Reports surfaced last weekend that AMD's graphics roadmap for 2013 would keep the Radeon HD 7000 series "stable" throughout the year. The news came out in an unusual way, via an interview with a Japanese website and then several tweets from AMD employees and the official Radeon Twitter feed. Naturally, we had questions about the state of things, so AMD held a conference call for the press today, in an attempt to clarify matters.
Prior to last weekend, we expected AMD to be introducing a new generation of graphics cards within the next few months. That expectation was informed by the firm's usual cadence for graphics releases and its public roadmaps. For instance, here's the slide AMD shared at its CES press conference, with graphics along the top.
The CES roadmap seemed to agree with the one AMD showed at its Analyst Day roughly one year ago, when it projected "Sea Islands" to supplant Southern Islands with a "new GPU architecture" and "major GPU architecture enhancements for graphics, compute, HSA" in 2013.
HSA, of course, is AMD's mixed-mode computing initiative that seeks to enable CPUs and GPUs to work together on shared computing workloads in more efficient fashion. Improved HSA support would almost surely involve modified silicon—either changes to the programming model, more bandwidth for data sharing, or both.
Thus, we expected the Sea Islands parts to be a top-to-bottom refresh of AMD's GPU lineup, likely to be called the Radeon HD 8000 series. We understood the "Solar System" chips to be the mobile counterparts to Sea Islands, likely based on the same silicon, with some possible re-branding of prior-gen parts in the mobile space, as often happens.
In fact, prior to CES, we published a performance preview of the Radeon HD 8790M, a mobile graphics processor based on brand-new GPU silicon. This GPU is derived from the same first-gen GCN architecture as current Radeons, but there's precedent for such mixing of architectural DNA. For instance, back when, the VLIW5-based "Barts" chip came to market alongside the VLIW4-based "Cayman." This new chip is code-named "Mars," placing it firmly in the Solar System naming scheme. Its desktop variant, "Oland," will be making its way into desktop Radeon HD 8000-series cards for large PC makers. And, yes, Oland is an island in the Baltic Sea.
The next step, we believed, was for AMD to announce desktop Radeon HD 8000-series cards based on new Sea Islands silicon and a second-generation GCN architecture. At CES, sources with knowledge of the matter told us to expect those products late this quarter. Instead, reports slipped out last weekend that cast doubt on that plan, and AMD scheduled its conference call with the press.
I'm not sure what to tell you next about the call that ensued. Although the goal was roadmap clarification, no roadmap slides were shown or discussed. AMD's Darren McPhee began the call by asserting that the Sea Islands code name refers to products for "desktop and notebook, with focus especially on the notebook" and that it's primarily aimed at OEM customers, or big PC makers.
He then affirmed that the Radeon HD 7000 series will remain AMD's "focus" on the desktop for most of 2013.
He explained that AMD is "not sitting still" with the 7000 series, that new products are coming in it. He then explained that AMD plans to introduce "a new series of products" by the end of 2013.
At this point, our sense was that the larger chips in Sea Islands, the ones with the new graphics architecture, had been delayed. AMD's reps were essentially saying that the plan for Sea Islands had never included such things, that Sea Islands was focused on mobile and OEM products. Along with others on the call, we began asking clarifying questions. I think perhaps the best thing to do here is to share some of the ensuing exchanges with you, so you can see exactly what was said. We were talking with several AMD reps, including VP Roy Taylor, as you'll see.
Let's start with this bit about what Sea Islands really is:
TR: Guys, it's good that you guys are selling the product that you have, but this marks a change from your approach of having a certain cadence for products. And we know that these chips take three or four years to develop, and it appears that Mars/Oland was actually a delivered chip. There have to be other chips in the pipeline. And now we're hearing they're not going to come out on schedule. This is a roadmap change—
Dave Baumann: Mars and Oland has already made it to the market. They're already shipping. In fact, you can buy notebooks actually in November—
TR: I know that, but I thought there were other chips in that family. Were there not?
Dave Baumann: Yes, there are. And they'll be announced in the coming months. We've already made roadmap presentations that show additional notebook products coming in Q2.
Darren McPhee: In Q2 we have more 8000M parts to round out the family, as an example.
Dave Baumann: And as we said, we're not done launching the 7000 series yet, either. And that's not necessarily just refreshes of current ASICs.
Next came this bit, a little later:
TR: Guys, I think there's some confusion here, because you're talking about new products, and products can be many things. They can be based on existing ASICs, they can be new ASICs. The expectation I think that we came into this year with is that you had a top-to-bottom ASIC refresh planned, and that that was called Sea Islands, and that we would see products based on Sea Islands relatively early in this year. Are you now saying that that was never the plan? Or has the plan changed?
Devon Nekechuk: The plan has always been to keep our 7900 series as the mainstay of our enthusiast product line from when we launched at the end of 2011 to the last part of 2013. So this was definitely an intentional plan.
