AMD recently had their first earnings call for 2015, where they talk to their shareholders and industry analysts about what’s happening inside the company, as well as future plans for their growth strategies and how they’re going to climb out of the well of debt they have dug for themselves. This earnings call was a little different, however. Not only was AMD clear about their future strategy and how to return to profitability, they also detailed their new products that they’re going to be releasing in the next 12 months, starting with the next-generation Radeon GPUs and their Zen processors, set to replace all their current silicon on the CPU market. Its a tough time for AMD and these will be their most difficult months to slog through, but things are looking up for the red team.
Putting all their eggs into one Zen basket
Its been no secret that AMD wants to get as far away from the Bulldozer-derived architectures that they’ve been pushing as possible and at the end of 2014, two code names were being bandied about by AMD’s CEO Lisa Su and CPU Head honcho Jim Keller – “Zen” and “Skybridge”. Skybridge is an initiative to make the ARM server world more accessible to AMD and its customers, allowing them to build a modular APU design that would be able to either integrate a set of ARM cores or a set of Carrizo-derived cores (with a later update to replace Carrizo with Zen).
That was a possible product for the desktop market but the plans have changed slightly. Skybridge will still arrive for the server and embedded markets utilising separate sockets for x86 and ARM products, but it will be Zen and x86 technology all the way for consumers and OEM notebook and PC vendors.
The details on Zen are pretty sketchy so far and no-one really knows too much about it. However, AMD has now promised that Zen, compared to the older Excavator cores that will be found in Carrizo processors, will make room for up to 40% more instructions per clock (IPC) to be fed through the processor, improving throughput and performance. This is a fantastic figure given AMD’s commitement to only 15% improvements year-on year over Bulldozer, and much higher than Intel’s 5-10% IPC improvements since the release of Sandy Bridge.
However, this might be one of the ways in which AMD improves their efficiency rating while still holding on to “good enough” performance, which I’ll detail in a little bit.
Zen also is expected to help fix AMD’s position in the CPU market, offering new products that fit into the budget APU market as well as the high-end desktop, previously catered for by AMD’s socket AM3+-based Vishera FX processors. DDR4 compatibility is now totally on the cards and AMD’s moving to a single socket, finally. This is going to be interesting because they would also be abandoning the old 900-series chipset found on the AM3+ platform.
My guess is they’re going to use the existing 800-series chipsets found on the socket FM2+ motherboards compatible with existing APUs, but the socket change will mean that you won’t have an upgrade path beyond socket FM2+. At least we’re getting PCI-Express 3.0 and possibly some SATA Express and M.2 love out of this switch.
Looking at upgrade paths again, the different colours for the product segments show that AMD is making a complete, clean break from their existing IP and moving on to something completely new. What’s interesting about this table is the scales – AMD expects that some of their Zen-based APUs will eat into some of the market share held by lower-priced Vishera CPUs, while the Zen-based FX lineup will aim higher than the current lineup. I think we’re going to see an eight-core, hyper-threaded Zen AM4 part that competes directly with Intel’s Core i7-5820K at the very least.
Again, notice that the sockets are going to be the same for the mid-range and high-end desktop markets. That is a massive departure for AMD and we’re going to have to get used to the possibility that AMD will now switch to using Intel’s strategy of a new socket every two years. Maybe backwards-compatibility between two sockets will remain, though.
As for energy efficiency, how are they going to achieve that? Well, AMD has a target of a 25x improvement on energy efficiency by 2020. That’s not impossible, but its highly improbable that they’re going to be able to achieve this on every part of their CPU and GPU architectures. Like Intel, they may have to design modular components inside the chip that can be powered down when not in use, saving on power and heat generation.
Another possible strategy they could use is to take advantage of the IPC improvements. Previously, energy efficiency in Kaveri came at the cost of using a 28nm production process that was suited for GPUs using high-density libraries. AMD compensated for that by increasing the IPC and dropping clock speeds to compensate for the low switching speed and this brought them significant power consumption improvements. However, overclocking Kaveri APUs is problematic because the transistors can’t switch as fast as their counterparts in Richland-based APUs, manufactured on a CPU-optimised 32nm process.
