IT LOOKS LIKE Nvidia is being it’s normal honest self with respect to the company’s high end Tesla compute cards. Yes, the specs on them dropped again, precipitously, and that is from the already castr^h^h^h^h^h downgraded specs released last fall.
If you recall, a year ago, Nvidia was telling people that Fermi would come out in October of 2009 at 1500MHz, have 512 shaders, and only take about 175W. At its triumphant launch during not-Nvision 2009, those specs creeped down a bit, finally finishing off at 1.25GHz-1.40GHz clock, 448 shaders, 1.8GHz-2.0GHz memory clock and only sipping a mere 225W. The ship date slipped from 2009 to Q1 of 2010, then Q2, and if Nvidia liked you, and you were a financial analyst covering its stock, Q3 for anything resembling real quantities.
Customers were not bothered by this change, they took it in stride. Everything was going well, just ask Nvidia. No problems. Can’t make chips? Feh, the architecture is fine on paper. Less than 20 percent yields? Not a problem, just obfuscate when asked about it, and telling the truth seems to be punishable at Nvidia.
Step forward to the ‘release’ Tesla cards, the C2050 / C2070, as seen by the spec sheet here. Remember the spec sheet that was here, but now has a page not found for some reason? Odd. The link may be dead, but the documents are pictured here, and we have saved copies of both that document titled BD-04983-001_v01.pdf and the datasheet titled NV_DS_Tesla_C2050_C2070_Final_lowres.pdf as created on 11/11/2009 and dated “NOV09”. They differ a bit from the more modern ones.
November stats, from Nvidia’s PDF
The new specs are a significant reduction, unless you are wondering about power, then they are higher. If you are an OEM, they are higher still, but lets not quibble about levels of dishonesty, I mean, if the SEC doesn’t care about what Nvidia is telling the analysts and the investing public, why should mere journalists bother to hold it to its statements? It doesn’t give analysts 30″ monitors.
The new stats, slightly reduced for quick sales
You can find all the new literature here, but the one you want is specifically the C2050 / C2070 data sheets from “APR10”, here. Those ‘puppies‘ only run at 1.15GHz, have 515GFLOPS DP FP performance, and 1.03TFLOPS SP FP performance. Memory is at 1.5GHz, and power consumption is now 247W. Here’s an analysis of the slippage.
Last year, November, and now.
What is there to say? The Fermi based compute cards are already a running joke, delivering only 68 percent of the promised performance, 88 percent of the cores at 77 percent of the intended clock speed,for 141 percent of the power. That turns out to be slightly less than half the promised performance per watt, the overridingly critical measure in the compute space.
In the end, Nvidia seems to have delivered about half of what it had promised, less if you consider memory speeds, but it is late, draws lots of power, runs hot and costs far more than hoped per chip. None of these problems are fixable, it is time for a new architecture.
With any luck, Nvidia will get to those favored financial analysts before they realize this. One thing for sure, it needs to get word out before some pesky journalists start raising inconvenient questions about threading, asynchronous transfer capability, and how much CPU time that takes versus what Nvidia promised. If word of that gets out to any analyst who understands the effect this will likely have on sales, things could get mighty awkward for the boys in green.S|A