It’s been a rocky year for AMD. In retrospect, Bulldozer’s disappointing performance was the tip of the proverbial iceberg. Since it launched, Sunnyvale has laid off thousands, is preparing to lay off thousands more, cancelled its first set of 28nm chips based on the Bobcat architecture, and has lost ground overall in the PC market.
Multiple executives have fled the company, margins have tumbled, and long-term prospects are dim, at best. In the midst of all this chaos, AMD’s engineers have plugged away at refining Bulldozer. The first APU based on Piledriver, codenamed Trinity, debuted last spring; the desktop variant hit shelves at the beginning of this month.
Today, AMD is rolling out the Piledriver core to Socket AM3+ motherboards as well. These new chips, codenamed Vishera, are supposed to deliver a bevy of improvements that address some of Bulldozer’s weakest areas.
Rather than running Vishera through a broad range of consumer tests, we’ve focused on a few benchmarks designed to highlight the low-level differences between Vishera, Bulldozer, and Thuban as well as Intel’s Ivy and Sandy Bridge. Does Piledriver fix ‘Dozer, or at least significantly improve its value proposition? Let’s find out.
Note: AMD’s Turbo Core and Intel’s Turbo Mode were disabled on all chips, in order to prevent them from adjusting the CPU’s clock speed and throwing off results. As a consequence, the results here will be lower than in a standard review, particularly for single-thread performance.
There are three key areas where Piledriver must improve on Bulldozer in order for AMD to win back the ground it’s lost to Intel.
Single-thread performance: Bulldozer’s single-thread performance was significantly lower than AMD’s previous generation CPU, codenamed Thuban. This fact, combined with the core’s inability to hit its target clock speeds, fundamentally crippled the processor.
The FX-8350 attacks one of these problems off the bat, as the chip’s base clock is 11 per cent faster than the 8150′s, and its Turbo Mode pushes four cores up to 4.2GHz (the 8150 topped out at 3.9GHz when four or more cores were active).
Power consumption: Bulldozer, like the original Phenom, was a power hog. This had direct consequences for the chip’s ability to scale. AMD was rumoured to be working on a new revision of Bulldozer that would reduce power consumption, but the core (if it existed) never saw the light of day. Bringing the chip’s TDP into line is essential to gaining design wins in smaller form factors.
Multi-core scaling: When AMD created Bulldozer, it promised that it could offer strong scaling while sharing core resources. In practice, Bulldozer’s scaling was around 80 per cent that of a traditional dual-core design. This wasn’t an outrageous penalty, given how much die space AMD had saved, but it further hurt the chip in comparisons against Thuban. Eight BD cores at 80 per cent of Thuban scaling often looked a heck of a lot like six Thuban cores.
On a high level, Vishera looks just like Bulldozer. Its caches are all the same size, the same set associativity, and the same functions are shared at the core/module level. AMD hasn’t disclosed the die size, but the transistor count is the same, at 1.2 billion. Given this, we’re assuming that the two chips are similarly sized, at 315 mm sq. All these similarities make it easy to compare TDPs and clock speeds between the new parts and the old:
One thing we can conclude off the bat – Vishera delivers some of the clock speed Bulldozer promised. Exact figures vary – the FX-8320 is 12-17 per cent faster than the FX-8120, while the FX-6300 and FX-6120 hit exactly the same clock speeds and TDP. Still, there are solid gains on the clock speed front virtually across the board.
Another thing that’s changed in 12 months is the CPU’s launch price. When AMD launched Bulldozer, over in the US it priced the chip at around $270 (£170). Given its lacklustre performance versus the X6 1100T, which was two-thirds of that price, and the drubbing it took from Intel CPUs in the same price bracket, the FX-8150 was a poor, poor value.
This time around, AMD has priced the top-end FX-8350 at $195 in the States, £125 in the UK. Whacking nearly 30 per cent off the launch price changes the points of comparison. This time, AMD is claiming that the FX-8350 is a strong match for Intel’s Core i5-3550 CPU – and that’s the Ivy Bridge we chose for our own tests. The Core i5-3550 is a quad-core CPU at 3.3GHz with a 3.7GHz Turbo Mode. Hyper-Threading is not enabled on this processor.
Both AMD CPUs were tested on an Asus Crosshair V using the latest BIOS, 8GB of DDR3-1600, and a Radeon 7950 GPU. Windows 7 SP1 (64-bit) was installed, along with all additional MS updates and the two scheduler patches Microsoft released for Bulldozer last year.