If AMD aims to improve energy efficiency by a factor of 25x by 2020, they’re going to be increasing IPC and dropping clock speeds to meet their target. If they do this and don’t move back to a CPU-optimised production process, they’re just going to have the same problems with existing Kaveri chips and high clock speeds, which will require extra voltage to overcome the slow transistor switching times.
Graphics gets a boost thanks to HBM and API improvements
AMD’s plan is to soldier on with the Graphics Core Next (GCN) graphics architecture, but this is going to be the fourth year that it features as their main GPU IP. Sooner or later, they’re going to have to switch out to something different, but my best guess is that will come in 2018 or thereabouts. For now, AMD will be reiterating on GCN with GCN 1.3-based products, which are similar to the Tonga architecture found in the Radeon R9 285. Tonga already meets their promise for “Advanced Memory Techniques” to be introduced, because it is the first GPU in AMD’s stable to make use of new colour compression techniques to reduce the amount of memory bandwidth needed to run a particular render for a game or application. Nvidia has this same basic idea in their Maxwell architecture and it is just as aggressive as AMD’s.
2016 promises a 2x improvement in energy efficiency and this could be down to new memory technologies such as High-Bandwidth Memory (HBM) being used in AMD’s GPUs. HBM is rather tricky and expensive to implement on lower-cost GPUs for now, though, so that improvement will also be helped along by switching to a new FinFET process suitable for GPUs. This could be at the 14-nanometer scale, so we can expect improvements in power consumption and the amount of voltage needed to drive the chips at their rated speeds.
AMD clearly thinks that HBM will be their saving grace, and that’s why they’re pushing for it so hard in their upcoming products. Basically, instead of connecting the memory to a bus that drives the chips separately using separate 64-bit-wide memory channels, HBM stacks the memory dies on top of one another and puts it closer to the GPU through a silicon interposer, reducing the travel distance needed on traditional GPUs and increasing the amount of memory bandwidth available, since the GPU can access the memory more directly.
AMD claims a 3x improvement in performance/watt and power savings of 50%, but this is because HBM memory runs at lower clock speeds, making use of a larger memory bus (1024-bit, single-channel rumored) to make up for the lack of throughput compared to higher-clocked GDDR5 RAM on a 512-bit memory bus. It will be a while before AMD and others like Nvidia can ditch the use of an interposer to make HBM even faster, but they’d have to then solve the heat issues that come with packing all the memory chips right next to a super-hot graphics core.
Its been rumored that AMD’s Radeon R9 390X will come in a 4GB and 8GB version, with the 8GB unit being shipped with a water-cooling kit, while the 4GB version makes do with air cooling. Adding on more memory dies increases the heat output and makes managing power levels more problematic, so we may see AMD also making significant changes to their GPU Boost technology to keep heat levels in check.
The question is, though, why is AMD pushing so hard for HBM adoption before Nvidia? Well, technically Intel has already figured out HBM on processors thanks to the L4 cache on their Core i7 chips with Iris Pro graphics, giving them sizeable advantages in memory bandwidth and performance improvements on the GPU side of things. Intel even envisioned Iris Pro graphics being used for GPGPU compute applications and with 128MB of super-fast eDRAM on-board with the CPU, that gives them a significant advantage over AMD and Nvidia off the mark. However, the L4 cache is on-die and not accessed through an interposer, so heat levels are much higher than normal and Iris Pro doesn’t end up being a very efficient chip overall.
AMD needs to figure out HBM on their GPUs before putting the tech into their second-gen Zen cores, because that’s the final piece of the puzzle to allow them to make their APUs even more suited to GPGPU applications. Using the Heterogeneous Systems Architecture programming model, they can go toe-to-toe with GPGPU hardware from Intel and they might even steal some market share away from Nvidia as well. Uses like Machine Learning and VR are all things we can already do today, but both applications benefit tremendously from increased memory bandwidth.
