Nvidia Tesla C2050, C2070 cut down to 448 shaders

Santa Clara, we have yield problems.

The upcoming Tesla 20 series cards from Nvidia will bring massive improvements over the previous generation of cards, the Tesla 10 series, and was also supposed to be released with 512 cores. This number was not born out of speculation but of actual documents from Nvidia and a big event sponsored by the company, the GPU Technology Conference(GTC), where the "Fermi" architecture was first unveiled.

Going back to November, with no press at all, a PDF was quietly released along with specifications:

Also, the following text is present, which provided less detail about floating point performance than before:

The Tesla C2050 and Tesla C2070 is capable of up to 600 GFLOPs/sec of double precision processing performance. Tesla C2050 comes standard with 3 GB of GDDR5 memory at more than 170 GB/s bandwidth. Tesla C2070 comes standard with 6 GB of GDDR5 memory.

Nvidia has had these plans for some time now, I just did the math the wrong way and counted on the 512 cores that they had announced before. Having to disable two clusters per core on a very high-end card can be attributed to very bad yields on the new chips - this is not a $500, $600 or $700 card, these are $2500+ HPC cards.

On the topic of performance, the previous math and frequency calculations, based on a 512 core card and more precise peak double precision floating point performance:

630 GFLOPs / (256 ops * 2 (FMA)) = 1.23GHz ~ 1.296MHz, a number more commong in Nvidia's graphics cards due to the clock generators generally used. The C2050 card will work at a slower ~1020MHz.

Now, rewind and readjust the parameters as we know it today: 630 GFLOPs / (224 ops * 2 (FMA)) = 1.4 GHz - the number released on the "new" PDF. It was not the rationale that was wrong, Nvidia was hiding some new developments about the said cards. The frequency of Tesla C2050 cards is now also clear in the above spec sheet: 1.25GHz. A 1.4GHz, 512 shader version of the cards would yield 716 GFLOPS/s in double precision and 1.8 GFLOPS/s in single precision, whenever it can come to existence.

This change in specifications seems like a preemptive move with the aim to bring the cards in early Q1 2010, a necessary step to correct what would really seems to point to underlying yield problems on the massive chip. The 512 shader card will inevitably be built but it may require a few more months of work on TSMCs problematic 40nm process that has also caused shortages on AMD's Radeon HD 5800 series cards.