Recent reports from unofficial sources contradict official statements from NVIDIA which say that the company’s upcoming Fermi/G300 GPU will be available by the end of the calendar year.
Unnamed board partners cited at DriverHeaven and [H]ard|OCP claim that NVIDIA is working to have “some” Fermi sample boards running at the January, 2010 Consumer Electronics Show in Las Vegas. The implication here is that NVIDIA may very well miss that window to demonstrate working silicon to the public.
Of greater concern is the overall state of the Fermi project. If the company is hoping, rather than guaranteeing, to show the first public demo of Fermi in the second week of January, then the G300 may be further behind than anyone had imagined.
After the project slipped its original September/October launch, many believed that G300 would launch in the November/December time frame. That date was based on a July tape-out of the company’s A1 (first) G300 revision.
A tape-out is the finalization of a PCB/IC design; it is at this point that the schematics for the product are sent for manufacture. It typically takes about five months to translate the specs into a physical product. This is why a year-end G300 launch was expected from a July tape-out.
However, a tape-out is not always the last stage of circuit design. The design could fail viability checks at the foundry, or the production board may fail functionality checks in its shakedown cruise. Should either of these scenarios unfold, the circuit designer must commit itself to a new revision–a respin–which takes about 4-6 weeks from debugging to the very first test boards. Should that revision get a green light, you can add another 6-8 weeks for production boards.
Not only does NVIDIA have a history of going retail with A2 or A3 revision silicon, but some reports say that A2 silicon has only just taped out. In a best case scenario, NVIDIA will have A2 back by the first week of December. If A2 is any good, then the very first boards will just barely make the CES window AIBs are talking about. That means retail availability is narrowing in on February, but it could be as late as March or April if the company must spin the A3 rev.
NVIDIA has stayed uncharacteristically quiet about Fermi as it toils to bring the silicon to life. Rival firm AMD is waging an unchecked “NVIDIA is neglecting gamers” campaign which mostly panders to the constituency, but somebody will buy it. Gullible folks are always listening. Follow that with AMD’s 5-6 month lead on DirectX 11 hardware, and it looks like AMD is simply leaving a silent NVIDIA on the sidelines. The “we have PhysX” argument is not going to win it for gamers who care about performance.
What is going on?
NVIDIA’s troubles with realizing Fermi are legion, and many of them are not an indictment of the company’s engineering prowess. First and foremost, the company is packing 3 billion transistors–30% more than Evergreen–into the Fermi’s 40nm die. If AMD’s 2.15 billion transistor Evergreen die is 334mm², then Fermi is near 430mm². That’s Xbox big. That’s Ron Jeremy big. Bringing a die of that size to bear in volume is hard, plain and simple.
Secondly, NVIDIA is a fabless company with no production facilities of its own. This saves a tremendous amount in capital costs (the business case for AMD’s spinoff of Glofo), but it forces a reliance on the expertise of an independent foundry. NVIDIA relies on the Taiwan Semiconductor Manufacturing Corporation for manufacturing duties, and they recently revealed that their 40nm process, the very same used by both Evergreen and Fermi, is a right hot mess. Fermi is already an architecture with yield issues, and TSMC is jamming its thumb in the wound with a crippling drop in throughput.
The last and most flagrant accusation is that Fermi is a leaky design. Sites like SemiAccurate have long railed on the architecture for spewing electrons like a fire hose, which would make Fermi difficult to produce and hot to run. Whether or not those accusations are true is entirely open to speculation, but that’s the going word, so we’re repeating it for the sake of discussion.
The final word
Fermi is a very large and complicated architecture, perhaps the most ambitious in NVIDIA’s history. It’s not terribly surprising that the architecture’s viability for a January demonstration is in question, nor is it surprising that we may not see the card in volume until March. The real question is why NVIDIA is staying mum on its woes.
We have debated the cause of the unusual silence over the last few weeks, and we’ve come to the conclusion that it’s an effort to deny the AMD PR machine additional ammunition, and an effort to preserve shareholder faith. Even if a product is behind, saying how and why it’s behind can more damaging than simply letting it quietly slip its schedule. NVIDIA is already suffering Fermi’s protracted development, and the outfit is smartly refusing to make the situation worse.
At this point, the speculative environment surrounding Fermi would be wise to take a “wait and see” approach. Estimating the performance of Fermi’s radically new design is a total gamble, and cards will appear when silicon is good and ready. If it so happens that it’s ready in time for CES, Icrontic will be there to give you all the pictures and information you can stomach.