As I mentioned in my previous blog, AMD and GlobalFoundries,it is rather difficult to interpret AMD’s financial results. At the end of the quarterly earnings conference calls analysts get to ask questions. It is the comedy relief portion of the call usually, so I blog this today to better prepare these so called analysts for their job of helping investors understand what is really going on insideAMD.
There are two main products in the server space: Lisbon/Magny-Cours (45nm SOI, introduced in 2010) ; and Valencia/Interlagos (“Orochi” Bulldozer core-based modules, 32nm SOI, introduced in Fall 2011).
Server revenues were up, a good sign. (The AMD x86 server market share is still very low, compared to levels achieved with the initial Opteron family, of course.)
However, the financials are puzzling:
“Server revenues up QoQ in double digits, Bulldozer-based units making up 1/3, but the ASP’s declined.”
The latest design should command a price premium. For ASP’s to decline, either the 45nm generation was discounted deeply, or the Bulldozer parts do not command the premium, or a combination of both. Given the report that AMD inventories at the end of the quarter declined, I’m assuming the 45nm parts were discounted to move them out.
Yet, the server performance benchmarks for the Valencia/Interlagos have not been stellar (although not as bad as the desktop benchmarks – more on that shortly).
And, AMD has pre-announced the second generation Bulldozer core – aka, “Piledriver”, also 32nm SOI – will be available in mid-2012, with micro-architectural improvements for performance. And, there are highly-anticipated Ivy Bridge product announcements from Intel in April.
I assume the server OEM’s are negotiating significant discounts from AMD, on both the older and new, Bulldozer-based parts. So, while the server revenues are up, it is concerning that the ASP’s are down, especially in light of a new product line refresh in September, 2011.
Based upon the pricing strategy that Intel adopts for Ivy Bridge (and existing Sandy Bridge) products in April, this may not bode well for AMD’s ASP leverage.
NET: I have a negative outlook on AMD’s ability to grow server market profitability.
There are several desktop family parts still available – but, AMD has committed to transition from 45nm (and older) designs to 32nm Bulldozer/Piledriver parts going forward, to shift GF resources to 32nm ramp.
So, the current desktop offering primarily consists of: Athlon-X2 (65nm), Phenom-II family (45nm, non-APU, introduced in 2009-10), Llano (32nm SOI APU, in 2010), and Bulldozer/Orochi (32nm SOI non-APU, in Fall 2011). Going forward, clearly the 65nm and 45nm product availability will diminish.
The financials reports indicate “desktop processor revenues continued to slide…”
AMD has held a significant market share (25%+) in this segment, due to the relatively strong performance of the Phenom-II family.
However, the recent introduction of Bulldozer/Orochi in this segment was an epic fail. The performance on single- and lightly-threaded applications typical with desktop users was on par or even below the existing Phenom II parts – I cannot recall a case where a new product in a market segment performed more poorly than the part it is intended to replace. Although designed as a server part for very highly-threaded workloads, the introduction of Orochi as an OEM desktop part is baffling.
To further add confusion to the execution of the Bulldozer/Orochi rollout, there was a response from AMD that:
“the Windows7 O/S is not optimized for the unique, asymmetric thread execution on Bulldozer… Windows8 will be better suited for this architecture, and in the interim, Microsoft will be issuing a Windows7 O/S patch.”
To me, this excuse is inexcusable… Bulldozer has been in active development since 2006. There has been plenty of time to provide Microsoft with performance models and engineering sample hardware to tune the current O/S. To release a new desktop product line, then indicate that the preferred O/S will not be available for 12+ months is poor execution. (I haven’t seen much about what the Linux community has done, in the way of attempting to optimize the O/S task dispatching algorithm for Bulldozer/Orochi.) And, although available in a desktop form factor from a few manufacturers, Llano parts are really not well-suited for the high-performance desktop market. Although the APU (CPU+GPU) architecture does support the addition of an external GPU, the die area dedicated to the GPU necessitated reductions to the cache hierarchy typically found on higher-performance desktop processors.
NET: The desktop market segment where AMD had a significant market share has declined. The transition away from Athlon parts for low-end desktops and from Phenom for high-end desktops to a Bulldozer-based offering has been poorly executed, to date. It’s difficult to foresee what will result in a turnaround, certainly not a “10% performance improvement with Piledriver in mid-2012”, given the upcoming Intel announcements.
This is the best result for AMD in the past quarter, although it was pretty much “by accident”. (Sometimes it’s OK to be lucky.)
At the low-end, the “Bobcat core” based Brazos APU products have done surprisingly well. Initially designed for a “netbook” or “ultra-mobile PC” (UMPC), neither of those product form factors appear to have much longevity.