Dave Baumann: Yeah, there's been some hangover from the Internet hive mind that assumes that there's going to be an 8000-series refresh immediately or a year after the previous product. That's not the case. That's just why we're clarifying today.
And then this:
TR: Let me ask a question, because you said that there are some late 2013 products coming. Are those part of Sea Islands, or are they something else?
Dave Baumann: [long pause] We cannot really make comments on what exactly the product stack are and what the roadmap names are, etc. etc. You know, there's plenty exciting new products to look at when they come later in the year.
TR: Are they based on GCN architecture or are they something substantially new?
Darren McPhee: Graphics Core Next is the architecture that we're committed to as a company for years to come. So Graphics Core Next will be at the core of all of our products for the next several years.
TR: Let me make sure that I get this, because I'm still not clear and I don't want to write the wrong thing. You're telling me that the 7000 series is going to remain stable throughout the year, and that there will be new products in the 7000 series, and that there will be new chips in Sea Islands in addition to Mars and Oland that come out during that span of time? Is that correct? Including larger chips than Mars and Oland? But they'll be under the 7000 branding?
Darren McPhee: With the exception of not giving you details on that last question regarding size, yes, yes, and yes.
TR: So it's possible that the only thing we'll see is smaller chips than Mars and Oland?
Darren McPhee: I can't comment on the size. It doesn't . . . [trails off]
TR: Will Tahiti remain your high-end product throughout this span?
Darren McPhee: [pause] The 7900 series will be a focus for the rest of 2013.
Roy Taylor: Why don't we have another one of these calls in a couple of weeks and we can tell you some more.
There you have it. Phew.
Let me see if I can explain what I think is happening. I wouldn't normally do this, but some of the info coming out of AMD seems contradictory, so I'll offer you my version of events. Time will tell whether I'm right.
My sense is that "Sea Islands" was initially a truly new generation of GPUs based on a refresh of the GCN architecture, just as the roadmaps from last year's Analyst Day showed. Since this new family was to bring "major GPU architecture enhancements" for both graphics and compute, it's extremely unlikely Sea Islands was planned from the beginning with only smaller chips. At some point along the way, the Sea Island chips based on the architectural refresh were very likely either delayed or canceled.
The surviving parts in the Sea Island/Solar System families are smaller, mobile and OEM-focused chips based on the first-gen GCN architecture, among them Mars/Oland with 384 shader ALUs and the larger "Neptune" chip with 640 ALUs that AnandTech outed a while back. AMD has already launched Radeon HD 8000-branded mobile products based on Mars and will soon introduce OEM-only desktop Radeon HD 8000-series parts based on Oland, the desktop version of the same chip.
AMD has an entire desktop Radeon 8000-series lineup based mostly on older GPUs, only for OEMs, because PC makers expect a yearly update cadence.
The "other members" of the Sea Islands family still to come, to which AMD referred in the conference call, are probably Neptune and its desktop counterpart. We'll more than likely see OEM-focused Radeon HD 8000 series products based on this chip. We'd expect to see retail/channel products based on "desktop Neptune" and Oland that carry 7000-series branding, as well.
Those may well be the only Sea Islands chips that ever make it to market.
What's coming at the end of 2013 isn't entirely clear, other than that it will be a full lineup refresh, perhaps called the Radeon HD 9000 series. We don't know whether it will include the architectural enhancements once planned for Sea Islands, whether it will be based on a newer fabrication technology than the current 28-nm TSMC process, or much else. We do know the architecture will be branded as Graphics Core Next, so we're probably looking at an evolution of the current GCN setup, not something radically new.
Between now and then, AMD's graphics division may well have to compete with rival Nvidia without the benefits of new silicon—and we do expect a silicon refresh from Nvidia in the coming months. Even though both firms will likely be sticking with a 28-nm process, there are benefits to building a new chip.
For AMD, those benefits could have been substantial. GCN is an all-new architecture, and it's not as efficient in terms of performance per die area or per watt as Nvidia's Kepler. AMD almost certainly has learned a lot about GCN performance from having real chips in hand, things it couldn't have fully understood using simulation when the chips were being developed. Refreshed silicon could have made substantial improvements. Beyond that, refresh generations offer the opportunity to build "tweener" chips like Barts whose mix of size and internal resources hits a sweet spot in the market. Also completely off the table is anything to rival the rumored GK110-based monster that Nvidia may use to capture the single-GPU performance crown.
AMD could very well remain competitive throughout much of the market with a combination of aggressive game bundling, discounts, product repositioning, and the promised driver updates in the pipeline. Without new silicon, though, it could be facing a rather difficult uphill battle.
Unless, you know, I'm completely wrong about what's happening here. Of course, we'll keep watching this story and update you on further developments as they happen.