Advanced Micro Devices' (NASDAQ:AMD) CS revenues depend on APUs and CPUs, and AMD just recently launched a completely new architecture at the high end of the APU stack, so a look at Kaveri is warranted.
So How Did Kaveri Do?
There are numerous Kaveri reviews available to read, and the results are pretty close to what I expected, with CPU performance remaining about the same at the high end, and iGPU and GPU compute would be where we see the largest improvements.
I am not downplaying the lack of CPU performance increase at the high end in the slightest. I personally would like to have seen an improvement on this front. Had AMD either had more drastic improvements leading to an overall more efficient CPU design (looking purely at the CPU), or had AMD achieved higher clocks without ballooning power, I think the reviews would have been more glowing.
But all in all, I think AMD made smart design choices with Kaveri, and will explain why by looking at the results.
ExtremeTech has the most complete suite of HSA specific benchmarks that I have seen. The two that I have picked illustrate two extreme cases for the chip.
This first case I think represents the potential, and challenges, AMD is facing in generic consumer workloads. In AfterShot, OpenCL and HSA acceleration greatly close the gap with Intel (NASDAQ:INTC), but don't quite close it completely. So if you're a software company, do you rush to release HSA software for AMD hardware, given that Intel owns the vast majority of the PC space?
Now, there is a silver lining to this.
There is a lot of potential in AMD's APUs from OpenCL acceleration alone, which is much more common. In a recent article, I touched briefly on the Mac Pro design win, and the importance in that this specific design win should help to accelerate the adoption of OpenCL accelerated programs.
OpenCL has really only begun starting to grow. I think this is a good way to demonstrate that even ideas that are great in theory (HSA, for example) take time to grow into something useful and meaningful. OpenCL adoption is much further along than HSA, so the compute advantages of Kaveri over Richland will likely surface in OpenCL workloads first.
MuseMage is a piece of software that uses GPU acceleration to aid image editing. Both the A8-7600 (in 45W mode) and the AMD A8-6500T are 45W chips, but you'll notice a ~38% performance increase for the 45W Kaveri.
Regardless of HSA software adoption, industry shifts toward OpenCL could tend to play more to AMD's favor (45W Richland vs. 45W Kaveri demonstrates this). This will likely not catch AMD up to Intel completely, but more serve to narrow the performance gap.
The above benchmark shows an example which is more of just a pure number crunching algorithm, and we see a massive speedup of the HSA compatible code.
I think this benchmark shows that in very, very specific instances AMD's new APUs perform extremely well.
This is another instance where AMD's APUs shine.
(click to enlarge)
Above, you are looking at percentage increases of various 100W and 45W APUs against previous offerings.
In the AnandTech review, the following comment is made:
"It is interesting to note that at the lower resolutions the Iris Pro wins on most benchmarks, but when the resolution and complexity is turned up, especially in Sleeping Dogs, the Kaveri APUs are in the lead."
To explain this sentence, think of doing the bare minimum amount of work, but accomplishing it very quickly. Iris Pro uses a massive graphics cache that drastically improves bandwidth. So if the problem the GPU is working on is a matter of bandwidth (cranking out as many frames as possible at low resolutions and minimal detail), then Iris Pro shines. However, when the detail and/or resolution is increased, the problem shifts from one of bandwidth to one of rendering power, and we see that Kaveri is able to pull ahead.
Comparing Kaveri against Iris Pro isn't something that I totally agree with either. If you search Iris Pro on Newegg, the only results you will find for Windows PCs are 2 laptops, each sporting $2k+ price tags. The most promising design for Iris Pro looks to be the Gigabyte Brix, which based on an article on AnandTech, will be launching at $530 for a Brix Pro with an Intel i5-4570R. The $530 does not include an HDD, RAM, or operating system. Admittedly, the form factor does look pretty good. But even these aren't available yet, despite Iris Pro being launched last June.
Note that I have seen articles stating that Broadwell could see some pretty significant GPU changes, so Intel will look to close the gap on the GPU portion of the equation in the back half of the year, so this is something to pay attention to as Broadwell launches. (Source: ExtremeTech)
To me at least, comparing Kaveri against Iris Pro is an apple to orange comparison given the price difference. But despite this, Kaveri does very well.
