AMD's A10-5750M Review, Part 1: The APU and Radeon HD 8650G Performance
by Dustin Sklavos on June 28, 2013 12:00 PM ESTConclusion: Surprisingly Potent Refresh
Without getting into the nitty gritty of the MSI GX60 that houses our review APU, AMD's A10-5750M, it seems like AMD's new APU is a net victory...with some caveats.
I imagine at the time of Richland's development, graphics performance wasn't a huge concern. With Haswell, that performance target shifts a little bit, but as long as we're not heavily CPU-bound, AMD continues to offer superior mobile graphics. Richland seems to be most aggressively focused on shoring up the primary weakness of AMD's existing architecture: the CPU itself. Piledriver was a decent improvement on Bulldozer and, ironically, a mild one on Llano, but it's nowhere near enough. Bumped up clock speeds and improved turbo help close the gap at least a little bit, but we're still dealing with 35W AMD APUs struggling to hit the same performance levels at 15-17W Intel chips.
Arguably more impressive is that this refinement was possible at all. AMD was able to take Trinity, tune the silicon ever so slightly, and extract a healthy gain in CPU performance from it. Graphics performance seems to have held flat from Trinity, but an essentially free performance boost on the CPU is welcome.
I am, however, forced to address a few elephants in the room that are getting glossed over. As consumers, we need AMD to succeed. Lack of competition is showing in a major way: desktop Haswell is a joke, Haswell's GT2 IGP is a minor improvement yet promises to be the most common one in Intel's lineup, and Intel seems to be planning to mostly coast on Haswell for two years while focusing on Atom's successor.
Yet the ugly truth about Richland and Trinity before them is that these reviews are covering the fastest models available. A8 chips lose a third of their GPU hardware across the board, and A6 chips lose another third. If you do some quick and dirty math, that means that anything below an A10 is going to be almost directly inferior to Haswell or Ivy Bridge. Graphics performance will at best be slightly above parity, while CPU performance takes a bath.
The other problem is that AMD is still targeting 35W as the mainstream TDP, but that's a target that's actually quietly shifted in the marketplace. This isn't 2008 anymore and Intel isn't charging a fat premium for its low-voltage hardware. The market that needs 35W CPUs is shrinking, being devoured at the low side by tablets and systems with ULV CPUs that still offer enough performance to handle the majority of tasks end users will need them for. And anyone who needs more performance than that can simply make the jump to a system with a quad-core Intel CPU that has more muscle. In this reviewer's opinion, 35W isn't the target, it's the halo. 15W-17W is the target, and while AMD has offerings at those TDPs, they're woefully uncompetitive.
AMD doesn't just need Kaveri. We need Kaveri. We need the Steamroller architecture update, and we need the graphics cores to switch over to GCN from VLIW4. Hopefully AMD will be able to produce a Kaveri part that has a fighting chance against Intel at 15W/17W, since Kabini and Temash are destined for smaller form factors. For now, the Richland A10-5750M is a good option and a solid offering for a refresh, but I don't think anyone can rely on it as a stopgap for too much longer.
Facebook
Twitter
RSS
82 Comments
View All Comments
Vi0cT - Friday, June 28, 2013 - link
Still don't think 5100 can do a lot to the AMD A10-5750M GPU, I believe that at best it can offer equal performance to it and AMD stills wins in the price/performance (take into account driver optimizations too since ). Also VLIW is a more parallel architecture, so it should fare better in non-mainstream GPGPU stuff, however who runs that kind of stuff on a mobile device is beyond me, the best usage case I can think of (for a notebook) is Nebula 3 and the Volterra kernels on CUDA(Professional Audio) but that is on NVIDIA hardware.Still believe the market for both is a bit different, I mean Intel is focusing into getting more mainstream devices meanwhile AMD is kind of stuck in this kind of devices (GX60).
Even being better than the Intel solution price/performance-wise any design aimed at mainstream will prefer the Intel's lower TDP and power consumption.
whyso - Friday, June 28, 2013 - link
Look at the 4770k igp results from the iris pro article. It looks like HD 4600 SV is quite on par with even the top range of AMD's apus (4770k is around 10-15% faster than trinity, richland is marginally faster than trinity [sometimes slower] so it looks like 8650G is barely faster than HD 4600).Vi0cT - Saturday, June 29, 2013 - link
In the desktop Richland is 34% faster than the HD4600 in avg.The only comparison point we have here for now is 3DMark 11 (Sadly :/) and that puts Richland at 29% advantage vs the HD5100 in the MBA.
Again the peak theoretical performance difference between the HD4600 and the HD5000 is 63%. Between the HD5000 and the HD5100 there's is only 18.3% even taking into account throttling the best I can see the HD5100 doing is being on par with Richland.
Also take into account driver optimization and the optimization for AMD/NVIDIA's GPU architectures.
Still with Richland we're still talking about VLIW, with CGN (Kaveri) AMD will gain ground against Intel.
If done right AMD can get this performance into the 15-20W targets and beat Intel at 28-35W with less problems in CPU intensive games.
Khato - Friday, June 28, 2013 - link
Pricing is definitely a legitimate concern... too bad that we have no clue how much AMD's mobile parts even cost isn't it? They don't even publish recommended pricing. If we compare similar models that differ only in their processor manufacturer (Lenovo Edge E531 vs E535) then you're talking a $50 premium for the Intel chip on the base offering... though Intel's recommended pricing is the same for the upgraded i5-3230m as the base i3-3110m, so let's figure that the i5-3230m actually costs that recommended $225. In that case we arrive at roughly $100 for the base AMD Trinity and then something around $200 for the A10.Regardless it doesn't matter too much. For comparison, Iris Pro models start at a recommended price of $440 for the i7-4750HQ. And then all the non-i7 28W Iris parts have a recommended price of $342. But yeah, who knows what actual prices are? The i3-4158U (28W Iris 5100 part) may well be cheaper than AMD's A10-5750m. And as esgreat already stated, expect Iris 5100 to be quite a bit faster than HD 5000.
Vi0cT - Friday, June 28, 2013 - link
I'm pretty sure AMD is giving lower prices to OEMs (else they will get almost 0 design wins) many OEMs will try to get extra margins from the AMD. Also don't underestimate 50 USDs. In developing countries (being from one myself) what looks like "50 bucks" makes a huge difference for the avg Joe.It's the difference between the guy actually buying or not the AMD one.
Khato - Saturday, June 29, 2013 - link
Oh, I wouldn't be surprised if AMD is giving lower prices to OEMs for their low-end models. But they want to make more money on their 'high end' SKUs same as Intel. Heh, and if Intel's i3-4158U outperforms AMD's top end mobile offering then I'm sure that the AMD part will be priced a bit lower than it. But it's hard to claim that AMD has a massive lead in price/performance if that's the case. Regardless, it's all speculation since neither AMD nor Intel give real pricing information for mobile parts.Vi0cT - Saturday, June 29, 2013 - link
However we can see it like this, maybe both sell the SKUs at the same price (let's say 5% more or less whoever gets it) and the OEMs get those "50 bucks" just cause it says Intel (there is a lot of change that is happening), then it's not Intel's pricing problem :PVi0cT - Saturday, June 29, 2013 - link
there is a lot of chance*FwFred - Saturday, June 29, 2013 - link
Why would you think the maximum turbo rate has anything to do with the average GPU clock for a 15W part vs 28W?kyuu - Saturday, June 29, 2013 - link
There is no mobile GT3e part, so what are you even talking about. Or does your definition of "fair" somehow involve comparing mobile Richland to desktop Haswell?