So Why is Raven Ridge a Potential Game Changer?

Well, to answer that question, we can simply go through our bucket list of issues holding back iGPU’s. Remember, information about Raven Ridge at this point is extremely scarce and comes from rumors, so much of this is simply speculation on my part. Rumors indicate Raven Ridge to contain 4 Ryzen cores, and between 12 and 16 CU’s based on the Vega architecture, and potentially having variants with and without HBM2 memory, all built on the 14nm LPP process from Global Foundries.


First problem would be the CPU. The 7890K’s CPU is simply not great, based on a failed architecture that did not impress back in its day, and did not stand the test of time at all. Raven Ridge on the other hand, is supposed to be the first APU built on the Zen architecture, a grounds up redesign first used in Ryzen CPU’s, which is supposed to bring AMD back to parity with Intel. Raven Ridge’s CPU is rumored to be a 4 core, 8 threaded one, with each core being as performant as Intel’s modern cores, therefore being more than sufficient for most gamers. Most modern Mid to High end gaming system are built on a foundation of Quad Core i5 or i7 CPU’s, so Raven Ridge won’t be behind on that account.

Die Size and Process

Raven Ridge, like Ryzen, Polaris, and Vega, will be built upon Global Foundries’ 14nm LPP process, a modern, efficient, and performant process that is a massive jump from the 28nm SHP that previous AMD APU’s had to deal with. What this means in practice is quite simple – more hardware can be put into the dGPU than previously possible under a certain die size. Assuming a 250mm^2 die size like previous AMD APU’s, Raven Ridge could theoretically fit a full 16 CU’s, being on par with the RX 460’s Polaris 11 GPU. Assuming a 1000MHz clockspeed, a conservative amount due to the potential in Vega to raise clockspeeds, this would put the iGPU inside Raven Ridge at 2 TFLOPS of performance, nearly doubling the 1.2 TFLOPS of the Xbox One, and being slightly above the 1.8 TFLOPS of the PlayStation 4, while also using a more modern GPU architecture than found inside of those consoles.

Power Restrictions

Let us assume a 65W TDP for Raven Ridge, a very reasonable amount. We need to split that power for the CPU and GPU, and in order to do that, we need to look at what AMD’s current 14nm LPP based products can do under certain power envelopes.

For the CPU, we can look at the soon to be released Ryzen 7 1700, a 65W TDP 8 Core CPU. Considering Raven Ridge is a Quad Core, we can more or less cut the power requirements in half to get what a Quad Core Ryzen can potentially do. With some optimization and tweaking to better match the requirements of an APU, the CPU could potentially fit under 30W with a similar clockspeed as the 1700. This would provide exceptional performance, while still leaving 35W for the iGPU inside Raven Ridge. Which brings us to that, the iGPU.

The RX 460 with its 75W TDP is a bit too much for us to bear, but we don’t have to. We have a far more efficient example that we can fit comfortably inside that 35W envelope left in our theoretical Raven Ridge APU. I am talking about the Radeon Pro 460, found inside the MacBook Pro 2016. Based on Polaris 11 with 16 CU’s, this GPU can provide 90% of the performance the 75W RX 460 offers, in just a 35W TDP package! This is in large part due to the way power scales with frequency and voltage, meaning there’s a certain sweet spot that can be taken advantage of. The Radeon Pro 460 sits at said sweet spot, with a 900MHz core speed and 16CU’s, as opposed to its RX 460 brother with 1200MHz and 14CU’s. This is actually fairly conservative, as the Radeon Pro 460 is based on Polaris, while Raven Ridge is supposed to be based on Vega, AMD’s latest graphical architecture which promises to bring higher clockspeeds and another efficiency leap. And there’s the fact that the Radeon Pro 460’s TDP includes the PCB and and memory system, something that isn’t included in our Raven Ridge 65W TDP number, as the Motherboard and System Memory provide those for us.

Overall, 14nm and AMD’s latest architectures really allow the processing units to stretch their legs within a power restricted scenario like an APU.

