The number of bugs in a chip is relatively proportional to the number of transistors - Bob Colwell, former Intel chief architect
DESIGNER OF WARM GPUS Nvidia has cut the performance of its Fermi based Tesla GPGPU aimed at the high performance computing (HPC) market.
The server vendors Appro and Supermicro recently announced 1U servers featuring up to four Tesla cards. We marvelled at the engineering brilliance of the two firms as they had to overcome the associated power and heat issues that come with graphics cards that use the Green Goblin's Fermi architecture.
It seems those engineering feats were helped, in a manner of speaking, by Nvidia itself, through a decrease in the shader count and clock speeds. The latest specifications detail a 18 per cent clock decrease, coupled with a previously announced 12.5 per cent cut in the number of "stream processors" from 512 to 448, and the board tipped up drawing 10 per cent more power than previous vague estimates that had been bandied about by the firm late last year.
Given that server vendors flog their units by stressing the performance per watt per square foot metric, Nvidia's cuts in performance and increase in power won't help Appro or Supermicro shift too many of their boxes kitted out with Tesla cards.
This is perhaps not surprising given the months of delays that Nvidia has had trying to get its Fermi GPUs out the door. Regardless of who you believe (though we do tend to agree with Charlie Demerjian's analysis over at SemiAccurate), even the most fortuitous yield reports are abysmal and Nvidia's repeated attempts to rein in the power consumption of a chip that has the same thirst for power as a tinpot dictator seem to have failed, with the firm even having admitted that its Fermi GTX480 GPU chip runs hot.
The whole sorry saga is set to hurt Nvidia in its latest push into the HPC market where it has previously done well against its competition. The problem is, with the power requirements being so high and the scale of the performance decrease, the numbers of extra servers that will be required to overcome the shortfall might be too much for prospective customers.
All of this could make Nvidia's efforts to become a big player in the HPC market go up in smoke like an insufficiently cooled Fermi chip. µ
Sign up for INQbot – a weekly roundup of the best from the INQ