AMD stokes up its engines for Steamroller

Just like its namesake AMD’s Bulldozer processors took a while to get up to speed, but now that they’re out there’s no slowing them down. By which I don’t mean they’re blazingly fast CPUs, of course, rather that after initial delays in the launch of its newest CPU range AMD is ploughing on with rapid reiterations of the core design at a fair pace.
Originally released as the AMD FX range back in October last year, Bulldozer turned into Piledriver this spring for the launch of Trinity hybrid processors. Now AMD’s CTO Mark Papermaster has revealed more details of the second revision due out in time for Christmas, codenamed Steamroller.
I shan’t be taking bets as to whether or not the newest FX chips will ‘steamroller’ their way to success (baboom!), but it does look as though AMD is working on the right things to improve the line. So far, FX chips haven’t impressed because they’ve simply not been able to keep up with equivalently priced Intel alternatives in terms of raw computing.
The headline change for Steamroller is that AMD is giving each execution unit within the CPU core its own decode unit, reducing the number of resources shared across the chip and improving the number of instructions that can be processed per clock tick. It’s also increasing the L1 cache sizes and introducing more power efficiency measures.
That’s a pretty big change, and a bit of a back pedal for AMD.
To recap, inside a Bulldozer ‘core’ there are two traditional execution units, but they share resources at the front end of the pipleline for fetching and decoding instructions. The result is that far from one Bulldozer core being equivalent to a ‘dual core’ CPU (although they are usually marketed as such), it gets heavily bottlenecked in multithreaded tasks.
It’s probably easiest to think of it as being a bit more efficient than Intel’s Hyperthreading, but not much more so.

Bulldozer v Steamroller

The hybrid variants, which have a built in Radeon graphics core, can mitigate this to a certain extent thanks to GPU acceleration for certain applications, but it’s not enough to keep up for things like gaming performance. However, it does mean that AMD-powered Ultrabook-alikes like the HP Envy 6 Sleekbook are incredibly good value for money (I’ve tested a couple now and been generally impressed, mostly by the price and design).
By splitting up part of the front end, Steamroller should open up the throttle a little for FX processors and get instructions to the execution units faster, although AMD has cautioned that it’s likely to reduce maximum clockspeeds as a result.
On paper it makes sense – my only concern is that with or without the extra capability to feed cores, my experience so far has been that lower end FX chips and their laptop A-series variants often spend more time at 100% utilisation than Intel equivalents. So the question will be whether or not Steamroller really does open up untapped power in the design or whether it runs up against a new bottleneck further down the chain.
We’ll just have to wait and see.
(Via Anandtech)