Moving onto a benchmark that is more likely representative of what Kaveri was designed for, we see even the A8 Kaveri delivering much higher frame rates than the Intel Core CPUs (and very substantially, to the tune of ~80% higher performance).
In general, my thoughts for integrated graphics performance is that Kaveri finally brings integrated graphics within spitting distance of the lower end of mainstream discrete performance. And despite being in completely different price brackets, Kaveri even holds its own against Iris Pro despite the price difference between the two. Kaveri is much smaller than the total size of the Iris Pro solution, so this is one instance where AMD's stronger graphics IP give the underdog a substantial die size advantage over Intel.
AMD Looks To Deliver The Most "Oomph" at Lower TDPs
Of Kaveri's benchmarks, this is the information that most intrigues me.
In a recent article, it was my speculation that we would see a slight regression in single threaded performance, slight improvement in multi-threaded workloads, but the biggest gains would likely be seen in GPU compute and iGPU performance. The reviews show that this guess on my part was pretty close to the mark. In the same article, I also stated that:
I think the above slide dictates the true potential of Kaveri, and that is a large percentage of the performance looks to scale to lower wattages. The 45W Kaveri chip retains 80% of the performance of the 100W chip regarding GPU performance in theoretical benchmarks.
A report from Jon Peddie Research confirms that AMD had been increasing desktop shipments while simultaneously losing notebook shipments.
Portable PCs comprise a larger overall percentage of PC sales, so a strong showing for Kaveri is doubly important in the mobile space. Increasing performance at lower TDPs should be an important focus for AMD, as this is where it was losing the greatest market share, and the overall TAM is larger.
I think here the results speak for themselves, where AMD is showing the largest performance increases when looking at these chips using a performance/watt mindset (source: SemiAccurate).
SemiAccurate notes the 45W Kaveri delivers ~80% of the performance of the 95W Kaveri chip.
AnandTech's demonstration of an approximate 40%-50% increase in iGPU performance and sizeable gains in CPU performance shows Kaveri's gains at lower TDPs.
To tie this back to notebook, notebook processors are much more power constrained than the desktop counterparts. AMD did not change pure CPU performance appreciably at higher TDPs, but at lower TDPs the story looks like it could be different due to higher clocks at the same TDP. AMD's iGPU performance also has a nice performance gain across the board.
More On Power Consumption And Relating It To The Data Center
Bit-Tech shows that under a workload that is likely unrealistically demanding (both the CPU and GPU cranking away as hard as they possibly can) that between the 45W A8 Richland and 45W A8 Kaveri, power consumption goes up by ~18%. However, base frequency is increased 47%, demonstrating that Kaveri is able to achieve higher clock speeds at lower TDPs when compared against Richland, signaling a nice boost in performance/watt at lower TDPs.
Again, this is stressing both the GPU and the CPU. Since the chip is almost 1/2 GPU, likely a good portion of the power consumption is coming from the graphics side, meaning that only part of the listed power consumption is coming from the CPU.
Source: AMD's Roadmap
Notice that AMD's roadmap has the server version of Kaveri listed as both a CPU and an APU?
The video in which the image above was taken was originally pointed out to me by SemiAccurate forum member "ptMd," and the video appeared on InsideHPC, but has since been taken down (my guess being that the conference was supposed to be covered by an NDA). WCCFTech was able to grab a screen capture of the point I wanted to discuss.
The reason I feel this is important is because Berlin as an APU will be used in clusters designed for big data, HPC, and media. But notice there is also a pure CPU version that is aimed at typical work center loads?
CPU-World shows 2 Opteron 3300 series quad core server chips (I, II) based on Piledriver architecture. These chips have TDPs of 25W/45W, with the 45W chip operating at a base frequency of 2800MHz, and up to a boost frequency of 3800 MHz, in a total die size of 315 mm^2.
