Processor Whispers: About cold chips and hot blogs

by Andreas Stiller

When Intel celebrates the IDF in Silicon Valley, traditionally, AMD isn't far away, presenting its newest products. This year, it was AMD's Bulldozer in particular that was shown to the multitude of visiting journalists and analysts.

With a lot of liquid helium, AMD managed to grab some attention on the opening day of the Intel Developer Forum. The helium was used to boost an FX prototype (with three disabled modules, 2 volts core voltage) to the new world record clock speed of 8.429 GHz – at a cool -230 degrees Celsius (-382 degrees Fahrenheit). With the help of experienced overclockers AMD passed the record set by an old Celeron-D, in 65 nm technology with P4 architecture, by a considerable 1.4 per cent and was allowed to register the new record in the Guinness Book of Records.

AMD's Mobile Bulldozer Trinity (on the right) challenges the Core i5 under Deus Ex. Just around the corner from where the IDF conference was held, some room-temperature Bulldozers could be admired, among them the Llano successor with the codename Trinity. This chip, expected to roll out at the beginning of next year, will feature Piledriver, the slightly improved Bulldozer core, and offers around 50 per cent faster graphics. In a demonstration of Deus Ex, the team of four Piledriver cores was able to produce a much smoother video output on its notebook's display than the Core i5-equipped notebook. However, the Piledriver notebook featured DirectX 11 with tessellation while its competitor had to make do with DirectX 9 without tessellation.

Next to this pair, an 8-core FX was pitted against a 4-core Intel Core i5 "in the same price range" and outpaced it by 19 per cent in the open-source Handbrake H.264 benchmark. For another demonstration, DiRT 3 running at a resolution of 2560 x 1600, AMD was a bit more generous with details: the competitors were an FX-8150 on an AMD-990FX board with DDR-1600 and an Intel Core i7 980 on an Asus X58 board with DDR3-1066. That both were roughly equal in performance, the first with 81 and the latter with 83 fps, probably had more to do with the two AMD Radeon 6970 cards with crossfire in each system. In any case, AMD pointed out, the components – processor, memory and board – only cost $515 for the FX-8150 system in comparison to a list price of $1279 for the Intel system.

But AMD doesn't just struggle against Intel. On the other front, there is NVIDIA. And so, an Acer Tab with AMD-Z-01-APU (Bobcat, dual core, Windows 7) was pitted against two "competitive" tablets in Futuremark's Peacekeeper browser benchmark: one equipped with an Intel Atom Z670, the other one with an NVIDIA Tegra 2, running Android 3.2. The results were 1367 points for Bobcat, 1089 for Tegra 2 and, far behind, the Atom with only 652 points. Intel had, however, published other benchmark results, like Linpack or Sunspider, at the IDF that placed the Atom Z670 50 to 80 per cent ahead of the Tegra 2.

Anti-Aliasing

Being debated more hotly than AMD's cherry-picked numbers are the new results from well-known benchmarks like Cinebench that are circulating on the Internet and which make the Bulldozer look much less top-notch. The Czech blogger Obrovsky has published many such results for the AMD FX-8120 at 3.1 and 4 GHz and the FX-8150 at 3.6 and 4.2 GHz on his site, thereby attracting the collective wrath of the AMD fan community.

Intel has equipped the Ivy Bridge with three DisplayPorts, but AMD has now shown five displays in a row with Eyefinity. For a direct comparison, he used the FX-8150 at 4.2 GHz and an equally clocked Phenom II X6 1100T, both with their turbo mode disabled. The Cinebench 10 and 11.5 results, in SuperPi and Fritzchess and with the same number of threads, are disillusioning but not really surprising considering that the old Phenom architecture has one universal pipeline more per integer core, larger caches, its own frontend and its own FPU. With an equal number of threads, Obrovsky's results for the Bulldozer prototype mostly are about 25 per cent below the Phenom's ratings. And even at full core performance the 6-core Phenom II X6 still beats the 8-core FX with 7.38 to 6.95 in Cinebench 11.5.

Perhaps, Obrovsky still has an old prototype. According to him, the stepping is B2, but the Bulldozer's steppings range from B2a up to the current B2g. Also, some performance reducing bugs are known that can possibly be fixed or reduced by a processor driver before B3 or C0 take care of them for good. For Linux, a kernel patch (not in the distributions yet) that avoids the frequent, mutual kicking out of cache lines (thrashing) caused by aliasing problems in the L1 instruction cache was released at the end of July. The patch is being discussed under "[PATCH]x86.AMD:Correct F15h IC aliasing issue" on gmane.org, by Linus Torvalds among others. Some SPEC-CPU2006 benchmarks are supposed to be especially affected by this aliasing problem.

Be that as it may, in his blog, Obrovsky also published results from tests with nominal clock speeds and enabled turbos, the way the processors would normally run. And here it looks much better for the FX-8150 at 3.6 GHz in comparison to the Phenom II X6 1100T at 3.3 GHz. Thanks to the turbo, the FX still gets 6.93 points in Cinebench 11.5 while the Phenom II X2 falls to 5.92. Also in most other benchmarks, the FX leaves the old Phenom behind and mostly contending with the Core i7 2600K and, at times, with the Core i7 990x. With TrueCrypt AES, which apparently supports the new AES instructions, the bandwidth is almost five times as high as with the Phenom. But that's a lot of confusion, so it's high time that we should get the chance to find out for ourselves.

(djwm)