Nvidia castrates Fermi to 448SPs

[evolucion8]

IT LOOKS LIKE we were right about Fermi being too big, too hot, and too late, Nvidia just castrated it to 448SPs. Even at that, it is a 225 watt part, slipping into the future.

The main point is from an Nvidia PDF , first found here On page 6, there are some interesting specs, 448 SPs, not 512, 1.40GHz, slower than G200's 1.476GHz, and the big '6'GB GDDR5 variant is delayed until 2H 2010. To be charitable, the last one isn't Nvidia's fault, they need 64x32GDDR5 to make it work, and that isn't coming until 2H 2010 now.

KEY FEATURES
GPU
 Number of processor cores: 448
 Processor core clock: 1.25 GHz to 1.40 GHz
 Voltage: 1.05 V
 Package size: 42.5 mm × 42.5 mm 1981-pin ball grid array (BGA)
Board
 Ten layers printed circuit board (PCB)
 PCI Express Gen2 ×16 system interface
 Physical dimensions: 4.376 inches × 9.75 inches, dual slot
&#61572; Board power dissipation: < = 225 W
External Connectors
&#61572; Single port, dual-link DVI-I
Internal Connectors and Headers
&#61572; One 6-pin PCI Express power connector
&#61572; One 8-pin PCI Express power connector
&#61572; 4-pin fan connector
Memory
&#61572; Memory clock: 1.8 GHz to 2.0 GHz
&#61572; Interface: 384-bit
&#9679; Tesla C2050
&#8213; 3 GB
&#8213; 24 pieces 32M &#215; 32 GDDR5 136-pin BGA, SDRAM
&#9679; Tesla C2070
&#8213; 6GB
&#8213; 24 pieces 64M &#215; 32 GDDR5 136-pin BGA, SDRAM
BIOS
&#61572; 2Mbit Serial ROM

I don't know if that's completely true in my opinion, but the whitepaper comes directly from nVidia, and I just hope that the card comes with full 512sp instead of 448, so we can have healthier competition and lower prices, but every second that runs out, the Fermi card looks less impressive.

 

[tviceman]

The tesla card is not the consumer card i.e. gtx 380 and therefore will have different specs.

 

[nitromullet]

Weird... That PDF is dated November 2009.

Didn't NV just do a con call about Tesla last week? Was there no mention of the number of processing cores?

 

[nitromullet]

The tesla card is not the consumer card i.e. gtx 380 and therefore will have different specs.

"NVIDIA's Fermi: Architected for Tesla, 3 Billion Transistors in 2010" mentions 512 cores.

http://www.anandtech.com/video/showdoc.aspx?i=3651

 

[Schmide]

Preemptive Wreakage reply

Or maybe it is so fast they felt a need to chop it down to not show their hand to early.

 

[evolucion8]

Tesla is supposed to be the most expensive SKU with the best GPU cherry picked cores, so considering that it may have only 448 SP, will mean that the consumer card will have the same SP count, I doubt that nVidia will use the best cores for consumer cards when the Tesla cards can be more profitable and mission critical. After all, nVidia can scale down and disable SP for cheaper SKU's and get some profits selling it for the regular consumer, pretty much the same bussiness tactics done before.

 

[yh125d]

We can't really definitively decide anything from those specs, as far from launch as they were written they probably couldnt have been very sure about it themselves.


Even if that tesla card does come out just as that sheet says, it could just be like a GTX375 equivalent on the consumer side

 

[thilanliyan]

Maybe the GTX360 is 448?

 

[Keysplayr]

Tesla is supposed to be the most expensive SKU with the best GPU cherry picked cores, so considering that it may have only 448 SP, will mean that the consumer card will have the same SP count, I doubt that nVidia will use the best cores for consumer cards when the Tesla cards can be more profitable and mission critical. After all, nVidia can scale down and disable SP for cheaper SKU's and get some profits selling it for the regular consumer, pretty much the same bussiness tactics done before.

You sound quite sure of this. Mind if I quote you and save it for later?

 

[evolucion8]

You sound quite sure of this. Mind if I quote you and save it for later?

Well, I was just speculating using common sense, but usually in bussiness, common sense isn't used wisely, do whatever do wanna do with the quote, that can make me famous j/k

 

[Idontcare]

You sound quite sure of this. Mind if I quote you and save it for later?

Actually since the OP is quoting entire passages en masse from Charlie's latest semiaccurate article without actually using quotes or giving credit or source references I think the OP just outed themselves as being Charlie.

