A few weeks ago the press was atwitter about Nvidia and Pascal along with some “incontrovertible proof”. You may have noticed SemiAccurate did not write anything on the topic, we just shook our head in utter disbelief.
This disbelief was not about how someone had found a real scoop or the press dug up some interesting news, just the opposite. This “incontrovertible proof” of Nvidia’s Pascal not only existing in silicon but shipping internationally was anything but, in fact it said quite the opposite. These sites who copied their material from a single source all misinterpreted the data, badly, if they even bothered to read it. So badly it made our heads physically throb, dozens of tech sites none of which actually had the barest technical knowledge all professed it in headlines.
But it gets worse than that. Months ago there was a report that Nvidia had taped Pascal out, and it too made the rounds of the ‘tech’ site echo chambers. With dozens of sites all saying the same thing because they all copied from the same source, it had to be right, right? So it was now “proved” that Nvidia taped out Pascal last June or so, someone on Beyond 3D posted it in a forum, 3D Center wrote a story on it, and at least a few dozen more repeated it. Proof Internet style.
So with Pascal “taping out in June”, it would have been an easy task to have silicon back, even if it was non-functional, by early August if not before. There would have been time for two spins and serious debug before CES, again more than enough time to get it booting, and unquestionably enough time to have non-functional die to show to the press and analysts at CES. For some reason Nvidia didn’t show silicon, they lied about it instead.
A few weeks later a whole raft of second tier and worse sites all claimed there was unquestionable hard proof of Nvidia having Pascal silicon in hand. SemiAccurate was a bit curious and read one of the dozens of echo chamber reposts on the subject. This is where our head started to hurt, and it hasn’t stopped since. The proof was in the form of a shipping database from Zauba showing parts with what is supposedly a Pascal part number, specifically BGA37.5×37.5, going from Taiwan to Bangalore. Hard proof, right, how can you question this?
Zauba list from our forum here
There was another listing on Zauba that purportedly has “proof” of some Pascal’s being shipped from Taiwan to Bangalore, but those are even more vague and questionable. Lets put aside the obvious problems of these shipping manifests showing nothing of the sort and having no mention of Pascal or Nvidia anywhere. Lets assume those shipping databases really are Pascal related, and that the 37.5mm x 37.5mm BGA socket materials were actually for Pascal. This means Pascal taped out in June like the forums claimed, Nvidia had Pascal silicon in hand but for some odd reason chose to lie in front of hundreds of press and analysts about it, had time for multiple spins before CES, and silicon is being shipped around the globe in late December, 2011, right? Wrong, dead wrong.
Again assuming this stuff is actually for Pascal, it says quite the opposite. Look at the manifest, they are all for things related to BGA37.5×37.5, or the aforementioned BGA socket. We will just call it BGA from now on to save time. What are these things, and why are they being shipped from Taiwan to Bangalore? Lets take a look.
The first line is a TEC cooler rated at 650W, something the echo chamber is using as proof that this is a big, fire breathing GPU for the enthusiasts. If you have even a vague clue about the technology involved, or did the _SLIGHTEST_ bit of research, two things would be obvious. First is TEC coolers are rated by input power, not cooling ability, and they are horribly inefficient. How inefficient? The last time SemiAccurate studied the tech closely was a few years ago and they were roughly 10% efficient at the time, or could pull 65W of heat out of a device for 650W input. Lets give technical progress the benefit of the doubt and assume efficiency has doubled since then, that brings us to 130W of cooling ability. This pretty much excludes any enthusiast parts, if this cooler is for a Pascal device, it is a low-end part, not the enthusiast part dozens of sites claimed it was. Line one precludes a big Pascal part.
Line two is a guide plate, and what it does is exactly what it sounds like, it aligns the silicon with underlying contacts. You use this to position a smaller die into a larger socket with precision and keep it where it should be when the lid is clamped down. This is used for testing initial silicon you almost assuredly do not want to solder down so you can probe, test and replace things. The next two lines are probes to test said initial silicon and for bringup and debugging. They too are just tools that do what they claim. These three lines show nothing much, they are just common tools used to bring up early silicon and test it.
Line five is a retainer, aka clip or clamp for the socket, a sophisticated, precise, and very costly version of clamp you use on a modern Intel CPU. It holds the silicon in the BGA socket, carefully, and likely with a lot of ports for probing and measuring the chip in question. This line is effectively a lid to hold the chip in the socket. Line six is related to this, it is a thermal head or ‘heat slug’, aka the metal plate fixed on top of all modern CPUs. This one is again more costly, more precise, and likely comes with ways to test the die that your production silicon lids don’t. It is a heatsink, a chunk of metal acting as a thermal sink and buffer. Line seven, again assuming this is all for Pascal, is a water cooling block for the socket in question, a physical cooler just like the thermal head but wetter.
The last line is just what it sounds like, a socket for the mystery silicon, and it is valued at a bit under 500K Rupees. That is roughly $7350 at the time of this writing which pretty much excludes an off the shelf PC related socket. Since we are on the subject the thermal head is valued at a little under $3400 and the water cooler a bit over $1500. Like we said, not typical PC parts, very precise, expensive, and made for testing. This again all proves Pascal silicon exists, right?
