AMD’s new A10/Trinity chip is not that impressive from a checklist point of view, but under the surface, it has a killer feature. Most people don’t understand energy use and performance per watt, but that is Trinity’s ace in the hole.
Trinity launched a few days ago with five chips mobile chips, capped by the A10-4600M part, a 2.3GHz CPU with a 497MHz GPU. These numbers seem a bit underwhelming, especially in light of Intel’s 600Mhz advantage with their Ivy Bridge based i7-3920XM. Intel’s GPU is also much faster at 650MHz, but lets face it, Intel’s GPUs suck, when they work right, and that is a rare day indeed. That said, AMD’s CPUs suck, but they do work right. AMD’s previous generation Llano CPU was woefully behind both new CPUs with the A8-3550MX coming in at 2.0GHz and 444MHz for the GPU.
Things get a bit nicer for AMD when you take turbo in to account, with the following numbers quoted as base/turbo. Trinity now is listed as a 2.3/3.2GHz chip with a 497/686Mhz GPU while Ivy is a 2.9/3.6GHz GPU, 650/1300MHz GPU. Llano takes up the rear at 2.0/2.7GHz CPU, 444MHz GPU, no turbo on the Llano generation GPUs. That said, these are just specs, and mean about as much to the average user as package dimensions.
Numbers aside, Intel’s Ivy Bridge embarrasses Trinity in CPU performance, it is in a different class, and Llano, well, it is notably slower than Trinity. On the GPU front, Llano smokes the non-functional last generation Intel Sandy Bridge GPU (Note: They never got it working before replacement, what a joke.) and runs tangibly faster than even Ivy Bridge. Trinity is much faster still. Net result, Intel wins on CPU performance, AMD on GPU performance, and both are runaway victories. The one area where both are were even was noxious, pointless, user abusive naming schemes. AMD was a mere one ahead with A8 vs i7, but Trinity pulls out a lead with A10, they are ahead of Intel by three now!
Back to the real world, one important factor here is energy use. The Ivy Bridge i7-3920XM is a 55W CPU, the A8-3550MX is a 45W CPU, and the A10-4600M is a 35W CPU. What if they were all 45W, how would it stack up? Interesting question, but there is no 45W mobile Trinity this time around, it only comes in 17, 25, and 35W versions. Stepping down to the 35W level, things get more interesting, far more interesting.
The top 35W Llano is A8-3520M, a 1.6/2.5GHz CPU with a 444MHz GPU, so we know where that generation of AMD products chew through energy, the cores. Intel’s top older generation Sandy Bridge appears to be (Note 2: Intel’s web site gets progressively worse every time we visit. Finding real information is actively being hindered by Intel now, quite sad, so there may be faster 35W chips we can’t find.) the i7-2640M at 2.8/3.5GHz CPU, 650/1300MHz GPU. This however is a dual core, Intel could not make a 35W mobile 4C Sandy Bridge, note 2 applies here though.
On current parts, Intel’s i7-3612QM is a 35W quad core mobile part, and about the closest competitor to the A10-4600M. It runs at 2.1/2.8GHz CPU, 650/1100MHz GPU, but has two threads per core vs AMD’s one. Please take note, AMD has a 200MHz base and 400MHz turbo CPU advantage here, but the GPU architectures are radically different so can’t be compared on this basis. Intel only appears to lose 200MHz on the turbo clock, but that is pretty substantial in terms of percentages, more so in light of Intel’s anaemic GPU.
Comparing the numbers, when you go from a 55W to a 35W Ivy Bridge, the lower wattage part has only 73%/78% of the CPU speed, 100%/85% of the GPU speed. Similarly, a 35W Llano runs at 80%/93% the CPU speed of a 45W Llano, the GPU is at the same clock on both parts. As we mentioned earlier, there is no 45W mobile Trinity to compare against, but it would obviously be faster than the 35W version. Interesting, no?
Most benchmarks show that, clock for clock, Trinity is a bit ahead of it’s older brother, Llano. Trinity’s meager performance per clock gains are what people notice, but the clocks went way up in general, and up massively at the same wattage level. AMD basically took the performance gains with Trinity, and manifested them in energy savings. While this is arguably the most important number for mobile users, it is also the hardest thing to explain to the average person. People just can’t wrap their heads around performance per watt, but raw performance is easy, as is battery life. AMD took the hard road for PR with Trinity.