He's back to his usual completely generic over-reaching statements regarding areas of technical expertise and business management with which he has absolutely zero credentials or experience:

The architecture is broken, badly designed, and badly thought out. Nvidia does not understand the basics of modern semiconductor design, and are architecting their chips based on ego, not science. The era of massive GPUs is long over, but Nvidia (mis-)management doesn't seem to want to move their egos out of the way and do the right thing. Now they are left with a flagship part they can't make.

 

[ViRGE]

Tesla is supposed to be the most expensive SKU with the best GPU cherry picked cores, so considering that it may have only 448 SP, will mean that the consumer card will have the same SP count, I doubt that nVidia will use the best cores for consumer cards when the Tesla cards can be more profitable and mission critical. After all, nVidia can scale down and disable SP for cheaper SKU's and get some profits selling it for the regular consumer, pretty much the same bussiness tactics done before.

You're assuming the issue is bad cores. Note that the card is speced for 225W; perhaps they had to cut off some SPs to make their power budget on the Tesla cards?

They're going to have no problem going to 300W on the consumer cards, if that's the case.

 

[Keysplayr]

Actually since the OP is quoting entire passages en masse from Charlie's latest semiaccurate article without actually using quotes or giving credit or source references I think the OP just outed themselves as being Charlie.

He's back to his usual completely generic over-reaching statements regarding areas of technical expertise and business management with which he has absolutely zero credentials or experience:

Hey, it wouldn't surprise me at all.

 

[tviceman]

"NVIDIA's Fermi: Architected for Tesla, 3 Billion Transistors in 2010" mentions 512 cores.

http://www.anandtech.com/video/showdoc.aspx?i=3651

It's an architectural overview though, not "here is what the gtx 380 will be and here is what the tesla 3800 will be." As we all know, Tesla != Geforce.

And as someone else posted, perhaps they needed to castrate tesla to meet certain a 225w TDP requirement. If that's so, then there really wouldn't be that much of a reason to castrate the gtx 380, so long as it stays under 300w. And yes, if this is all true, then Fermi would be sucking an awful lot of power and would need some major future revisions to get it under control.

 

[GodisanAtheist]

Who's to say the extra SPs were even doing anything? Perhaps Nvidia realized their card was a bit unbalanced and the extra 64 SPs were adding very little to the actual performance of the card while still sucking up juice, pumping out heat and adding a few mm^2 to the die.

448 SP version could be every bit as fast as the 512sp version in the real world, it just sounds bad in print.

 

[toyota]

Who's to say the extra SPs were even doing anything? Perhaps Nvidia realized their card was a bit unbalanced and the extra 64 SPs were adding very little to the actual performance of the card while still sucking up juice, pumping out heat and adding a few mm^2 to the die.

448 SP version could be every bit as fast as the 512sp version in the real world, it just sounds bad in print.

the die wouldnt be any smaller with 448 would it? its designed with 512 so disabling 64 still leaves the same size chip.

 

[mwmorph]

the die wouldnt be any smaller with 448 would it? its designed with 512 so disabling 64 still leaves the same size chip.

It would help yields quite a bit though. TSMC is having yield issues so it could happen.

 

[Idontcare]

It would help yields quite a bit though. TSMC is having yield issues so it could happen.

Look at it the other way, every processor out there has extra cache, on-purpose, for yield-enhancement purposes.

(AMD has popularized the laymen's term "harvesting" but this tactic was in-use across the entire industry for decades prior)

Think like CELL B.E. and the PS3. 8 cores, one is for redundancy so they get the yields up and lower costs.

A 448SP Fermi with an extra 64SP's of redundancy for yield-enhancement will always have higher yields than a 448SP Fermi (albeit smaller in size) with zero SP's for redundancy.

As the process matures the yield differential (which will always favor the chip which has more redundancy) will slowly reduce and eventually become so marginal that the cost differential between the larger chip with the redundancy and a theoretical chip with less (or no) redundancy begins to be a problem that needs solving.

Think Regor versus Callisto when it comes to the two different ways AMD skinned the Athlon X2 and Phenom X2 cat(s).

Since redundancy is a yield-enhancement (read "cost reduction") feature and not a performance-enhancement feature (which would fall under "ASP enhancement") you don't want to go about building in too much redundancy into your IC's otherwise it will actually come to represent an elevated cost structure (the opposite of what you wanted) relative to a hypothetical chip which has less redundancy (smaller, so you can make/sell more).

That NV would harvest Fermi's which have fewer functioning SP's is not exactly unexpected, is it? Especially early on in a product ramp.