Lastly on the GPU front, AMD took a little time to remind people where they stand with API adoption. They already have three generations of graphics cards compatible with the Mantle API and Vulkan, so any improvements to those APIs benefit people running any version of a GCN graphics card. As far as DirectX 12 compatibility goes, all of their GCN graphics cards are again ready for DX12, but the only GPU that is feature-level compliant is the Radeon R9 285. That’s something that needs to be addressed soon in AMD’s GPU product stack and having even one GCN 1.0 or 1.1 product anywhere in the mid-to-high-end markets is a big issue they can’t afford to spend time downplaying.
I’m not sure why Metal is listed here, though. Metal is Apple’s Mantle-like API for their iOS devices and doesn’t even figure into OS X yet, so its possible that AMD is doing some work in the background to prepare themselves for the possibility that they’ll have to support it. If their future ARM products with GCN graphics ever are tasked to run an iOS device, it’ll be to their benefit and Apple’s that everything is in place already to take advantage of Metal.
Financial plans from now until end 2016
Those of you who follow my System Builders Guide articles will know how far AMD has fallen in their respective markets since they waved the white flag to Intel. Not only has the PC industry shrunk worldwide, revenues on products have also dropped, margins have become thinner and the only two players earning buckets of money from hardware sales are Intel and Nvidia, who use their market dominance to set the price of their products. But lets settle the one argument that goes around the net every year – AMD isn’t going to go away any time soon.
The company’s plan is to build up their cash reserves, currently sitting at around $800 million, to just over $1 billion. That will give them breathing room to lose about $150 million every quarter or so while they get the ship right in about a year’s time, a strategy that might just pay off because they’ll need to do a lot of cleaning up in the retail and distribution channels to make sure that all their old stock is gone or removed. This puts them into a similar position as a company like Nintendo – able to make huge losses on their products, but still stay in business because they can use money from the war chest to stay operational and work on fixing whatever they didn’t do right in the first place.
Fun fact – Nintendo as of 2015 has about $17 billion stashed away in their war chest, comprised of company assets and money. If they started today, they could take operational losses of $300 million a year for 57 years before they have to start selling off IP. If they released absolutely no consoles from now until 2020, they’d still have war chest money left over to fund development and production of another hit like the N64.
In the long-term, AMD has some interesting ideas of where they’re going to see their money flowing in. They’re expecting less and less growth in the PC market and only mid single-digit (4-7%) growth in the GPU market year on year. Still, they’re going to be grabbing about 4-7% profit margins on those markets, which tells us that they’re planning to make their next-generation of GPUs as cheap as possible. We should expect a Radeon R9 390X to fetch about $500 at launch because they need to compete against Nvidia as well.
Other markets stay in the green because growth and profitability is what AMD expects given their work with HSA and OpenCL acceleration, as well as investing and inventing new GPGPU technologies. The semi-custom market growth numbers are also interesting because that also includes growth from consoles like the Playstation 4, Xbox One and the next Nintendo console. That mid single-digit growth is probably overly conservative, but they’ll need to tread carefully in that area because they don’t want to be caught in the middle of a big jump in growth and not enough leeway to increase production to meet demand. They had that exact problem during the GPU mining craze and avoiding another two years of that mess is necessary to ensure future growth.
Videos
You can access all of the videos that contain these slides as well as the talks given by some of AMD’s key employees like Lisa Su and Mark Papermaster from the source link below. Su and Papermaster’s presentations cover everything I’ve discussed today, while the other videos available delve into aspects of AMD’s business and strategies that I don’t intend on delving into – they’re also pretty long-winded and I’m not in the mood to spend that much time on presentations that don’t have much to do with what really interests us – graphics cards and processors.
Source: AMD Financial Analyst Day Webinars