Memory Bandwidth

Now here’s the really big issue, perhaps the biggest one of them all. The Radeon Pro 460 operates with 80GB/s of bandwidth, provided by 5GT/s memory and a 128bit interface. Raven Ridge, as opposed to the 7890K, will have the benefit of using a more recent memory standard, DDR4. A fairly viable, if somewhat expensive, memory configuration would be dual channel 3200MHz DDR4 RAM, providing a theoretical 51GB/s of bandwidth. Already a massive improvement over the mere 34GB/s found on the 7890K, and far closer to its dGPU big brother than before. However, that’s not all, Raven Ridge still has more tricks up its sleeve, in the form of the Vega architecture.

The Vega architecture promises not just improvements in power, but also in memory bandwidth efficiency. Just like Polaris, it should have Delta Color Compression, but new to Vega is the Draw Stream Binning rasterizer, theorized to be similar in function to mobile GPU’s, and NVIDIA’s Tile Based Rasterizer, found in their architectures since Maxwell, and believed to be a major source of power and memory bandwidth efficiency improvements found in that architecture. The end result should be more work done on chip cache than before, reducing memory system accesses, which both saves bandwidth and power. The Render Back Ends inside Vega, unlike in Polaris where they were clients of the memory controller with an L1 cache sitting in between, are now clients of the L2 cache. This too should help reduce memory system accesses, yet again saving bandwidth and power.

And finally, in case all of this isn’t enough, AMD has a new weapon. High Bandwidth Memory, introduced with the Fury X and Fiji. Using an interposer and a single stack of HBM2, AMD can provide up to 256GB/s and 4GB of memory to Raven Ridge. This is an absolutely massive amount, and complete overkill, not to mention the expense involved, but it is an option. A single Hi stack with very loose binning (Could make do with even half of the potential max speed), would be far cheaper than a 4 Hi stack at full speed, potentially reducing costs for AMD in case HBM2 needs to be used for Raven Ridge.

And Everyone’s Worst Enemy: Economics

There’s no denying that 250mm^2, a size equivalent to the die found in the RX 480, is a fairly large die for RX 460 levels of performance. However, if you consider the RX 480 being sold for 200$, and at least half of that going to PCB, cooler, memory, and AIB margins, you will realize that there’s potential for great profit in such an APU. An RX 460 costs 100$, a Quad Core Ryzen CPU according to rumors is supposed to cost at the minimum 130$, or 180$ for one with Simultaneous Mutl-Threading. The combined cost is already greater than that of an RX 480. However, an APU wouldn’t have a PCB, it wouldn’t have GDDR5 memory to worry about, nor would it have a heatsink in addition to the existing CPU heatsink, and it wouldn’t have an AIB cut to worry about. All of these amount to a potential 200$ APU providing 300$ of value, while having far higher margins for AMD than the 200$ RX 480. For the first time, an APU as a value proposition actually makes sense over buying a CPU and GPU alone, AND it makes economic sense for the manufacturer!

Even if one assumes a worst case secnario of having to use HBM2, the cost of the interposer and a single stack of HBM2 shouldn’t be higher than an entire GPU assembly, so margins should still be superior than an RX 480 despite using an expensive solution.

And that’s not all. Such a solution would be a major boon for OEM pre-built machines, not having to worry about a separate graphics card to provide better than the bare minimum of graphical performance, which would mean cost reductions in multiple places along the chain of production in every aspect of the machine, from the motherboard, to the power supply, and through validation and warranty. Such a part would sell by the truckloads for pre-built desktops.

And let’s not forget the power and space sensitive laptops. Raven Ridge could provide the highest amount of graphical performance ever possible in such a tiny package, and due to not having a separate dGPU, would be a major power efficiency and weight improvement. Again, in laptops, Raven Ridge could become the de-facto APU, invalidating nearly the entire stack of Intel’s mobile CPU’s, and making any mobile dGPU with less than Radeon Pro 460 of performacne irrelevant for manufacturers.

For the first time in a long time, AMD’s problem would be being able to supply the massive demand, rather than creating said demand.