this gives the refresh and derivative parts quite a bit of hope, then. I mean, if the GTX 4x0's are basically underwhelming but good enough, then this means there's still quite a bit of headroom for later parts.