Actually no, it proves exactly the opposite. When SemiAccurate saw the above manifests, we immediately thought, “They don’t have silicon”, others had a very different view. So we called around and multiple sources with extensive knowledge of early silicon all said the same thing, these are bring up tools, confirming our initial thoughts. We then asked the obvious next question which we didn’t have a good feel for, “How long before you get silicon do you want this stuff in-house?” They all answered in the same ballpark, a few weeks.
Stepping back a little lets explain the bringup process a bit more. When you tape out a chip it goes to the fab and depending on how much you pay, it goes through the system with priority and is often called a hot lot. This takes between 6-8 weeks depending on the process node, metal layers, and the financial fortitude of the designer. This is a minimum time, it could take longer but with mask costs in the $10M+ range, and time to market usually critical, sane people don’t pinch pennies here.
With early silicon time is critical, every day that is wasted is a day’s delay to market. Silicon doesn’t age well, delays are painful and costly so any company capable of bringing complex silicon to the market lines up everything possible in advance to minimize delays. When the silicon is (usually) hand carried from the fabs to the testing and bringup facility, you want it to go right into testing as soon as you walk it in the door. If there is a problem you want it debugged and the next spin to the fab again as quickly as possible. Each respin will eat at least two months time to market, more than a few are death for a GPU with about a year of market life.
So if you have done bringup before, and Nvidia unquestionably has, you know the drill. You know the size of the device you are going to get back down to the fraction of a mm. You know the pad placement, you know the socket design, and you know the tools you will need to bring it up. Since these tools are custom-made for each chip you order them well in advance of first silicon production and you get them back before you have the chips, hopefully well before.
If the cargo company misplaces something or there is a weather delay or whatnot, you don’t want your multi-million dollar first silicon to sit on the test bench waiting for a heatsink. You get the tools listed above weeks before the silicon arrives, set it up, test them where possible, and fix or replace anything that may need it. Basically you get your proverbial ducks in a row because it costs you almost nothing to do before hand but potentially millions afterwards. No sane company waits on bringup and testing tools, period.
So back to Nvidia and Pascal, assuming these manifests are for Pascal related tools. What does it mean? Does it mean Nvidia unquestionably has Pascal silicon? Nope, it means Nvidia unquestionably does not have Pascal silicon and probably wouldn’t for at best several days after they got these tools in-house. In short these tools shipping pretty much prove that Nvidia was anticipating Pascal early silicon, and not an enthusiast part, as the last few days of 2015 passed. If you assume a day to ship these parts from Taiwan to Bangalore, a few days to set it up, and a few more to test and prep, and a few more just in case something goes wrong, there is no way Nvidia had Pascal silicon when Jen-Hsun claimed they did, and he had to have known it.
It also proves the ‘reports’ of Pascal taping out in June were unquestionably wrong, if it taped out in June there is no way silicon would be getting back to Nvidia in early January, best case. ~7 months is enough for a full production wafer to go through TSMC, a month of debug, and another full production run after that. If you hot lot the silicon as they most assuredly did, there would be room for 3-4 full cycles, more than any GPU in recent memory bar Fermi needed. They would have shown it off months before AMD showed running Polaris’ in early December.
So dozens of sites all echoed the June tapeout date, and since it was on so many sites it was definitely true, only it really wasn’t. Pascal likely taped out in early November and wasn’t back from the fabs when Jen-Hsun gave his CES speech, at least that is what the data shows. The only thing the tapeout story really proved is none of the dozens of tech sites has any real sources and will all copy from anyone without bothering to so much as verify a fact. It isn’t news or journalism, it is plagiarism in a socially accepted form. And it is wrong.
The same happened with the Zauba manifests, again assuming they are for Pascal. These manifests unquestionably prove that there is no Pascal silicon in hand, not that Nvidia would be insane enough to ship such valuable parts via commercial cargo anyway, they prove that it didn’t exist at the time of shipping. Once again if the dozens of sites that reported Pascal silicon arrival had any technical knowledge, any sources, or even the most basic urge to do their jobs and verify the data, they would have known it was false. Literally no site other than SemiAccurate did this, they just copied from the nearest source and posted it as fact. And they were once again all wrong.
This is the sad state of technical ‘journalism’ and ‘news’, no sources, no checking, just copying. A random post on a forum with questionable validity turns into dozens or hundreds of articles, all of which were wrong. If any of the people involved had any sort of technical knowledge, or even read the data they were writing ‘facts’ about, they would know it was wrong. Now do you see why my head hurt?
So what do we have? AMD showed off functional Polaris silicon with multiple working devices in early December. It wasn’t rough, it wasn’t a static demo, it had drivers, cards, and all the things you would expect from non-first silicon. A month later Nvidia did not have silicon and lied to a room of analysts and press about it, and there was no way their CEO would be unaware of such a major milestone at his company. Tech sites took random forum posts, made up stories around them, and then copied each other until they got bored by doing so. None had sources, none had information, none did their job, few if any actually read the data they posted, and they were all dead wrong. Welcome to the Internet, you get what you pay for, and for sites reporting the tapeout and silicon, it isn’t worth the price.S|A
Latest posts by Charlie Demerjian (see all)
- Intel kills off the 10nm process - Oct 22, 2018
- HyperX ships it’s 60 millionth enthusiast memory module - Oct 15, 2018
- Bittware/Nallatech water cools 300W of Xilinx FPGA - Oct 12, 2018
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018