NVIDIA Fermi: Architecture discussion

[Razor1]

You think they're saving some high-end bins for the graphics card over the Tesla?

nope thats only for Tesla. Telsa has much more power contraints then Geforce, reasons amount of ram and closed casings and cluster configs.

If anyone listened to the web conference about Tesla, there was mention about power usage and the flop numbers given in recent documentation, but it has nothing to do with the other lines of cards.

 

[compres]

So they will sell the 512 alus version for 500$ and the salvage part for 2500$?

 

[mboeller]

So they will sell the 512 alus version for 500$ and the salvage part for 2500$?

Yep! Thats how a few think it will work. Sell the scrap parts for US$ 2500-3500 and the good parts for US$500.

 

[Razor1]

Yep! Thats how a few think it will work. Sell the scrap parts for US$ 2500-3500 and the good parts for US$500.

These scrap parts are validated at much stricter levels then the other lines because most of them will be used for a 24/7 on rate.

 

[seahawk]

The more intresting question is if we will ever see a non castrated version of the chip hit the market in any numbers or form. Fermi must be the biggest blunder @ NV since NV30 at least, but it could be easily worse.

 

[neliz]

The more intresting question is if we will ever see a non castrated version of the chip hit the market in any numbers or form. Fermi must be the biggest blunder @ NV since NV30 at least, but it could be easily worse.

I have no doubt we'll see the full 512 part on the market, only question is how much cheaper than the HD5970 they will sell it.

 

[compres]

I am just wondering how much a single alu cluster lowers power compsumption. I suspect this count is more realted to yield issues than to power limits.

Edit: A simple linear assumption would yield very little power reduction: 448/512 = 0.875

But then this is only for the ALU related logic. How much of the rest of the die would they cut as a result of this if anything?

 

[Kaotik]

I am just wondering how much a single alu cluster lowers power compsumption. I suspect this count is more realted to yield issues than to power limits.

Edit: A simple linear assumption would yield very little power reduction: 448/512 = 0.875

But then this is only for the ALU related logic. How much of the rest of the die would they cut as a result of this if anything?

According to the PDF the max board power is still the same <=225W

 

[Silus]

I have no doubt we'll see the full 512 part on the market, only question is how much cheaper than the HD5970 they will sell it.

That's the point, because if it's very near it performance wise and it's priced lower, the HD 5970 will have to come down in price

 

[Silus]

I am just wondering how much a single alu cluster lowers power compsumption. I suspect this count is more realted to yield issues than to power limits.

Edit: A simple linear assumption would yield very little power reduction: 448/512 = 0.875

But then this is only for the ALU related logic. How much of the rest of the die would they cut as a result of this if anything?

From the specs on the pdf, nothing else seems to be cut, except the number of Stream Processors.

 

[neliz]

As usual, yes. There was never any promise that Tesla parts would have 512 SPs.

the GTC conference clearly says: Over 2x the cores (512 total) one of the 4 bullet points of the Fermi architecture is the 512 Processor Cores.
Andy Keane (GM Tesla) said Fermi has 512 Cores. What's the use talking about their next gen Tesla has 512 Cores, while it doesn't?

The theoretical performances in the whitepapers are also based on the 512/256 FMA's

 

[DegustatoR]

the GTC conference clearly says: Over 2x the cores (512 total) one of the 4 bullet points of the Fermi architecture is the 512 Processor Cores.
Andy Keane (GM Tesla) said Fermi has 512 Cores. What's the use talking about their next gen Tesla has 512 Cores, while it doesn't?

Fermi is not a product.

 

[Alexko]

As usual, yes. There was never any promise that Tesla parts would have 512 SPs.

Oh come on! Now you're just splitting hairs in 4. Nvidia has talked several times and at several computing events about Fermi's 512 "Cuda cores" and how that's just so great for GPGPU.

 

[Silus]

the GTC conference clearly says: Over 2x the cores (512 total) one of the 4 bullet points of the Fermi architecture is the 512 Processor Cores.
Andy Keane (GM Tesla) said Fermi has 512 Cores. What's the use talking about their next gen Tesla has 512 Cores, while it doesn't?

The theoretical performances in the whitepapers are also based on the 512/256 FMA's

The full chip has a max of 512 Stream Processors. That's all they said. Where is the quote saying that the Fermi chip used in Tesla parts was the full chip ?

 

[Silus]

Oh come on! Now you're just splitting hairs in 4. Nvidia has talked several times and at several computing events about Fermi's 512 "Cuda cores" and how that's just so great for GPGPU.

How is that splitting hairs ? Must assumptions that Tesla would use the full Fermi chip, be considered "facts" now ? I'm not the owner of "lessthanaccurate" you know
Until now, there was nothing that proved that Fermi based Teslas would use the full chip.

Also, everyone assumed that Fermi missed its target clocks, based on the announcement of Tesla parts, which had lower DP capabilities than expected. We know that Tesla parts go through a much stricter validation process, which usually makes them be clocked much lower than their GeForce counterparts. Now, we also know that Tesla won't be using the full Fermi chip which lowers its DP capabilities even more (448 SPs @ 1400 Mhz = ~627 GFLOPs)

 

[DavidGraham]

the GTC conference clearly says: Over 2x the cores (512 total) one of the 4 bullet points of the Fermi architecture is the 512 Processor Cores.
Andy Keane (GM Tesla) said Fermi has 512 Cores. What's the use talking about their next gen Tesla has 512 Cores, while it doesn't?

The theoretical performances in the whitepapers are also based on the 512/256 FMA's

Those could be the lower models of Tesla cards , they could be planning to release higher models further down the road .

It seems that they are saving all 512-core parts for Geforce GTX380 , and planning to use the 448-core parts for Tesla and GTX360 .

 

[dkanter]

Fermi is not a product.

As much as I am loathe to admit that I agree with someone as oblivious as Degustator...even a broken clock is right twice a day.

Fermi is a microarchitecture, and that is different from productizations. Nehalem as a microarchitecture has SMT, but not all SKUs do, etc. etc.

It's dishonest bait and switch type stuff in one regard, but ultimately if NV hits their performance numbers, does it matter if they chose to disable a few SMs? I don't think it's the end of teh world, but it sure doesn't bode well...

DK

 

[Sontin]

The full chip has a max of 512 Stream Processors. That's all they said. Where is the quote saying that the Fermi chip used in Tesla parts was the full chip ?

That's right. This is from the nVidia Tesla site:

Based on the new NVIDIA CUDA™ GPU architecture codenamed “Fermi”, the Tesla™ C2050 / C2070 computing processor is designed from the ground up for high performance computing. It supports “must have” features for technical and enterprise computing including C++ support, ECC memory for uncompromised accuracy and scalability, and 8X the double precision performance compared Tesla 10-series GPU computing products.

http://www.nvidia.com/object/product_tesla_C2050_C2070_us.html

So how many guys will complain when the Geforce cards come without ECC support?

 

[no-X]

The full chip has a max of 512 Stream Processors. That's all they said. Where is the quote saying that the Fermi chip used in Tesla parts was the full chip ?

Did ATi present RV770 as a 900 SPs part?

 

[Silus]

Did ATi present RV770 as a 900 SPs part?

Uh?
What does that have to do with Tesla not using Fermi's full chip ?