The first thing we wanted to compare was whether or not Piledriver would scale more effectively than Bulldozer in various threading scenarios. We measured the performance impact of running quad-threaded tests across all four modules and eight cores (4M/8C), four modules with one core active per module (4M/4C) and two modules, with both cores active (2M/4C). We’ll start with Cinebench, a benchmark AMD points to as evidence of Piledriver’s superior performance.
Piledriver’s single-threaded performance in CB11.5 is actually a bit better than Bulldozer’s clock-for-clock, as we’ll see, but the chip still takes a significant multi-threading penalty when we run all four threads across just two modules. Microsoft’s scheduler improvements don’t do much to help here.
Maxwell Render, on the other hand, does point to some improvements between the two processors. Bulldozer’s quad-thread/octal-core render time is 8 per cent worse than the ideal 4M/4C configuration, whereas Piledriver drops just 3 per cent performance there. Both chips take a 20 per cent performance hit when run in a two-module/quad-core configuration.
Next up, there’s DIEP. DIEP is a chess simulator that calculates the potential position of every piece on the board through a sequence of moves. A ply depth of one means the program has calculated every potential move a single turn into the game; a ply depth of 15 means every potential move 15 turns deep. The program spins off a pre-defined number of independent threads and uses no floating point code, which makes it useful for examining Bulldozer’s and Piledriver’s integer performance in different configurations.
Again, we see a much smaller drop off between 4M/8C and 4M/4C configurations. Good news for Piledriver on this front – now we turn to comparative performance.
Whether this graph is good news or bad depends on how much you care about single-thread performance. The good news is that the FX-8350 finally pulls unilaterally ahead of its closest Intel competitors as well as last year’s 2600K. The bad news is that a 1.06 single-threaded score at a clock speed of 4GHz is pretty terrible compared to the Core i5-3550′s 1.55x score at 3.3GHz. The X6 1100T (not shown here) turned in an even 1.0x last year, at 3.3GHz.
Despite multi-threaded improvements, Vishera still lags the old K10 core in single-thread performance, at least in some cases.
DIEP’s performance is strictly in line with the clock speed improvement between Bulldozer and Vishera. Unfortunately.
Maxwell Render is another win for AMD, and shows a bit more than a linear performance increase. The FX-8350 is only clocked 10 per cent faster than the FX-8150, but it finishes the render in 83 per cent of the time. Again, it looks like AMD’s latest chip is well matched against the Core i5-3550.
Finally, there’s PC Mark 7. AMD’s latest SATA driver significantly improved performance here. The FX-8350 doesn’t quite catch either of the Intel solutions, but it closes the gap.
Power consumption was one of our main criteria that Piledriver needed to improve relative to Bulldozer – so how do things look on that front?
Power consumption is a glass half-empty (or full) sort of moment. The good news? The FX-8350 is running at 4GHz; the FX-8150 at 3.6GHz. Given that both are drawing 233 Watts, AMD has improved the architecture’s power consumption. But, that said, the Core i5-3550 hammers the FX processors here.
The good news on Piledriver is that the eight-core Vishera CPU is an unambiguous improvement on ‘Dozer. L2 cache latency is back within range of Intel’s CPUs rather than literally half as fast. Multi-core scaling has improved. Power consumption is down, which allowed clock speeds to rise. This, in turn, allows Vishera to cleanly pull away from Thuban. At £125, the FX-8350 is a much better deal than Bulldozer ever was. I’ll go so far as to say that if AMD could’ve launched the FX-8350 as the first iteration of a Bulldozer-type architecture 12 months ago, it might’ve changed analyst opinions on the company’s long-term prospects.
Under different circumstances, I’d be willing to end the review right here, and close it on an up note. Piledriver is a good step forward, finis, the end. But given the state of AMD – the stock closed today at $2.10, having shed 25 per cent of its value since last week’s earnings call – it’s impossible to avoid the very real question of whether AMD has a future to look forward to.
Vishera is a good step forward given what AMD has had to work with. If Rory Read wants to keep his company afloat, he’ll start talking – in real, concrete terms – about what comes next. In October of 2011, we knew Piledriver was set to debut in the spring of 2012. In October 2012, Kaveri Steamroller – the chip that’s supposed to be Piledriver’s successor – is completely MIA. AMD gave a presentation on its architecture in August, but has been completely silent on when to expect the next-generation chip. There’s no word on whether or not the next-gen core will support Socket AM3+, whether it’ll offer PCI-Express 3.0, or if Steamroller will debut in both CPU and APU form, or be strictly an APU.
AMD has to answer these questions if it wants Vishera to generate momentum as opposed to postponing the inevitable.
Published under license from Ziff Davis, Inc., New York, All rights reserved.
Copyright © 2012-2013 Ziff Davis, Inc