One sickly sheep infects the flock - Isaac Watts
|
|
|
Intel is confident of its Sandy Bridge integrated GPU
CHIPZILLA'S Shmuel "Mooly" Eden of Intel's Israel team is one of the most famous recent microprocessor architects, having led the original Centrino architecture push and evolved it into efficient, fast "Core" CPUs that not only replaced the troubled Netburst, but also abruptly ended AMD's performance dominance in 2007. His team created the Sandy Bridge generation, expected to launch in mobile, mainstream and high end markets early next year.
A lot is already known about the initial mainstream desktop and mobile Sandy Bridge part, including its four CPU cores together with one GPU core, all sharing 8MB of obviously shared L3 cache, which, just like other on-chip caches, is expected to have lower latency than the current Nehalem or Westmere parts.
Besides the architectural core execution efficiency improvements, there are also AVX 256-bit vector floating-point instruction extensions with the new 3-operand format, like on the good old RISC machines, and matching doubled vector floating-point units for substantial performance improvement once the code is recompiled. The official DDR3-1600 dual channel support and improved overall interconnects and power management round up the story.
The speed of that integrated GPU in Sandy Bridge was the target of a lot of speculation. Intel's supposed confidential slides circulating on the web mentioned around twice the speed of the current Core i5 6XX series CPU, which would be decent, but not exactly easy to be competitive against a low-end Radeon HD 55XX series derivative expected to be integrated into the top end AMD Fusion parts in the same timeframe. And AMD people I spoke to on condition of anonymity were confident of having better GPU performance in those parts than what Sandy Bridge would have, resulting supposedly in "better overall user experience" despite likely slower CPU cores.
When confronted with that opinion, Mooly told us that he is confident of the Intel part's top performance, processing technology advantage and much better CPU to GPU integration. The latter two do make sense since the tuned 32nm CMOS process is now mature in Intel's fabs. Also, as far as we know Sandy Bridge CPU and GPU cores communicate tightly at the shared L3 cache level, while AMD's Fusion components supposedly talk one level below, at the integrated memory controller and crossbar level.
On the other hand, to really beat AMD's HD55XX series GPU cores in its Fusion parts, and in the absence of Larrabee, Intel will have to work hard on that GPU core anyway, and I'm sure it knows that. Well, we will know too, in about six months, how far ahead both Intel and AMD have gone. µ
Share this:
Share
Everyone go and search for "Intel 8087" on Wikipedia. Sounds familiar? We are back where we were three decades ago only on a much larger scale. The future is definitely "fusion". Does Intel's Sandy Bridge integrated graphics support DirectX 11 in order to be compatible with DirectCompute (Microsoft's API) or OpenCL (everyone else's API) for GPGPU applications? If not, will Intel's AVX performance be enough to match DirectCompute/OpenCL implementation? How about in terms of robustness in implementation and reachness in features from software development perspective? By the time AMD Fusion APUs arrive, AMD would have full OpenCL software framework support already available to developers for over a year. It can be done in two ways too: 1. using Fusion APU, or 2. using AMD CPU and Radeon HD 5000 series GPU in PCIe slot. AMD already has GPU market share. The future of x86 is homogeneous GPGPU (somewhere along the lines of Intel's own Larrabee) and what AMD is preparing with Fusion 2.0 by 2015. Both Intel and AMD know it but the software support is what makes it or brakes it. Software support is crucial and a lot still has to be learned about GPGPU computing before it can become homogeneous. This is why Larrabee got scrapped - it was ahead of its time.
...one underestimates the impact of software on a successful product. Intel has proven time and time again it does not know how to do software. There is no reason to expect that their drivers for this GPU will be any better than past unsuccessful efforts from Haifa, Chandler, Hillsboro or Folsom.
Ok, so we're talking about integrated GPUs here, and people are posting about how long a socket lasts. Here's the thing, there are two types of users involved here:
1) People using an integrated GPU. People upgrade these machines sometimes, but they aren't know as the big upgrade crowd. Keep in mind, this article is about integrated GPUs.
2) People using discrete GPSs. First, why are they commenting on this article? Second, just because they will support future CPUs doesn't mean they will support future RAM specifications.
If u want a socket system that will last go with AMD.
Their boards have a really long life. For example they released boards with support for the Phenom 2 about 6 months before any CPU's came out. These boards were widely available for only 75$. The fact that they weren't needed yet avoided the usual price hike and shortages, and when the CPU's finally came out, there was a large enough supply of MOBO's available that the resulting price hike was minimal.
Add on the fact the boards support chips way back (original Athlon x2's or earlier), and far into the future (Like the new 2.8 GHz 6 core for 200$) and these boards have a life span of 5 years easy.
Intel doesn't even need Sandy Bridge to compete with AMD. As soon as Otellini resumes his under-the-table bribes and threats to HP and other computer manufacturers, AMD will be kicked to the sidelines of the market again.
Cheating pays. Intel is the master.
Since Intel are pulling the douche one-socket-per-generation move, I'm going to wait until Sandy Bridge and Fusion hit the market to upgrade. Hopefully I can look forward to a socket that won't be obsolete the minute I buy it.