WHAT IS THE latest on DAMMIT’s, soon to be just MAD again, Radeon 6000 series? The story is long and complex, with some things coming into view with a bit more clarity.
Lets start out by saying we got something very wrong at SemiAccurate, the Southern Islands name. No, SI is real, it is the 28nm version of the chip that we have been calling Northern Islands for a few months now. The dates, functions and all the rest we had correct, we just reversed the family names.
Why? Lets blame TSMC for that. From what we are told, TSMC canceled their 32nm node less than a month before AMD taped out Northern Islands, then a 32nm family. Some people at DAAMIT were MAD as heck, in a polite Canadian manner of course, but what can you do? Word has it that Nvidia was similarly affected, but they were nowhere near as close to tapeout.
That is where the confusion started. The 32nm NI parts were called Cozumel, Kauai and Ibiza,and those were unceremoniously flushed. The next stopgap parts came about at the same time we heard the name Southern Islands, and that was the week our normal crystal ball was at the shop. The backup, with a government supplied HD decoder box confused the new code name with the stopgap family.
The stopgap family had no code name at first, and then was called NI-40, then just NI once again. This happened about the same time 28nm info started floating in, and as bad luck would have it, the decoder started acting up. We got the family names wrong, but everything else right.
Internally, new code names, Cayman, Barts et al were put into place, and then eventually those re-became the NI line. Still with me? If not, it is OK, I was lost too. In any case, that is where we screwed up, and why.
Back to the NI family, what are they? Well, that part is easy enough, they are a serious re-do of the Evergreen family. The biggest change is in the shaders, they have gone from a 4 simple + 1 complex arrangement to a 4 medium complexity arrangement. This should end up no slower than the old way for simple calculations, the overwhelming majority of the workload, but also be faster for most of the complex operations.
The reason for this can be summed up by saying that the new ‘medium’ shaders can’t do what a complex one can in the same time, but there are more of them, and they can more than make it up in number. Since a GPU is a throughput machine, not a latency bound device, you won’t see the difference, it will just work a lot faster in several kinds of operations.
There will probably be a pathological case or two that will be a bit slower, so look for the attack slide decks to float as soon as samples leak. Remember the Nvidia slides from CES about how Fermi was many times faster than HD5870 on a specific section of the Heaven benchmark? Remember how well that turned out in practice, and in sales? Wait for real benchmarks, and don’t worry about the desperate sputters from the big green FUD cannon.
Since the shader count is 80% of the old grouping, there is some space saved, and on top of that AMD has had a lot of time to optimize area. On the down side, each shader is marginally bigger, but the end result is a cluster of four new shaders that is smaller than the old 4+1 group, and faster too.
The uncore, or at least unshader is beefed up as well, with some very notable efficiency gains as well. The net result should be vastly better utilized shaders as well, so performance should go up incrementally there too. Throw in a few more shader groups, and you have a notable speed increase from the NI family.
The down side is size. More shaders, smaller though they may be, and a beefier front ends adds up to a larger die. If this part was on 32nm, it would be smaller and probably would have had more shaders as well, but the backport had a price. The net result is that the die of NI will grow by about 10-15%, lets call it around 380-400mm^2. Performance on the other had should grow disproportionately, with the few weak spots of the Evergreen architecture smoothed over. That is what the game is about, isn’t it?
When is the HD6000 series launching? There have been rumors of October 12th flying around, but that is not the launch date. On October 12th, there is an event related to the 6000 series, but there won’t be cards on sale until October 25 +/- a day or two depending on schedules.
What is going to launch then? The one area where AMD has a hole in their lineup is the $175-250 range, and that would be filled by the hypothetical 6700 line. In a stunning bit of coincidence, that is exactly what will launch first, shortly followed by the 6800 and 6900 line. In early 2011, the lower end parts will backfill the lineup, and that will be it. Until 28nm. Then things get really interesting.
Back to the near future, the launch of the 6000 line will basically bury Nvidia’s chances in the market. The green team already has a higher cost structure than the current AMD 5000 line, and has to resort to price drops and kickbacks to achieve their current tepid sales. AMD in the mean time is holding prices steady on the top end, but the 57xx line is starting to show a little price softness.
It is almost as if AMD is making room in their lineup for a new chip to slot in and step on Nvidia’s hopes and dreams. In fact, that is exactly what AMD is doing, crushing Nvidia’s margins while still making money, and clearing out their inventory at the same time.
AMD is currently in the driver’s seat with the HD5000 line, and will roll out the replacement HD6000 line within two weeks of Nvidia completing their 400 series. AMD can drop prices on the older parts to fill in holes in the 6000 line until it is completed. During that time, Nvidia has no response.
AMD handily trounced Nvidia with their HD5000 line to win the first match of the DX11 tournament. With the HD6000 series, they will win match two by even greater margins, both financial and performance. Until the TSMC 28nm process is ready, Nvidia has no chance, but from the information already trickling out, don’t expect things to get any less abusive at that node.S|A
Latest posts by Charlie Demerjian (see all)
- Globalfoundries 7nm process isn’t even close to the name - Sep 26, 2016
- ARM upgrades realtime offerings to v8-R and adds Cortex-R52 - Sep 21, 2016
- Everspin and Globalfoundries team up for embedded ST-MRAM - Sep 15, 2016
- Intel’s Xpoint is pretty much broken - Sep 12, 2016
- ARM adds 2048-bit vectors to v8A with SVE - Sep 7, 2016