 

[Zap]

The architecture is broken, badly designed, and badly thought out. Nvidia does not understand the basics of modern semiconductor design, and are architecting their chips based on ego, not science. The era of massive GPUs is long over, but Nvidia (mis-)management doesn't seem to want to move their egos out of the way and do the right thing. Now they are left with a flagship part TSMC can't make.

Fixed.

So, how can Nvidia management "move their egos out of the way and do the right thing?" What is the "right thing?" I'm not so sure that the author quoted above should be the one lecturing others about ego.

A 448SP Fermi with an extra 64SP's of redundancy for yield-enhancement will always have higher yields than a 448SP Fermi (albeit smaller in size) with zero SP's for redundancy.

This was my first thought as well. All companies who design really complex chips do this. Heck, I'm sure many of us remember unlocking GeForce 6800 LE or Radeon X800GTO.

 

[dflynchimp]

Dialing fermi down to 448sps is an economical move. Dropping to 448sps might effect fermi's competitiveness with AMD's high end, but that's where pricing comes into play, and it improves overall yields and probably means Nvidia is going the old route of 7800GTX->7900GTX, 8800GTX->8800Ultra, except this time with sps as the major variable. They'll bring on the XTPEUltraXtreme's once yields are better. Right now Nvidia just needs to get their DX11 chips out onto the market before ATI soaks up too much of the demand with their 58XX's

 

[akugami]

I thought the same thing as IDC in that the stream processors (SP's) were reduced to 448 for redundancy purposes to raise yields. Much like the Cell as noted by IDC.

In many ways, Tesla is nVidia showing off the GPGPU power of their new architecture so they're not reducing SP's because they felt the card was too powerful. All of us are looking in from the outside so we don't know exactly why the SP's were reduced and whether it was cut out from the final chip design or merely disabled. Likely just disabled to increase yields IMHO.

Regardless, having Tesla, which again is the GPGPU/CUDA flagship, shipping with reduced SP's would actually point to the GTX380 having only 448 SP's initially and likely won't have the 512 SP's until a product refresh.

@tvicemam

If I recall correctly, most of the difference between a pro card like a Quadro and a consumer card like a Geforce is in the drivers. There's usually a difference in the quality of the parts used like better caps, maybe higher layer PCB for added stability, etc.

Previously there were cases where a consumer grade video card could be modded into pro cards with a simple firmware upgrade or in some cases minor hardware changes. Though I believe that since the Geforce 6 days, it's been pretty much impossible to do. Likely nVidia has a few tricks up their sleeves to prevent you turning a Geforce into a Quadro to protect the lucrative prices of the Quadro line of cards.

Either way, Tesla being "reduced" in specs does suggest that we might see lower spec Fermi's. At least initially. But at the same time there is so little concrete information out there that any and all analysis of what Fermi can or can't do or what the final specs will be is just a random shot in the dark. You might hit something but you're much more likely to miss. The only thing we do know is that it's not coming this year.

 

[nitromullet]

It's an architectural overview though, not "here is what the gtx 380 will be and here is what the tesla 3800 will be." As we all know, Tesla != Geforce.

And as someone else posted, perhaps they needed to castrate tesla to meet certain a 225w TDP requirement. If that's so, then there really wouldn't be that much of a reason to castrate the gtx 380, so long as it stays under 300w. And yes, if this is all true, then Fermi would be sucking an awful lot of power and would need some major future revisions to get it under control.

Agreed, but isn't it somewhat reasonable to expect at least one of the Tesla cards to have 512 cores when NV shows off an architectural overview that indicates 512 cores?

 

[BFG10K]

Depending on how powerful the new SPs are compared to the old ones, having 448 of them could still offer very competitive performance. It’s all about the work per clock cycle.

 

[tviceman]

Agreed, but isn't it somewhat reasonable to expect at least one of the Tesla cards to have 512 cores when NV shows off an architectural overview that indicates 512 cores?

Yeah I do agree with this. I think the geforce gtx 380 or whatever will still have 512 SP's, but I also think this entire situation boils down to one of three things: either Fermi is too power hungry in it's current state for workstations, Nvidia is worried about yields with 512 fully functioning stream processors, or as BFG stated the performance of 448 at a certain clock speed has hit their goal for what they wanted out of Fermi initially (I doubt this) which gives them the luxury of not chasing after low yield 512 sp parts.

 

[tviceman]

Depending on how powerful the new SPs are compared to the old ones, having 448 of them could still offer very competitive performance. It’s all about the work per clock cycle.

It's possible, but I think this is more or less becoming a situation of Fermi being "competitive" and not performance leading. Nvidia has an excellent track record though, so even if Fermi is disappointing, I'm sure they'll be able to bounce back with revisions.