Business, Hardware

AMD Delays Bulldozer Citing Performance Issues

Late last week, AMD officials got back to us and cited that remarks made by Rick Bergman were true and that the Bulldozer is not delayed. Day zero of Computex Taipei painted a different story…first we got bits’n’pieces from AnandTech, who went live with information that AMD delayed Bulldozer i.e. AMD’s FX processors:

"AMD originally wanted to launch Bulldozer at Computex but performance issues with its B0 and B1 stepping chips pushed back the launch. Now we’re looking at a late July launch with B2 silicon, but performance today is a big unknown. Apparently the performance of B1 stepping silicon doesn’t look too good."

At around the same time, XbitLabs came out with the story about AMD delaying Bulldozer for clocking issues:

"The currently available B0 and B1 stepping Zambezi/Bulldozer processors can function at around 2.50GHz/3.50GHz (nominal/turbo) clock-speeds and at such frequency they cannot deliver performance AMD considers competitive, a person with knowledge of the situation said on Monday. As a consequence, AMD needs to tune the design of the processor and create B2 stepping of the chip with better clock-speed potential amid similar thermal design power (TDP), which will take several months to complete. Therefore, the Sunnyvale, California-based chip designer will release its highly-anticipated Bulldozer processors for desktops in September, not in June, as planned."

According to our sources the performance was non-competitive, meaning all those rumored scores with Bulldozer pummeling Core i7 were incorrect. We later learned that while some of those scores were the result of pure fiction and fanboyism, some scores were leaked from a performance document which estimated the clock-per-clock performance with Sandy Bridge generation of processors.

Leaked Table showing clocks AMD was targeting with "Zambezi" B1 silicon

The realities are that AMD planned a launch a product running at 3.8GHz with Turbo mode up to 4.2GHz (FX-8130), with the limited, OEM/SI-targeted unlocked edition (FX-8130P) running at 3.8GHz with Turbo up to 4.2GHz. Prices were incredibly aggressive and there was no doubt in anyone’s mind that those processors would make an impact on Intel’s position. Unfortunately, the company hit a speed bump and was not able to achieve that goal with B0 and B1 silicon (staying over 1GHz short of target). Remember, these are targets that need to be achieved in worst possible thermal conditions (OEM Qualification i.e. "baking test" with air set to 70C etc.).

Last year, we had the opportunity to use a system running the very first silicon that AMD allowed outside of the pearly gates at Sunnyvale/Austin and the performance was abysmal. This was not only due to the clock of the CPU cores, set at 1.8GHz. The memory bandwidth with dual-channel DDR3-1333 was only running at about 50-60% of theoretical maximum, on the levels of the original Phenom (K10, "Barcelona" architecture). Read, we’re talking about scores being around 9-12GB/s out of theoretical maximum of 21.33GB/s. Following our story, BSN* was not invited to the launch of Radeon HD 6000 Series citing that our publication was "biased". Note that we did not disclose exact performance scores. Interestingly enough, the blue company did not accuse us of biased attitude even when our contributing editor was subpoenaed by the FTC in their now settled anti-trust case.

Getting back to Bulldozer B1 silicon scores, in one of our discussions with former AMD executive, we were told that B0 silicon was more akin to "Man with a Shovel" rather than being a "Bulldozer". Interestingly though, another highly-ranked source close to heart of the matter told us that the partial blame lies on manufacturing side of the matter, with some of key engineers left the 32nm processing team and focused on 28nm and 20nm process nodes, rather than a 32nm one – one used by Bulldozer. This was a consequence of AMD spinning off GlobalFoundries which were under new management that had only one goal: to capture as much customers as possible. Capturing customers meant focus on more manufacturing processes than before and the consequences are visible today.

Regardless of hearsay that we are all unfortunately, going to experience over the next couple of months, the result is the same: AMD is now forced to work on B2 silicon which may or may not resolve the internal bandwidth issues that plague otherwise excellent concept. The interesting bit to us is that while K8/10 generation of Opterons lagged behind Intel in integer and pummeled them in Floating Point tests, the result right now is exactly the opposite: clock-by-clock, Bulldozer will walk all over Sandy Bridge in integer and get hammered in Floating Point tests. The reason for this is the fact that second generation Bulldozer (so called Bulldozer Enhanced) will unite the Bulldozer cores with Radeon HD graphics which will take over the floating point operations as much as possible through OpenCL and other GPU Computing APIs.

And of course, we all know what it means to compare a CPU FP performance to a GPU one: if a CPU FP would be a horse carriage (quad-horse, sexa-horse, you get it) at full speed, GPU FP would end up being a Red Bull Racing Formula One car.

We hope that we will see Bulldozer processors for Back to School period, so that Taiwanese motherboard vendors, who all heavily invested in Bulldozer ecosystem with serious high-end motherboards – can finally see the top performance of their Bulldozer-tweaked designs.