AMD’S ‘NORTHERN ISLANDS’ family of GPUs was perhaps the most frustrating product I have ever had to dig out information on. The reason for that is a long and complex development tale that has more twists and drama than your average mystery novel.
When AMD let out some information during the NDA NI press briefing in Los Angeles a few months ago, I had to go around congratulating some of the more devious folk at AMD on what they did. It went something like this, “You did a good job with keeping this thing under wraps. F[bleep] you very much. Well done, but f[bleep] you very much”. I didn’t [bleep], they just grinned.
Dave Baumann’s eyes covered to protect his identity
The story goes way way back to the fall of 2009 when we first heard the name Northern Islands. We were told some names and details about the line that came from a very good source, but as always, we try and verify before we print any code names. This took months, and you will notice the lack of code names in that article, mainly because we weren’t sure what the heck was going on at that time. Our confusion was quite warranted though.
TSMC canned their 32nm process last fall as we exclusively reported. To say the chipmakers were not pleased was an understatement. When we first heard about the news, we started our usual blitz of confirmation emails and calls, and were quickly greeted with the same story from several people at several companies.
They all said the same thing, that TSMC had canned 32nm with basically no warning, and had severely screwed with the lives of multiple product teams at multiple companies. One GPU maker said, “Those [long string of creative expletives that would make a Russian sailor blush], they really screwed us over. We were two weeks away from tapeout.” That now ex-product was ATI’s Northern Islands, Ibiza to be specific.
As you would expect, there was mild panic in the hamlet of DAAMIT, quickly followed by a lot of meetings. Some of Northern Islands was just killed by external forces, so what do you do to fill the late 2010 product lineup? The product that went bye-bye was Ibiza. Cozumel and Kauai died under other mysterious circumstances. These three were replaced by the ones you now know, love and can now buy some of, Barts, Cayman, Turks and Caicos.
For a while, the new chips were family code name free, and then three new code names started making the rounds, NI-32, Sourthern Islands, and a new and completely different Northern Islands, sometimes called NI-40. To make matters worse, some sources said that Northern Islands were VLIW4 shaders, others that are normally as equally well informed said it was the same old VLIW5. Yet others said it was a hybrid of VLIW4, 5, and parts of the front end from several families.
Sorting through this was beyond frustrating, clear as mud would have been a welcome step up from the info coming out of AMD. You could get completely contradictory information from people in a position to know, and they would insist on the accuracy of the information. They had a story, and stuck with it. The problem was that they were all right, a journalistic worst case scenario.
As you know by now, Turks, Caicos and Barts are hybrid parts, they use most of the NI-32 uncore, like the improved tessellator, newer memory controllers, and other updates, coupled to the older VLIW5 shaders. Cayman, and only Cayman, uses the new VLIW4 shaders with a front end that is more or less what Barts uses, with a few added features. ‘New’ Northern Islands (NNI) is a family composed of two distinct DNA lines.
The confusion at SemiAccurate stemmed from the fact that not a single source, in about a year of leaks and exchanges, mentioned that the ‘family’ was in fact two distinct lines. B[bleep]. In fact, during the aforementioned Barts briefing, when the part was announced, I was stunned when they said the GPU was a VLIW5 architecture. Days before, I still thought they were all VLIW4 + the Barts front end.
That “Oh crap, I got this one _SO_ wrong” feeling lasted for about 30 minutes until they showed this slide.
Clarity through slideshows
Things went from feeling really bad to feeling really confused. Really really confused. It took a few minutes to realize what had happened, and by then I was obliged to shut up about the chips until launch. To say that it sucked is a severe understatement.
How did the two disparate lines come in to being? That one is easy, it is all TSMC’s fault. No, not the knifing of the 32nm process, but the pricing of the 32nm process before it was turned into sushi while likely ending a few careers. TSMC priced the process too high.
To step back a little, a foundry normally wants a customer to move to the new process, it tends to make them more money. Foundries make it a win/win by pricing the new wafer a bit cheaper than the cost of making the same number of transistors on the older process.
