We’ve spent some intense and serious time with the AMD folks recently, in Las Vegas, Texas, Santa Clara, here in our labs, on the phone and by email, and we’ve gotten a pretty good understanding of the company’s plans and people—or so we’d like to think (or maybe they’d like us to think…). Putting aside Machiavellian schemes and ideas for novels, being able to understand a company and its people keeps you from making toady comments, misstating facts, and hopefully making bad forecasts. It also helps you better understand other companies in the market.
AMD is into sweets-SSSS
In June of 2008, ATI introduced their Sweet Spot Strategy (SSS) in graphics, basically changing the rules of engagement on how GPUs would be built and employed in AIBs and elsewhere. It seemed a brilliant concept at the time and we all waited to see how it would play out. Turns out it’s played out pretty well and, except for some supply constraint issues, would have propelled ATI to the forefront in market share.
In February of 2010, ATI introduced their Server Sweet Spot Strategy (SSSS.) Basically changing the rules of engagement in how Servers would be built and employed in systems and elsewhere. It seems like a brilliant plan and there’s some discussion about it in this issue.
In August 2008 AMD introduced their Sweet Spot Strategy for singularity (SSSS) in chipsets. Basically, the company began to show signs of being a true platform provider.
This week, the company introduced their latest, best, and last IGP, the 890GX, and there are, I think, some tales in it. IGPs, as you know, have to be small and inexpensive and ATI’s are. They also have to provide as much performance as the silicon and price budget will allow in order to justify their existence—and for the most part the suppliers have been doing that with “good enough” offerings, and sometimes better (i.e., Ion, and the 790GX).
The 890GX is remarkable for a variety of reasons, and one might ask why AMD would put out such a powerful chip, with 40 unified shader processors in a low cost, limited lifetime product?
First, one more data point. You will recall that ATI invented the unified shader before Microsoft implemented it with DirectX 10. ATI invented the unified shader for Microsoft for the Xbox360, and used it again in the Nintendo Wii. ATI/AMD now has several years of making very small unified shaders.
The next sweet thing
Later this year, AMD will introduce their first Fusion product, Llano—the monolithically integrated multicore CPU and graphics processor.
This will be another shot at a sweet spot—namely the value and midrange of desktops and laptops and will be AMD’s Sweet Spot Strategy for Synergy (SSSS.)
Llano will be built in 32 nm, at Global Foundries—AMD has shared that much—in fact they are being built now in preparation for engineering sample deliveries to interested OD and OEMs. And from what we’ve heard those OD and OEMs are very interested.
Specifications have been more difficult to come by and so we’re left to our speculations. I believe Llano will be based on a shrunken Phenom II with a much larger and coherent cache for communicating with the GPU portion of the device. And I think the 55nm built 890GTX shows us that Llano will have well north of 40 unified shaders in it, so I’m guessing 64 is the right number.
Also, I’m guessing Llano will have Direct Connect 2.0 and the ability to attach a sea of memory to it for those folks who want to do heterogeneous computing.
Let’s C if we can compute
As of now, AMD is relying on OpenCL as the programming environment for its heterogeneous offerings. And, as I’m sure you noticed, AMD recently announced they have ported OpenCL to their CPUs, as well as their GPUs. Of course, it makes common sense for AMD to do that. Intel will do it too, but the most interesting part is that it is a perfect prelude to an integrated heterogeneous product isn’t it? Sweet.
And now the bad news. Whereas AMD is doing just about everything right on the hardware side, they are way behind on the software side. Relying on OpenCL as their programming environment will put them in good stead with the programming and OEM community, but it’s like watching paint dry before you have anything useful. Say what you will about Nvidia’s CUDA being proprietary, and how much everyone hates such things—it is nonetheless the current industry standard and more than that it’s the industry’s darling if the enthusiasm at SC09 was any measure.
What if you threw a party and no one showed up? That could be AMD’s fortune if they don’t get their programming act together. They might show up this fall with the amazing and sweet Llano and no tools to use it with—that would not be very sweet.