AMD to return to mid-range CPU’s – FM2+ as the great unifying platform

AMD has a problem: The AM3+ platform with its 9xx chipset is an evolutionary dead-end whose feature-set is not going to be improved and is barely competitive in the low-end, let alone the mid-range where it purports to belong. Worse still, it is tied to a processor ranger that does not have the performance to compete at the mid-range (>£120), and no longer has the cachet of a halo product in public consciousness. Taken together, CPU and chipset form a platform that is heavily integrated into technologies now defunct on the desktop, such as hypertransport, with high manufacture costs combined with low returns.

In short, it does nothing for its future, represents only a support cost in the present, and some nostalgia in the past. AM3+ is dead.

There will never be an AM3+ SteamrollerFX, and it seems like AMD are content to ride Piledriver into the grave to milk as much ROI from its 32nm facilities as it possibly can.

USB 3.1 = no, SATA Express = no, PCIe 3.0 = no, Fast fabric interconnect = no, On chip PCIe = no, IOMMU 2.5 to enable HSA = no.

What AMD have said is that they will be making CPU’s at 20nm, 28nm, and 32nm right through 2015. They have a ton of 32nm fabrication technology, which is the reason why they won’t tell us that AM3+ is a dead platform, because they need people to keep on hoovering it up.

This creates a problem; how to keep the average selling price of your silicon high-enough to turn a profit after investing in the next generation of tech that must inevitably follow?

AMD have FM2+ for their low-end, and they have their A-Series “Kaveri” APU’s to sell in this low-end. By way of explanation, I divide the market segments as follows:

Low-end = £60 – £120 / Mid-range = £120 – £240 / High-end = £240 – £480

However, Kaveri was targeted as a performance/watt design, with four cores, 512 shaders and a average power budget of 55W (scaling 15W – 95W), this was designed for mobile usage and low-price computing where an add-in graphics card is unlikely. Something of a shame for a company that covers the whole computing ecosystem, and would love to sell you a high-end “FX” brand CPU/APU and platform to go along with their high-end ~3000 shader GPU (~£400).

Gaming does drive high-end PC sales, and a rough metric to work to is a 1:1:2 cost ratio for a single WQHD screen at 2560×1440. If you are spending £380 on a shiny new AMD 290X you want a £180 CPU and a £180 motherboard to properly complement it. Yes, in theory you could buy one of their 220W ‘specials’ – such as the AM3+ 9590 – but if you are spending this much in late 2014 would you really go for a end of the line platform with no upgrade path and a huge deficit of features to be found on the competition?

Remember, in an era of annual architecture and process upgrades AMD has given us:

32nm Piledriver for 2012, 32nm Piledriver for 2013, 32nm Piledriver for 2014, and 32nm Piledriver for 2015…

Q – What is the solution?

A – Make your FM2+ platform sufficiently feature rich and perfofmance flexible that it can scale the low-end and the mid-range (up to £240).

47% of the kaveri die space is GPU. If you consider that roughly 20% is uncore, that leaves roughly 33% as CPU. Give our take, 8 shader cores is fifty percent larger than 4 cpu cores. You could double that cpu portion to 66%, and still leave 14% for shader cores. Or, make the total die size just 10% bigger and you have an 8 cpu core APU with 4 HSA enabled shader cores ready to grind through FPU work. Pretty much die-size neutral.

Wouldn’t that be worthy of an “FX” title?

The problem with this is what we know about Carizzo, AMD’s next big-chip architecture using the Excavator Core, and whatever the latest variant of their GCN shader architecture is. Four cores and 512 shaders, using DDR3 and and a 28nm fabrication process does not an exciting “FX” APU make, not when you consider that AMD traditionally use the same silicon for both mobile and desktop. Just like Intel.

Here is where it starts to get interesting (sorry it took so long):

Intel isn’t doing that right now. Right now we know that while they are introducing their new 14nm Broadwell uarch for mobile in Q1 2015 it isn’t going to hit the desktop until at Q3 2015. Technically, they will be using separate designs for desktop and mobile silicon.

Added to this we have four tidbits of information from AMD that all point in a specific direction:

1. AMD’s is rumoured to use stacked DRAM for Carizzo. Fine on mobile which is always a semi-custom platform, but how do you implement this on the bog-standard FM2+ desktop?

2. Carizzo is rumoured to arrive with a grand total of 8x PCIe 3.0 lanes on-die (kaveri has 24x). Fine for mobile, but a difficult cross-sell for expensive GPU’s and Sata Express SSD’s.

3. Oops, another rumour that says that Carizzo is a mobile only product!

4. The final rumour is that AMD is teasing us with a new “twelve core” APU product…

WCCFTECH talk about it being possibly 12 GPU cores, or even 12 CPU cores, but don’t make the logical leap that it could be 8 CPU cores [and] 4 GPU cores, i.e. the mirror opposite of Kaveri/Carizzo. It could even be six and six.

None of this is advanced tech, or using niche and bespoke solutions for marginal purposes.
It is bog standard steamroller cores, as used in kaveri. It is bog standard GCN shaders, as used in kaveri. It is bog standard 128b DDR3 2133 interface, as used in kaveri. It is bog standard uncore elements, as used in kaveri.

High x86 performance, in excess of their existing Piledriver products, and a modicum of transistor dense GCN shaders sufficient to keep HSA relevant across the AMD product range (but not eating so much of the power budget that it would retard CPU speeds). The FM2+ 100W power budget is plenty to provide a 12 ‘core’ APU using a mature 28nm process.

All this would rather indicate they must have different plans for the desktop, plans involving more PCIe lanes (remember Kaveri has 24x), and presumably higher performance parity with Intel in order to keep the average selling price reasonable. ASP being the end goal here; there is little business efficiency in AMD servicing the high-volume/low-margin market of the CPU/platform market when it does not allow them to cross-sell with their low-volume/high-margin GPU’s. They, like every business selling entire systems, would like you to buy the whole package, and [not] sacrifice the high-value halo products to their competitors.

The logic of this Kaveri-on-steroids product is that much like intel AMD would bring new architectures to mobile first, and then transition them to the desktop.

Yes, I’m stretching here, but that is what this blog is for.

Update 10.01.2015 – Carrizo is a SoC, not an APU with a separate southbridge, and will not be arriving on AMD’s APU platform FM2+