AMD is fielding a 45W Kaveri APU with 384 graphics cores, along with a quad core CPU, operating at 3.1 GHz/3.3 GHz boost. The same chip operating in a 65W TDP runs the CPU at 3.3 GHz base/3.8 GHz boost. If the GPU isn't being used, this allows even more power (higher clocks) for the CPU.
Berlin may end up being a good fit for AMD's SeaMicro servers. However, before I get my hopes up here I would like to see some design wins. AMD has very little server market, so any growth in this area would be welcome.
Now that we know Kaveri's performance, the only outlying questions that remain for me is yield and design wins. During the earnings call, CEO Mr. Rory Read stated that AMD was "over indexed" in notebooks, so the tablet encroachment on portable computing really hurt AMD during 2013. If AMD can score some nice design wins in mobile, 2014 could shape up to be a better year for Computing Solutions than 2013.
I think many make the mistake of looking at AMD and Intel through Highlander ("There can be only one") goggles.
The relationship between companies would be better modeled by a predator/prey type situation. When one company tips the ecosystem towards its favor, the population (market share) will shift, dragging revenues and earnings along with it.
To quote Thomas Ryan's article on SemiAccurate:
We all knew going into this review that AMD would have a performance advantage in video games, what we didn't know is that it was going to be so pronounced. The A10-7850K offers between 1.8 and 4.5 times higher framerates than the HD 4600 graphics solution on Intel's i7-4770K. That's a massive performance advantage that averages out across our testing suite to 2.5 times higher framerates on AMD's hardware versus Intel's.
One thing to call specific attention to is that Thomas' review of Kaveri is at playable framerates, and the A10 Kaveri offers 150% higher frame rates over an i7-4770k; that's a massive difference.
League of Legends, World of Warcraft, and Dota 2 are among the most popular games in the world, and Kaveri delivers more than playable framerates at full HD resolutions and high settings in all these titles. Note that in Dota 2, HD 4600 does come close to the average FPS, but the Kaveri chip delivers a substantially higher minimum FPS, giving a much smoother game play. In the other 2 titles (and most games in general), AMD smokes Intel's HD4600.
Kaveri looks to further improve AMD's graphics lead. At the highest TDPs, CPU performance isn't drastically improved, but at lower TDPs Kaveri's ~10% clock for clock advantage over Richland, along with higher overall clocks, will make a much stronger mobile offering. And I often find reviews on American websites irksome, simply because testing is done at ludicrous settings for APUs using the most demanding titles. This is somewhat odd, as some of the most popular games in the world do not require the most demanding hardware to run, so consumers that only want to play Dota 2 or WoW and surf the web would overspend by not buying an APU, unless they just really wanted the extra performance, price be damned.
Having competitive silicon is only part of the equation however. Now AMD needs to market and sell these products. So while right now I think Kaveri could be a great chip in mobile platforms, my hopes will be much higher once I see the design wins.
Specifically for servers, I'm more on the fence until I see some verifiable traction and design wins. Comparing the volume of the server market to the consumer notebook market, in my opinion the notebook market represents the biggest opportunity for AMD.
I would like to see Kaveri laptops more aggressively marketed by AMD and OEM partners at launch. Hopefully we'll see the company greasing the skids by sending out 10 or so laptops to review sites, along with requests that the reviewers test the games these chips are best suited for (F2P MMOs and MOBAs).
Finally, it seems like many are already putting the nail in the coffin for Kaveri, based on the 2 highest wattage APUs.
The 35W A10-5750M clocked at 2.5 GHz/3.5 GHz boost 32nm part. The 45W A8-7600 is a 3.1 GHz/3.3 GHz boost part 28nm part. If AMD manages to field an APU that clocks higher on the CPU and has a higher performing GPU, this will be a step in the right direction.
AMD has had declining notebook shipments in 3 of the past 4 quarters, with the exception of Q2 when the company had explosive growth when Kabini and Temash were launched.
Given this, a strong showing for Kaveri in mobile would be a nice step in the right direction, but it's up to OEMs to build designs that will actually sell.
Disclosure: I am long AMD, INTC. I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.
Additional disclosure: I own shares and calls in AMD, along with puts that I will either exercise or sell. I actively trade my AMD position, and may add or sell part of the position at anytime.