On paper, the overwhelming Ivy Bridge CPU speed advantage is still an overwhelming CPU speed advantage when stepping down to a 35W part, but less so. The overwhelming Trinity GPU speed advantage gets even more so vs a 35W Ivy Bridge, and things are likely disproportionately worse for Intel because of the attendant CPU speed loss in games. The few games where Ivy does well are largely CPU bound, not GPU. Basically, this 20W drop is a double whammy for Intel in games and heavy mixed utilization scenarios, aka the real world.
One factor that matters a lot, in both good and bad ways, is turbo. The idea behind turbo is that a chip has a set power use limit, also known as a TDP. The base clocks are a conservative estimate of what can be done at those power limits, and therefore thermal limits. Pre-turbo, that is what you got, but turbo adds smarts to the mix. If a CPU or a GPU is not using its allotted percentage of the TDP, that extra can be funnelled to the unit that is using its maximum. Basically, Turbo allows the CPU or GPU to exceed its limits when the other unit is at low utilization.
One nasty side effect of this is that traditional benchmarks that stress components one at a time, are the best case scenario for a turbo wielding chip. They show the absolutes of the chip correctly, just not what is real world achievable. Those theoretical numbers are so rarely achievable in real life that it is laughable. Benchmarks that stress a single function basically have become a test for turbo functionality, they no longer represent the components they purport to test. This is the long way of saying they are worthless on a turbo CPU.
Real world scenarios, multitasking benchmarks, and the like are all being thrown out there to try and address this problem. They fail badly. Some, like BAPCO’s SYSMark, are so thoroughly gamed that they have become nonsensical, and others try futilely to define average. Most however are just plain dumb, and that is the state of benchmarking a modern turbo enabled chip, dumb. SemiAccurate’s position on the subject is, well, complex, but some call that dumb too, it isn’t simple enough for the masses.
In the end, what do we have? We have an AMD chip that, on paper, appears to better Intel where it matters most to the user. If you think Trinity is underwhelming on paper, look at comparable numbers. AMD looks to use energy more efficiently than Intel, but until we get an Ivy Bridge at 35W in the lab, we can’t say for sure. Several tests like Tech Report’s, seem to show AMD leads Intel quite handily in battery life for real world usages. Those same tests also show AMD crushes a 45W Intel Ivy Bridge in GPU, and is crushed by the same chip in CPU tests. That means what CPU you chose is more a question of your workload, do you run HPC tests or a web browser? Movies and games or multi-hour statistical analyses more often?
Would a 35W Intel CPU change the results? Yes. Would AMD take the lead in CPU against that part? Not a chance. Would those 20W make a difference in GPU tests? Undoubtedly, Ivy would go from borderline working graphics to a lot slower, when the drivers actually worked. Who would have expected this state of affairs before Ivy launched last month?
And all of this summarized says one important thing, AMD has a very credible chip on their hands with Trinity, and Intel should be very worried. What looks to be a minor gain in top line performance is a massive leap forward in performance per watt, and that is what matters to the mobile user. AMD can now get in to the thin and light notebook category, and appears to have a better ‘Ultrabook’ CPU than Intel does, but we will have to wait for 17W parts in the wild to say for sure. AMD will be vastly cheaper too, and have working graphics, can Intel address either one? Did performance per watt increases, not top line performance, give AMD a win this round?S|A
Update: Had a 55W Intel CPU mis-identified as 45W, and therefore a 10W delta in some places when it should have been 20W. Fixed.
Latest posts by Charlie Demerjian (see all)
- Altera puts hard FP functions in Arria 10 and Stratix 10 - Apr 22, 2014
- Samsung lays down the design law to Apple - Apr 22, 2014
- What kind of numbers are Intel contra-revenue funding? - Apr 21, 2014
- SemiAccurate was right about Apple’s foundry plans - Apr 17, 2014
- Global Foundries and Samsung sync up 14nm processes - Apr 17, 2014