AMD TOOK THE worst possible route in naming their new 68×0 cards, and the explanations for it only make things worse. No, scratch that, it isn’t the worst possible route, but it does destroy a long standing and well thought out convention.
This isn’t by any means a G92 like debacle, if you recall that chip went from the 8800 to 9800 to 250GT to 300-something without anything more than an optical shrink. AMD just does a little name inflation that they seem to think is innocuous, but it has some nasty repercussions.
The short story is that the new ‘Barts’ chip fits nicely in a market segment hole between the older 5770 and the 5800 series. That $200 segment is a hotly contested part of the market where sales are decent and margins are too. It is a rarity, and caps the mid-range of the market where the GPU makers make the majority of their profit.
So, faced with calling the new chip a 6700 or a 6800, AMD decided to make it a 6800. You could make a case for it going either way, but the 6800 spec inflation has many more cons.
The most glaring problem is historical. The 5870 is about twice as fast as the 5770 chip, the 4870 was twice as fast as the 4770, and the 5870 is about twice as fast as the 4870. The 5870 has 1600 shaders, the 5770 has half of that, 800. The 4870 has 800 shaders, half of the 5870, and the 4770 has half of that at 400. See a pattern?
Update: The 4770 has 640 shaders, it was a 40nm test part.
In comes Barts/68×0 with 1120 shaders, not quite 150% of 5770. Added tweaks and higher clocks make it perform very close to the 5850 which has 1440 shaders. That is what AMD justifies the ‘name upgrade’ on, it is in the same performance class as the 58xx chips. AMD is entirely correct here, it is in the same performance class as the 5800 line.
What they didn’t take into account, or more likely did and ignored it for marketing reasons, is that the 5 in 5870 has been supplanted with a 6 as in 6870. That generation advance usually comes with a 2x leap in shader count, performance, and everything else. The 6870 is not a 2x advance over the 5870, it is actually slower.
So, instead of a 3200 shader (or the equivalent) card for the top of the 6800 line, we get a 1120 shader part. The 6700 line should be 1600 shaders (or the equivalent), and that is what we got. By all rights, Barts should be a 67xx card. That is dumb move #1.
To make matters worse, as we said before, the 6870 is SLOWER than the 5870. When was the last time you can recall a GPU coming out that was slower than it’s predecessor? By a measurable margin? On the same DX level?
I think the count there is zero. All the memorable flops of yesteryear like the 5800 and 2900 involved a transition to a new ISA that, errr, didn’t work out out as, umm, intended. On the same DX level, I can’t recall anything that was slower than it’s predecessor.
Mr. Moore has a succinct explanation as to why, but sadly that does not apply to marketing departments. AMD botched this one, plain and simple, dumb move #2.
The third reason is a little less problematic, and involves the decimal numbering system. AMD has traditionally named their high end single chip part the x800 line, as in 3870, 4870, and 5870. Mid-range cards with half the shader count lose one off the second digit, and lower end parts see that second digit get progressively lower.
With the last generation 5000 series, the highest end dual GPU card went from having an X2 suffix to being their own line the 5900. There was only one part in that line, the 5970, but it represented the pinnacle of the market, and did not carry an X2 suffix. x970 is the new x870X2, fair enough.
Name inflation means that the obvious place for high end single GPU part is taken by the low end. Since the decimal number system can only count to 9 with one digit, the upcoming high end Cayman GPU has to be the 6900 series.
That is fine, but where does that leave Antilles, the dual GPU card? Will they call it the 6A70? The 7090? According to decodes of the drivers, that card is likely to be named the 6990, and the two Cayman chips are going to be the 6970 and 6950.
The difference between the Cayman and Antilles boards, a hugely different market segment, has barely any name differentiation. The mystery and mystique of the uber-dual GPU card is now basically lost, and there is no headroom for future expansion. While all this technically fits into ATI’s naming scheme, that scheme is bent to the breaking point and a little beyond. Dumb move #3.
So, what are the explanations as to why the 6800 moniker is warranted? There are three that I have heard. First is that the new chips are in the performance class of the 5800 line, maybe a bit slower, but definitely in the right ballpark. That is totally correct, the performance is in the same class as the last generation high end, but that is exactly what you would expect from the mid-range of any new generation. It doesn’t wash.
Next is a more marketing oriented explanation. The new ‘Northern Islands’ parts are a family of GPUs. Barts is the mid-range GPU, and that will soon be followed by higher end, then lower end parts until the family is fleshed out. That is the way GPUs are launched, a bit at a time.
Because Barts is a little higher up the price structure than the replacement for the 5770, AMD claims the 6800 moniker is warranted. Barts based cards sell for $179-$239 while the older 5770 cards started at $179 and went down from there.
The higher price is based on two things, Barts is a much more expensive chip to make than the Juniper that powered the 5770 and because the competition is, or was, priced there. That level of performance ‘dictates’ that price. If there wasn’t a generational advance, this might be the case, but Mr. Moore still continues to beat his drum, even if TSMC is a little tone deaf. Price doesn’t justify the change either, Barts should be cheaper.
Lastly we get the most compelling reason for the change, and it is summed up by the phrase, “Because we can”. Yes, AMD can, and, in fact, AMD did. It is their card, and they can name them whatever the heck they want.
If AMD was feeling cheeky, they could have named it the “Twinky Winky 90210 GT-R WTFLOLBBQ”, but they didn’t. Even an idiot can see this would likely have a mildly adverse affect on sales because the name is stupid, and name inflation tends to have a positive effect on sales, especially when you don’t have a compelling product.
Just ask Nvidia about that one when the GTX580 paper launches. The biggest problem with name inflation is AMD did have a compelling product. Barts should be a 5790 or a 5860, not a 68×0.
In the end, AMD did three really stupid things and put out a few excuses that seem, at best, phoned in. They made a generational advance without any advance. Then they put out a card that was slower than it’s spiritual predecessor, a first in the industry. Last, they totally borked a quite clear and functional naming scheme. Worst of all, they don’t have a single good excuse for doing what they did.
While none of this is rises up to the level of Nvidia’s deceptive practices or Intel’s sadistically broken random name generation system, it still sucks. The worst part is that AMD didn’t have to do it, this hat trick of own goals was entirely optional, and easily avoidable. Fscking brilliant.S|A
Latest posts by Charlie Demerjian (see all)
- SemiAccurate has Skylake-SP die shots, sizes, and more - Jul 21, 2017
- Skylake-SP has a diverged core - Jul 19, 2017
- Intel’s Purley platform architecture is a step forward - Jul 13, 2017
- Intel launches Purley aka metal Xeons - Jul 11, 2017
- AMD’s Epyc has lots of connections - Jun 26, 2017