The buzz around this benchmark stems from the low 'SDP' envelope of Intel's Bay Trail. I am long both AMD (AMD) and Intel, so I typically follow much of the news regarding Temash and Bay Trail, as I feel these are both important chips for their respective companies.
In this article I will focus mainly on power consumption and performance per watt, as these are key metrics for mobile chips.
Two important terms to mobile chips are 'TDP' and 'SDP'. Intel defines TDP as "...the near maximum power a product can draw for a thermally significant period while running commercially available software." AMD uses essentially the same definition.
Intel uses an additional definition -- 'SDP'. 'SDP' is defined as "...an additional thermal reference point meant to represent mainstream touch-first usages. It balances performance and mobility across PC and tablet workloads to extend capabilities into thin, thermally-constrained designs."
So to compare AMD and Intel power consumption envelopes, you must understand the differences between these measurements.
Intel's i7-3689Y chip is I believe an 800 MHz chip when operating in its 7W 'SDP' envelope, a 1.5 GHz chip in its 13W envelope, and can boost to 2.6 GHz for short periods. TDPs and SDPs have nothing to do with actual power consumption however -- think of them more as limits.
Note the power consumption between the A4-5000 Kabini part from AMD and Intel's i3-3217U ULV part, courtesy of THG.
The 17W 'TDP' i3 used ~75% more power than the 15W 'TDP' Kabini. The dual channel memory configuration of the i3 did however help it achieve a higher frame rate during game play. I am using this example only to demonstrate the disparity between TDP and actual power consumption.
A Peek At Performance
Earlier I had estimated that Bay Trail would have roughly a 40% CPU performance advantage over AMD's Temash, while Temash would have around a 15% GPU performance advantage.
As of right now, the only benchmark we have as far as I know to make a somewhat direct comparison between Bay Trail and Temash is Cinebench.
According to notebookcheck, the 8W A6-1450 Temash scores a 1.02 while consuming between 3.4W to 4.6W. Since we are talking about very small TDPs, it's important to note that the measurements here are not adjusted for inefficiencies in the voltage converter, so the wattage I have shown is likely slightly higher than actual power consumption.
Based on Bay Trail's score of 1.47 compared against the A6-1450's score of 1.02, my 40% guesstimation was not too far off. But actual power consumption of Bay Trail is still missing.
Why Am I Skeptical of Bay Trail's Power Consumption?
The above information was posted by VR-Zone in early July. The one thing that stands out to me is that the 4 core Z3770 chips are rumored to be around 2W 'SDP'. Here I will point out that all of the chips above are rumored to be built on the same 22nm process based on the Silvermont architecture. The Celeron N2910 is the closest description to the z3770, and it is a 7.5W TDP chip. Careful binning and power management along with a lower clocked GPU could allow for the z3770 to fit in a lower thermal envelope and maintain most of the performance of the Celeron N2910.
Secondly, the GPU supposedly consists of 4 EUs, which are Intel's IP blocks for the GPU. HD4000 consists of 16 EUs, and on the mobile i5-3317U for example the GPU frequency varies between 350 MHz and ~1 GHz. Based on leaks as reported by CPU-World, we can see the rumored max GPU frequency of the Bay Trail parts is 792 MHz. Given that Bay Trail is rumored to feature 1/4 of the EUs of HD4000 and looks to run at 80% of the max frequency of the i5-3317U, GPU performance could end up somewhere around 20-30% of HD4000.
Bay Trail is an important chip for Intel. I feel it will be very competitive, but at least to me there is still some uncertainty. At the end of the day design wins and pricing power are orders of magnitude more important for investors, but the performance factors into both of these. Some of the uncertainty will be lifted next week at IDF, but before that I wanted to give some points to keep in mind.
Anand measured the power consumption of each platform above with the battery removed and the display disabled.
We can see two things from AMD's Kabini PR slide: idle platform power consumption of the 15W Kabini is around 4.08W, and that platform power consumption goes up somewhat during load. Also, the idle power consumption is .77W. Power consumption goes up by about .3W while web surfing, and by just over 1W while watching video.
This footnote explains the testing details of the platform used.
While it is true that Bay Trail ties with the 15W TDP Kabini in at least one benchmark, I have shown TDP is not indicative of power consumption.
The reason I am comparing this chip to the A4-5000 Kabini is because many tech writers (I,II) have pointed out that the Cinebench score of Bay Trail is equal to that of the A4-5000. The fact that it ties the 15W 'TDP' Kabini isn't nearly as big of a deal if it consumes almost as much power while doing so, and we don't know the power consumption of Bay Trail yet - just a guess at the 'SDP' envelope.
The last thing I will point out is that there are 2 main parts to an SoC: the GPU and the CPU. There are benchmarks designed to stress both, but the best benchmarks in my opinion are real world applications. Gaming is a great example because it taxes both the CPU and the GPU. Real world benchmarks, in my view, are by far the most important because they're a little more difficult to spoof.
I went into some GPU expectations above to show the possible performance of the GPU. The unknown here is what kind of power consumption and performance we can expect when both the CPU and GPU are taxed. Intel's newest 15W Haswell i7-4500U (with an even larger GPU) struggles to manage any sort of gaming experience in many modern titles, and Bay Trail's GPU will probably be far less powerful than the GPU in the i7-4500U. It may seem like an unfair comparison, but the competition from AMD is the Temash A6-1450 which has a higher TDP.
There is only so much performance that can be crammed into a 2W SoC, and that 2W could be a hindrance depending on the application. I feel this is a valid concern as we already have an idea of the performance of the Gen7 EUs that are rumored to be used, but have no idea how they will behave in such a thermally constrained environment. Given that it's a 2W 'SDP', it's possible that under certain conditions this figure can be exceeded, but this in turn would cause Bay Trail to lose part of its appeal.
Intel's Bay Trail is an important offering in that it brings a nice leap in performance over Clover Trail and gives Intel a product that can be used in fanless tablets, allowing the market to deliver x86 tablets at much cheaper prices than their bigger Core based CPU counterparts.
As such, it is highly anticipated and there are likely several hardware reviewers foaming at the mouth ready to benchmark this thing. I look forward to next week when we finally see the veil lifted and the ensuing reviews. The tablet market represents a 200M+ unit opportunity for Intel.
Additional disclosure: I actively trade my AMD and INTC positions. I may add or liquidate shares at anytime. I am long AMD both in shares and options.