NVIDIA GF100 & Friends speculation

[neliz]

Q4 2011? At least Charlie boy hasn't lost his sense of humour

What good is GF108 over GT218 though? I didn't expect anything in that segment until 28nm. I guess the DX11 tickbox is worth another chip.

Is this a good time to bring up Arun's speculation about GF108, since he was right about GF100 being a 480SP part as well?

 

[corduroygt]

Low end accounts for the majority of GPU's sold, and checkboxes are very important, especially with laptop OEM design wins.

 

[trinibwoy]

Is this a good time to bring up Arun's speculation about GF108, since he was right about GF100 being a 480SP part as well?

What was it? I wouldn't be surprised if GF108 is 16 shaders, 4 TMUs, 1 tessellator, 8 ROPs and a single 8x rasterizer on a 64-bit bus. No ECC, no DP. Token DX11 support is all that's required at that pricepoint. It would be silly IMO for them to make a bigger chip than that.

 

[KimB]

What was it? I wouldn't be surprised if GF108 is 16 shaders, 4 TMUs, 1 tessellator, 8 ROPs and a single 8x rasterizer on a 64-bit bus. No ECC, no DP. Token DX11 support is all that's required at that pricepoint. It would be silly IMO for them to make a bigger chip than that.

Err, I'm pretty sure they'd go for the mid-range before they'd go for the extreme low-end.

 

[neliz]

What was it? I wouldn't be surprised if GF108 is 16 shaders, 4 TMUs, 1 tessellator, 8 ROPs and a single 8x rasterizer on a 64-bit bus. No ECC, no DP. Token DX11 support is all that's required at that pricepoint. It would be silly IMO for them to make a bigger chip than that.

Talking of derivatives... FWIW, I'm currently naively expecting NV's line-up to look something like this:

GF100: 480SP, 384-bit GDDR5 (no SKU with 512SP, some down to 416/448) [Q2]
GF102: 320SP, 256-bit GDDR5 [Q3]
GF104: 160SP, 128-bit GDDR5 [Q2]
GF106: 64 SP, 128-bit DDR3 [Q3]
GF108: 32 SP, 64-bit DDR3 [Q4]

Given how wrong I've been about every single of my family predictions in the last trillion years, I heavily suggest everyone to ignore this or even make a clear attempt to change their own predictions if they seem similar to mine

GF100 was 100% Bingo

 

[trinibwoy]

Err, I'm pretty sure they'd go for the mid-range before they'd go for the extreme low-end.

Maybe but you don't believe GF108 refers to anything but the $40 SKU right? That's not really interesting. What I really want to see is how they handle that huge ALU:TMU ratio in midrange parts where performance is slightly older titles is arguably the most important consideration.

 

[rpg.314]

According to our dear Charlie, nVIDIA has tapped out GF108:



Even with somewhat good news, he cant resist to make it look bad :



http://www.semiaccurate.com/2010/03/17/nvidia-tapes-out-gf108/

At this rate, there is a good chance that atleast one member of the fermi family will be caught in the NI launch cycle.

 

[KimB]

Maybe but you don't believe GF108 refers to anything but the $40 SKU right? That's not really interesting. What I really want to see is how they handle that huge ALU:TMU ratio in midrange parts where performance is slightly older titles is arguably the most important consideration.

Well, the thing I'm thinking here is that typically they have a cascade of parts, starting from the high end and progressing down to the low-end. Now, I suppose it's possible that they have an ultra low-end part winding up, but it seems rather more likely to me that if there were the case, then they'd also have mid-range parts there as well.

Or they could be bucking with past behavior. We shall see.

 

[neliz]

Maybe the current product planning is also based on TSMC's capacity constraints?

 

[trinibwoy]

Or they could be bucking with past behavior. We shall see.

Oh yeah, I don't think GF108 or whatever will land before the mid-range stuff. I was just speculating on its configuration, not launch schedule. I see no reason why GF104/6 wouldn't drop first.

Maybe the current product planning is also based on TSMC's capacity constraints?

Do they have an urgent need to replace GT218? Right now their most vulnerable parts are in the $100-$300 range. Even if it was easier to launch GF108 there's no real reason to do so.

 

[rpg.314]

Well, the thing I'm thinking here is that typically they have a cascade of parts, starting from the high end and progressing down to the low-end. Now, I suppose it's possible that they have an ultra low-end part winding up, but it seems rather more likely to me that if there were the case, then they'd also have mid-range parts there as well.

Or they could be bucking with past behavior. We shall see.

a) The window of opportunity would be smallest for gf108 if it arrives at the end, because of Fusion.

b) They are trying to learn their 40nm lessons (albeit late) with the smallest die they can make and sell.