So, OEM’s started putting the dual-core “Zacate” member of the Brazos family into very low-cost notebooks, which clearly has done well. (I don’t anticipate the single Bobcat-core “Ontario” design has gotten much traction, though… mostly dual-core, low-cost, low-performance notebooks – 15.6” notebook from Toshiba with Zacate, 3GB DDR3, 250GB HDD for $329 at Fry’s. There’s definitely a market for very low-cost notebooks.)
Although used in a product line for which it was not intended, Brazos is doing well. When that fortunate situation is presented to a company – a high-volume part in a niche market segment – they typically look to extend that success through either product enhancements, or more appropriate in this case, cost-reduction efforts.
However, in this case, AMD has once again not executed well. The original plan to migrate the 40nm bulk Brazos designs at TSMC to 28nm bulk at GF was recently scrapped. Although there is no official word from AMD, every indication is that “Wichita” and “Krishna” projects were recently canceled. It is unclear how AMD intends to extend the life (and improve profitability) of these designs.
For mid-range laptops, the Llano APU has also done well, vindicating (to some extent) the decision to integrate mid-range GPU and CPU technology together, into a cost-effective offering. Ironically, Llano is also a bit of a “mistake” – it was not originally on the AMD product roadmap, as the Bulldozer core was going to be the first in a “Fusion APU” design. Yet, the delays in developing the Bulldozer core necessitated adding a new, stopgap design on the roadmap, which has done admirably well.
The Piledriver-based APU, aka “Trinity” (32nm SOI), will be announced mid-2012, to displace Llano. Frankly, AMD has a lot riding on this introduction, as there are no plans to extend the (stopgap) micro-architecture core used in Llano.
Intel’s product announcements coming in April will set a new bar in laptop/mobile processor performance. It may be difficult for Trinity to differentiate itself from these offerings, as the “on-die GPU performance gap” between AMD/ATI and Intel will likely be diminished with the new Intel parts.
NET: No clear path for improved Brazos profitability, through any cost reduction. A new product in Trinity to displace Llano in mid-range desktops (APU-based, no separate GPU) mid-year will need to hold the market share that Llano has achieved – this will be a difficult task.
“Graphics revenue fell 5% QoQ, due to a decline in shipments blamed on HDD issues…”
According to the financial reports, rising HDD prices resulted in fewer laptop manufacturers integrating discrete GPU’s, affecting both AMD/ATI and Nvidia.
Actually, in AMD’s case, I have to believe that there is ongoing “cannibalization” of the mid-range GPU offering by the APU’s. Nvidia has realized that the high-end discrete GPU parts (desktop and laptop) are a tough market, and has made a concerted effort to expand the reach of its offerings – e.g., high-performance computing opportunities, non-GPU (Tesla) processors.
Despite some initial ventures into the HPC area, the strategy for AMD/ATI to expand the discrete GPU revenue is unclear. So, I expect an ongoing leapfrogging of Nvidia and AMD market leadership with their upcoming (28nm) offerings, in an increasingly niche market area.
NET: Although certainly competitive with Nvidia as the next discrete GPU product generation will be available in 2012, the long-term growth aspects of this segment for AMD are unclear. Ongoing “cannibalization” creeping up from low- and mid-range integrated graphics units will also put a pinch on revenue.
To me, the biggest issue for AMD is “execution”, which their new CEO has mentioned repeatedly. The introduction of Orochi/Bulldozer for the desktop segment was a mess. The lack of a cost-reduction strategy for low-end x86 parts is a mess. The (pre-mature?) shift in production emphasis from 45nm to 32nm parts at GF is concerning. The relationship with GFI is concerning.
However, IMHO, the biggest AMD execution issue is the “blame” that was put on GF for “yield issues, constraining product volumes”.
What most so called analysts don’t know is the importance of Designing for Manufacturability (DFM) and Designing for Yield (DFY) has continued to grow, since these concepts were first emphasized in 90nm and 65nm nodes. It is no longer adequate to meet the “design rules”, send the DRC-clean layout database to fab, and expect production level yields. Design and foundry must collaborate closely (especially for an early node part) to reduce the yield sensitivity of circuit layouts, and to understand the requisite design margin needed to accommodate process variability.
Assuming the “defect-limited yield” (manufacturing defect density) at GF was in control for 32nm, the “design-limited yield” requires a concerted effort to optimize where to push the available lithography, and where to relax the design to reduce yield sensitivity.
In all the comments AMD made in 2011 about “GFI yield issues”, AMD never really addressed what steps they were taking to improve the design-limited yield. The current financial comments indicate “GF yields have improved”, and supply will be less of a concern (for existing parts) in 2012.
Yet, the lack of a plan describing what steps AMD has taken and will be taking for DFY is the biggest “execution issue” of all and the importance of this collaboration with GF is just going to increase.
Again, this blog is based on the most recent AMD Q4 2011 call. I post this as an observation and discussion rather than a report of facts and figures. I respect AMD technology and hope they succeed but I do not agree with how AMD executives communicate with investors on these conference calls. But then again, I’m just a blogger so help me out here.…..