I loathe to write articles that mention Intel (INTC) and Advanced Micro Devices (AMD) in the same title - let alone in the same article - but when I see a particularly important development in the competitive landscape, I am always compelled to write about it for my loyal readers. In this column, I would like to talk about a rather shocking development that may be even more interesting than even the ARM versus Intel debate. Intel just beat AMD at its own game - integrated GPU computing.
AMD Tried To Fight A Different Battle
When AMD bought ATi Technologies, one of the two last major players in the discrete graphics space, its grand vision was to integrate the graphics processing unit and the central processing unit onto one piece of silicon in an initiative known as "Fusion". The idea was that GPU architectures were much better suited for massively parallel computing workloads than CPU architectures were, so combining such an architecture on the same piece of silicon as a CPU that is great at handling serial workloads would lead to the next phase of computing.
I do not think AMD is wrong - this paradigm is, in my view, an inevitability. What I believe will be a problem for AMD is that massively parallel GPU architectures are very well suited to a boatload of transistors within given power/area constraints, and given that AMD's most potent competitor - Intel - has the ability to cram more, higher performing transistors in a given area than AMD can, and further given the dramatically increased level of investment that Intel is allocating to its on-die GPU architectures, I don't see how AMD can really "win" this game long term. The trend in this case is much more important than where we actually are today.
Up until now, Intel has trailed AMD when it comes to integrated graphics. With Haswell, Intel breaks through AMD's final line of defense and finally provides some very potent competition even in higher power/wattage situations.
Iris Pro - Doing Very Well At GPGPU
Anandtech, one of the world's most respected technology review sites, published a set of benchmarks for Intel's highest end "Iris Pro" graphics. To give you some background on how Intel does things with the Haswell generation of graphics, there are three "main" levels of graphics for Intel's parts:
- GT2 (20 execution units)
- GT3 (40 execution units)
- GT3e (40 execution units + 128MB embedded DRAM cache)
GT2 is known as "Intel HD Graphics", while GT3 comes in "Iris" and "HD Graphics 5000" flavors. GT3e is marketed as "Iris Pro".
So, in terms of gaming, the 47W TDP Core i7 (quad core, hyperthreaded Haswell + GT3e) is tangibly faster than the 100W TDP AMD A10-5800K. I expect the gap to shrink with "Kaveri" (and 100W parts will probably offer superior GPU performance), but since AMD will be stuck on Global Foundries' 28nm process, I do not expect too much in the way of a density improvement over the current 32nm parts. There could be a power efficiency and perhaps area efficiency improvement in moving to AMD's more sophisticated GCN architecture from the rather antiquated VLIW4 architecture (borrowed from the "Cayman" class GPUs from a couple of years ago) that the APUs currently employ in addition to the shrink from 32nm to 28nm, but if AMD cares about its gross margin profile and TDP limits, then I remain skeptical that AMD can regain the crown outside of the ~100W desktop space.
That being said, the more interesting use case is GPGPU computing rather than pure graphics/gaming applications. The hype about "HSA" and "GPGPU" is what AMD bulls typically present as evidence against Intel's mounting pure CPU lead. I use currently shipping hardware to extrapolate the potential competitive position of AMD's upcoming 512 GCN core "Kaveri" APU against the current 40EU + eDRAM "Iris Pro" in GPGPU computing applications.
The Hypothetical Comparison
AMD is not shipping any "graphics core next" based "Kaveri" APUs today, and it is likely that given the launch schedule of AMD's "Richland" that "Kaveri" will not be available for purchase by end users until about 1Q 2014 if not later. To try to estimate potential performance for "Kaveri" on the GPU side of things, I refer to the following slide from AMD's financial analyst day:
Doing some back-of-the-envelope math, we see that performance for "Kaveri" is about 32% greater than what is found in "Trinity". I also add an additional 10% to account for the fact that GCN probably allows for more of this compute performance to be "usable" over its Cayman based VLIW4 micro-architecture, implying 42% uplift over the current "Trinity" chip.
