AMD’s Financial Analyst Day packed a great deal of information into a few hours. It was, at times, frustratingly vague as far as the company’s long-term strategic plans, but one fact came through clearly. The performance war between Intel and AMD is over. Intel won.
Whether or not this is a bad thing is very much open to discussion. Let’s have a look at the big picture.
CEO Rory Read has said for months that the company’s focus would be in “low power, emerging markets, and the cloud.” New details on that plan, and Read’s comments yesterday, indicate that AMD will no longer focus on delivering high-performance products built on the latest process technology. The following slides lay out the current plan and future focus in stark terms.
Missing from the second slide is any mention of enterprise products (apart from cloud servers), discrete graphics, or conventional desktops. That’s not to say that AMD plans to abandon these markets, but none of AMD’s executive presentations or breakout discussions discussed high-performance CPU or GPU products. Instead, the company’s focus is on moving into SoC (System on a Chip) designs and leveraging that IP to create new products both for the 2012-2013 timeframes we’ve seen and beyond.
This move actually makes a lot of sense. The trend towards SoCs is driven by continued growth in transistor densities. AMD wants to be very active in the low power market, and bringing components on-die reduces both power consumption and communication latencies. This dovetails with the company’s decision to abandon its competition with Intel at the upper end of the market in several ways. AMD has no prior experience in building SoCs (unless you count the failed Krishna/Wichita designs), and such parts are inherently more complex than CPUs or even APUs. Combining new architectures and new processes in 2011 got AMD a chip it couldn’t manufacture, a chip nobody wants, and an APU it had to cancel. The only APU to hit its targets was Bra os — and it was built on 40nm.
Foundry economics are likely another factor. The press pays a great deal of attention to TSMC and GlobalFoundries’s production at their leading nodes, but that’s only a small component of their business. TSMC’s figures for Q4 2011 indicate that nearly 75% of the company’s revenue is built on process technology that debuted five years ago. With the cost and difficulty of each node transition rising, the foundries have little incentive to push the envelope. It makes much more sense for TSMC and GlobalFoundries to ramp nodes slowly — and that doesn’t fit well into a narrative of going head to head with Intel. AMD, in turn, isn’t willing to pay the hefty premium associated with driving early risk production at either foundry.
Ambidextrous AMD
AMD made multiple comments yesterday about being “flexible around ISA” and portraying itself as a “solutions” provider rather than being wedded to a single type of solution. At the same time, the company’s roadmaps through 2013 demonstrate a commitment to x86-based products.
Even if AMD is off designing an ARM core, it’ll take several years to bring a solution to market. Sunnyvale’s remarks around ISA (instruction set architecture) are best understood as part of the company’s strategic shift towards SoC’s and the mobile market.
What AMD is saying here is that if there is third-party IP it can license to improve its products, it’ll license it. If it sees an opportunity to develop its own IP for cross-licensing, it’ll do so. Part of the “flexibility” the company is discussing refers to HSA (Heterogeneous System Architecture) and the use of the GPU to handle specific tasks. The company is angling for shorter design cycles and a faster time to market, with new products shipping in 18-24 months as opposed to the 3+ years between CPU architectures. Again, the goal is to re-use synthesi able IP blocks rather than to create new products from the ground up.
HSA: The last bright spot
Nearly all of AMD’s presentations yesterday revolved around the need to reposition the company, build SoCs, and transition away from challenging Intel at the high end of the market. The one exception to that trend was AMD’s discussion of heterogeneous computing. Here, the company’s unification plans for CPU and GPU could still produce exciting products. AMD’s goal for HSA is to create solutions that are easier to program, optimi e, and load balance while offering higher performance in a lower power envelope.
Here is AMD’s HSA roadmap. In 2012, we gain user mode scheduling and bi-directional power management. User mode scheduling is a Windows feature that allows applications to schedule their own threads without involving the system scheduler. Microsoft recommends it “for applications with high performance requirements that need to efficiently run many threads concurrently on multiprocessor or multicore systems.” AMD’s slide mentions C++ support for GPU compute, but that’s a feature of the HD 7900 family — APUs will have to wait for 2013.
The features scheduled for 2013 make it easier for the CPU and GPU to share data. A fully coherent memory model will allow the CPU to spin threads off to the GPU (or vice versa); the unified address space and use of pageable system memory simplifies communication between the two. By 2014, the boundaries between a “CPU workload” and a “GPU workload” have essentially become invisible. The idea is that a developer can target the HSA model and leave the decision of where code should be running up to the hardware.
HSA isn’t a replacement to OpenCL — AMD refers to it as an “optimi ed platform architecture” that uses OpenCL. The HSA ISA is virtual, meaning that it’s not tied to x86, ARM, or any particular language. According to Joe Macri, CTO of AMD’s client division, code written in HSA is compiled for the underlying CPU or GPU using a JIT (just in time) compiler.
This isn’t the first time we’ve seen this information from AMD, but the company never put dates on anything before. If AMD can build support for HSA, it could give its consumer products a substantial visibility kick.
End of an era
The presentations and speeches yesterday made it clear that the old AMD — specifically, the AMD that prioriti ed competing head-to-head with Intel and building products that could challenge the best Chip illa had to offer — is dead. As an enthusiast who bought his first K6-233 in 1997, that’s hard to write. I cut my overclocking teeth on a Duron 600 with a pencil modded voltage of 1.85v that enabled the chip to boot at 160MH on a KT-133A-based IWILL KK266; I still own the Tonicom PC166 RAM I bought for the system.
A moment of silence, if you please.
As a journalist who’s covered the company for a decade, however, it’s hard to ignore the fact that AMD isn’t giving up so much as it’s acknowledging that it lost this fight years ago. The company isn’t folding — it’s retrenching around the idea of building mainstream products for the mass market using proven process nodes. As strategies go, it’s still a risky one. Sunnyvale’s decision to focus on SoC integration makes sense given where the market is headed, but its 28nm Bra os will have to compete against 22nm Atom and 28nm ARM chips. If Piledriver performs well and AMD can ship its 2013 products towards the front of the year, it should do well. If its 2013 roadmap piles up in the back half, it risks being left behind altogether.
To be honest, Read’s decision to wrap up his presentation with this slide doesn’t engender confidence. What comes after “Shaping the future” — “Terrori e Tokyo with my radioactive breath weapon?”
AMD could still have a role to play in driving the evolution of mobile products and helping keep prices low by providing competitive performance in mainstream and value market segments. If the company’s HSA initiative takes off, it might even give Sunnyvale a long term launch point back into the higher-end desktop, mobile, and server markets. For now, however, the company is retreating from these markets — and leaving Intel Inside.
Goliath wins: AMD retreats, retrenches, and seeks to reinvent itself
0 comments:
Post a Comment