Nvidia castrates Fermi to 448SPs

Unmanufacturable, hot and slow

Nvidia world iconIT 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.

Update: Nvidia has contacted us and declined to respond.

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

Fermi new specs

Fermi – Now less than promised

The rest however is Nvidia’s fault. It designed a chip that was more or less unmanufacturable, and we have been saying as much for more than six months now. It is big, 530mm^2 at the minimum, likely 10+mm^2 more, and way too hot. The 448 SP version is listed at 225W TDP, 190W ‘typical’, and that is with two of the 16 shader clusters clusters fused off. With them on, it would likely be well above 250W, far too hot for a single GPU card.

Then there is the whole problem of 512 SPs promised versus 448 SPs delivered. Fermi is arranged as 16 clusters of 32 shaders, and given that it is turning off 64 shaders, it looks like the minimum granularity it can fuse off is a single cluster of 32. This means it is having problems getting less than two unrecoverable errors per die, not a good sign.

Before you wonder if Nvidia is doing this on purpose, the Fermi based Tesla boards cost $2,499 and $3,999 for the small and large ones respectively. If there was ever a case to cherry pick the good parts, it is for low volume halo parts that sell for about 10 times what a gamer card goes for. Nvidia can not get enough yields for the cream of the crop, so how do you think the mainstream will fare? Look for a green spray-painted X800XTX-PE to not grace shelves before this is all said and done.

The architecture is broken, badly designed, and badly thought out. Nvidia does not understand the basics of modern semiconductor design, and is architecting its 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 the company is left with a flagship part it can’t make.

If Nvidia has a saving grace, it is the upcoming 28nm process node from TSMC, but that is for all intents a 2011 process for any sort of volume. Given that ATI tends to beat Nvidia to any new process node, and does it better, that makes it likely that this “light at the end of the tunnel” will be ATI’s Northern Islands chips running over it, rather than anything that fixes Nvidia’s problems.

Unless Nvidia has a radical new part waiting in the wings for Q2, it really doesn’t have much of a chance to win much of anything. Fermi based cards may beat Cypress by a little, but at a hugely higher TDP and a much higher cost. Architecturally speaking, this generation is basically lost, press stunts aside.

If you are waiting to see the fallout of such architectural snafus, look no further. 448 SPs at 225W TDP, massive yield problems from day 1, and an end result that will be barely manufacturable. Look for more backpedaling soon, count on it.S|A

The following two tabs change content below.

Charlie Demerjian

Roving engine of chaos and snide remarks at SemiAccurate
Charlie Demerjian is the founder of Stone Arch Networking Services and SemiAccurate.com. SemiAccurate.com is a technology news site; addressing hardware design, software selection, customization, securing and maintenance, with over one million views per month. He is a technologist and analyst specializing in semiconductors, system and network architecture. As head writer of SemiAccurate.com, he regularly advises writers, analysts, and industry executives on technical matters and long lead industry trends. Charlie is also available through Guidepoint and Mosaic. FullyAccurate