 

[RecessionCone]

What was it? I wouldn't be surprised if GF108 is 16 shaders, 4 TMUs, 1 tessellator, 8 ROPs and a single 8x rasterizer on a 64-bit bus. No ECC, no DP. Token DX11 support is all that's required at that pricepoint. It would be silly IMO for them to make a bigger chip than that.

I would expect it to be 1/16th of a GF100, which would give it 32 shaders. It would actually be a lot of work for Nvidia to make a part with only 16 shaders, since it would mean a completely different set of schedulers. Making a part with 16 shaders would mean either:
1. Reducing the physical SIMD width from 16 back to 8. This would make the schedulers loop 4 times over a warp instead of 2, changing the way the register file works, etc. Probably not going to happen for a derivative.
2. Removing the second scheduler from the SM. GF100 has 2 schedulers, potentially each dispatching a warp every cycle. In order to get the benefit from 2 schedulers, you need 32 shaders (since the SIMD width is 16).

I think it's highly unlikely, and runs counter to historical precedent, that Nvidia would make these kind of drastic changes for a cut-down derivative, instead it makes more sense for GF108 to have 32 shaders.

 

[rpg.314]

I would expect it to be 1/16th of a GF100, which would give it 32 shaders. It would actually be a lot of work for Nvidia to make a part with only 16 shaders, since it would mean a completely different set of schedulers. Making a part with 16 shaders would mean either:
1. Reducing the physical SIMD width from 16 back to 8. This would make the schedulers loop 4 times over a warp instead of 2, changing the way the register file works, etc. Probably not going to happen for a derivative.
2. Removing the second scheduler from the SM. GF100 has 2 schedulers, potentially each dispatching a warp every cycle. In order to get the benefit from 2 schedulers, you need 32 shaders (since the SIMD width is 16).

I think it's highly unlikely, and runs counter to historical precedent, that Nvidia would make these kind of drastic changes for a cut-down derivative, instead it makes more sense for GF108 to have 32 shaders.

Cedar has 40 alu's instead of the usual 80/simd.

 

[Jawed]

Maybe but you don't believe GF108 refers to anything but the $40 SKU right? That's not really interesting. What I really want to see is how they handle that huge ALU:TMU ratio in midrange parts where performance is slightly older titles is arguably the most important consideration.

If GF100 has only half the TMUs it's supposed to have (half present but turned off), then ALU:TEX on the smaller GPUs may indeed be quite different if they don't have half of them turned off.

Jawed

 

[entity279]

But it can happen that GF108 be ready before other GF10x chips, just like Juniper was a bit ahead of Cypress. So nV just decided not to wait any longer? Or is it a consequence of the (at some point in the past) unanticipated Fermi full respin?

 

[Tchock]

But it can happen that GF108 be ready before other GF10x chips, just like Juniper was a bit ahead of Cypress. So nV just decided not to wait any longer? Or is it a consequence of the (at some point in the past) unanticipated Fermi full respin?

I'm betting on the invisible war with Fusion and Sandy Bridge (the latter being scarier if it just so nicely thwarts Optimus with a lowend discrete).

And probably 3 chips instead of 4- at G92/<G92b (28nm) size GF104 will last for quite some time I guess, rebrands and hodgepodge whatnot.

 

[Silent_Buddha]

But it can happen that GF108 be ready before other GF10x chips, just like Juniper was a bit ahead of Cypress. So nV just decided not to wait any longer? Or is it a consequence of the (at some point in the past) unanticipated Fermi full respin?

It could make sense that GF108 is still on track. Smaller die means the 40 nm problems at TSMC would still yield a fair amount of chips compared to larger dies.

Considering this is roughly 6 months after GF100 should have launched, I'd bet they'll just decide not to hold it back.

Larger chips could be running into similar issues as GF100. And rather than trying to fix all of them at once, they are focusing on fixing GF100 yield issues and then hoping that experience will allow them to quickly get the midrange parts outs.

Add to that GF108 would be mobility friendly, and releasing sooner rather than later gives them a better chance to get OEM notebook wins for the back to school season.

Regards,
SB

 

[KimB]

a) The window of opportunity would be smallest for gf108 if it arrives at the end, because of Fusion.

Uh, why? As near as I can tell, they address quite different markets.

 

[mczak]

Cedar has 40 alu's instead of the usual 80/simd.

Yes, but I think it should be much easier to change simd length of r6xx+ parts (4/8/12/16 length) rather than doing that for fermi. nvidia have not done this for any of their chips, while there were quite a few chips from AMD which did this.
If Arun is spot on, that Q4 date though he had seems a bit too late if GF108 taped out already.

Oh and I must have missed that half tmus disabled rumors. Any credibility to that? That would be very very strange imho.

 

[Jawed]

Oh and I must have missed that half tmus disabled rumors. Any credibility to that? That would be very very strange imho.

I dunno, I can't work out if people are taking the piss each time it comes up

It sounds bonkers, frankly.

Jawed