The Intel data is taken from Anandtech.
|LuxMark 2.0 (Raytracing)||CLBenchmark 1.1: Computer Vision||CLBenchmark 1.1: Fluid Simulation||Average % Relative to "Iris Pro"|
|Radeon 7660D ("Trinity") 100W||112||1539||2036||54%|
|Iris Pro 5200 (47W)||329||2234||3496||100%|
While this is not exactly a perfect comparison, especially given that I am working with next to nothing for data on Kaveri, it serves to underscore that in terms of GPGPU performance, Intel currently has the lead today with its "Haswell" based GT3e. Further, given that "Kaveri" will more than likely be arriving just as Intel gets ready to introduce its 14nm "Broadwell" tick which should have a new graphics architecture as well as a significant transistor budget/power consumption advantage thanks to being 2 nodes ahead of what AMD will be using.
Does It Really Matter For AMD? Sort-Of...
From a performance/watt and performance/area standpoint, things are about to become rather tough for AMD. Fortunately, Iris Pro equipped chips are about $650, so it is not as though AMD is losing any of its ~$100 chip business to Intel's $650 beast - but that's been the problem all along, hasn't it?
Intel's pricing power up to this point has been primarily based on CPU performance/watt in notebooks/Ultrabooks, and now that the focus has shifted towards performance in ultra low power situations, pricing power will come commensurate with this particular metric. The integrated graphics in a 45W+ power envelope is pretty much a niche (and I expect that Iris Pro was just designed for Apple (AAPL)) market as the Anandtech tests showed that a discrete Nvidia (NVDA) GT 650M is, in gaming, generally faster than the ultra-expensive Iris Pro-equipped i7 chips (and probably cheaper, to boot).
The real battle happens in the 15W and below space. The issue for AMD is that while at 100W, the gap between its own Trinity parts and Intel's Ivy Bridge parts is rather large, the picture is rather different at 17W, with AMD's Trinity holding only a ~20% lead in gaming over the comparable Ivy Bridge part, according to Anandtech:
Given that Intel moves from 16 execution units to 40 execution units, beefs up CPU performance, and lowers TDP to 15W for the top end Haswell Ultrabook parts, it seems as though that at these levels, AMD could have a very tough time competing in the premium segment on the graphics front:
(click to enlarge)
AMD's strategy really seems to be to send its Jaguar-based Kabini parts to do battle in the "Ultrabook" space, although there has been some fanfare surrounding low voltage "Richland" parts and even 15W "Kaveri" parts. While AMD bulls point out that AMD will corner the market by virtue of the fact that its parts are cheap, it is tough to understand how lacking pricing power is a bullish argument. The battle between "Richland" and "Haswell" in the 15-19W space should be interesting when AMD's partners get these chips out to the wild, and for AMD's sake, the parts really need to be competitive in both performance and power - just one doesn't cut it in today's mobile world.
But at the end of the day, AMD's turnaround isn't predicated on "beating" Intel, but instead simply selling more parts. If "Kabini" and "Temash" are priced right, then AMD could see a nice increase in volumes, total revenues, and even gross margin over the current "large dies sold at small die prices" strategy that Intel has backed them into.
It is clear that Intel is investing heavily in graphics, and it is even further evident that Intel thinks that it can get paid for said graphics. Until now, Intel has viewed integrated graphics as something that "needs to be there" but wasn't particularly important, but in the low power, SoC oriented world, there is no room for any significant deficiencies in any subsystem of the chip. And, for the first time, it is clear that Intel has taken graphics incredibly seriously with Haswell, and that there is a lot to look forward to with the Gen 8 graphics in "Broadwell", then Gen 9 graphics in "Skylake", particularly for ultra-thin and tablet form factors.
AMD is smart to focus on its lower cost APUs, as I believe that these will have superior graphics performance to Intel's upcoming "Baytrail" Atom/Celeron/Pentium lines, while at the same time coming in cheaper than Intel's higher performance Haswell. As long as AMD doesn't take any more "take or pay" charges, and as long as gross margins improve (this means at current ASPs, sell smaller dies, or raise ASPs for larger dies - the latter of which has not worked), then AMD should continue its recovery and shares should go higher.