If a full process shrink halves the die size of a chip, the new process would be priced at a little less than twice the cost. This allows a chip designer to save a little money on the same effective area, and possibly get a little more performance out of the deal. Some people refer to this as price per transistor, others as price per area equivalent, but it all works out the same, the new process is cheaper per chip.
This time though, TSMC didn’t follow the usual rule of thumb, they priced 32nm higher than 40nm, presumably to cover ever increasing costs. Fair enough. For a bleeding edge chip that can’t be made on an older process, you have to use the new process, it simply can’t be economically made on the older one. This may be because of die size, yield, power consumption, or board area constraints, but is is almost always the reason why high end GPUs must shrink.
If the shrink isn’t available, then the high end chips just don’t happen, or more to the point, they can’t happen. One nice side effect of AMD’s ‘sweet spot’ small die strategy is that the company can skip a process half-step, for example 40nm shrunk to 32nm, and still make the design work on an existing process.
That is exactly what happened to Ibiza, the chip you now know as Cayman. Ibiza was the top end NI-32 chip, and it was backported to TSMC’s 40nm process. Instead of being about 2/3rds the size of Cayman, about 250mm^2 on 32nm, it ended up 50% larger at 389mm^2 on 40nm, but was ironically still cheaper to make than Ibiza on 32nm. Cayman did lose a lot of efficiency, so it probably uses 10-20% more power than Ibiza would have, but that is life. Alternatively, it lost some clocking headroom to heat.
If ATI had taken a large, inefficient chip strategy, this simply wouldn’t have been possible, and instead of a new generation of chips, they would instead be producing a bug-fixed older generation with barely any new features. Instead they got a real new chip. If they had been up against a power wall, instead, it would have resulted in the same thing. Luckily, ATI has good management with foresight, and it was a minor stumbling block.
That explains Cayman, but why weren’t Cozumel and Kauai shrunk too? Why does NNI/NI-40 have four members while NI-32 only have three? That one comes down to pricing. Even before the 32nm knifing, AMD ran the numbers, and realized that for a mid-range chip, the VLIW4 architecture, while completely doable on 32nm, would not be as cost effective as one made from selectively borrowing from NI-32, and backporting to 40. Cost per area equivalent bit TSMC, and cost them a lot of wafer starts.
Long before TSMC surprised the world about 32nm, or lack thereof, Cozumel and Kauai were killed and replaced by Barts, Turks and Caicos. In a nice bit of synergy, when 32nm was killed and the decision to backport Ibiza was made, a lot of the key pieces of the chip were already ported, or well into the process of being ported to 40nm. Give and ye shall receive.
So in the end, Barts and smaller chips were not really delayed by the 32nm knifing, but Cayman was. Instead of the usual new product introduction schedule of big chip, followed by successively smaller ones, the mid-range part came first. Instead of a September Cayman launch, we got a December Cayman launch. Barts came out before Cayman, basically on schedule, and the rest will come out in Q1/2011.
It all worked out tidily enough, but you can be sure that there were more than enough panic attacks, heart attacks, things thrown in rage, and enough long meetings that no one really enjoyed to fill up many an ATI product planner’s schedule. But it did work out.
From the muck-raking journalist side of the equation, it didn’t work out so well. You had two different product lines with the same name, sometimes without a name, the successor product line, SI, with a very similar name, products moved across fabrication processes, and two completely separate product lines with the same name. Frustrating? Yes. Conflicting? Yes. Annoying? Yes. Maddening? Yes. Fun? At times, but mostly not.
All of that came to a head on October 14th in Los Angeles. Things went from, “I know what they are going to say” to, “Crap crap crap crap crap”, to “WTF, why does my head hurt”. This was quickly followed by a flash of clarity, and the feeling of being had. It ended with, “You did a good job with keeping this thing under wraps. F[bleep] you very much. Well done, but f[bleep] you very much.” and this story.S|A
Latest posts by Charlie Demerjian (see all)
- 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
- Why Can’t Intel Supply Enough 14nm Xeons? - Sep 10, 2018