Slashdot Mirror
AMD Releases Image of Phenom/Barcelona Die
MojoKid writes "A few weeks ago, AMD
released information on new branding for their desktop derivatives of the Barcelona core, now dubbed the Phenom FX, X4 and X2. If you're unfamiliar with Phenom, the processors will be based on AMD's K10 architecture. They've been tight lipped about specifics, but we know that it will feature a faster on-die memory controller, support 64-bit and 128-bit SSE operations, and they'll be outfitted with 2MB of on-chip L2 cache (512KB dedicated per core) in addition to 2MB of shared L3 cache. This week, instead of revealing some more of the juicy details regarding those enhancements, AMD just sent over a tasty photo of a Phenom die. At least it's something."
On-chip connectivity can be much broader and lower-latency than off-chip connectivity. The two-dual-core in one package "quad cores" of Intel have to talk via the off-package north bridge. As you can see from the AMD Barcelona/K10/10h snapshot, the cores live together on a single piece of silicon.
According to Intel engineers though, communication between the chips was never a bottleneck, so the avantages of one vs the other design are questionable. I'm not a processor engineer, but that holds true everywhere: throwing resources in improving something that's not a bottleneck.. actually does NOT help performance. Logic 101.
And BTW:
AMD just sent over a tasty photo of a Phenom die.
All right! So I can print is and try it on a compatible motherboard, right? Right..
Indeed so. Anyone having bought or buying an AM2/AM+2 desktop motherboard can drop in Phenom processors. When you have a performance AMD 4x4 (1,207-pin Socket F) board with FX processors, you can drop in the new quad core FX chips as well. Similarly, when you have a DDR2 Opteron server/big-iron, you can also upgrade.
That makes the current AMD platforms attractive: you can buy a cheap Athlon X2 chip to get good performance now, and later upgrade to a Phenom chip and get excellent performance and four-way multiprocessing. I plan to wait with my upgrade until the price comes down a bit.
The single die - four cores processor architecture from AMD could be a result of their collaboration with Sun which already has 8 cores in a single die general purpose processor UltraSparc T1 for more than a year. It's surprising though that the two chip makers, Intel and AMD, still lag behind Sun in terms of cores per die.
I understand your cynicism. Especially considering how a lot of processor intensive applications that most consumers use (games and other multimedia, and to a small margin running outlook, internet explorer, and MSN messenger together) get absolutely no benefit from SMP. However, as multi-core chips are rapidly becoming the defacto must have for everyone, I think developers are going to start coding to take advantage of this. We've already heard plenty of rumors about offloading most/all of the physics processing onto a chip that does it's computations in matrices rather than in any sort of linear fashion, streamlining the process both in method of computation and by freeing up your cpu cycles any number of other tasks (potentially an increase in game artificial intelligence, so it behaves less predictably, maybe do away with all of the nested tree structures and boolean choices) The only potential problem is increasing the complexity of development. Applications to take full advantage of all the new widgets will also take exceedingly more development time, support time, QA time, ect which will (alm0st) inevitably lead to a rise for consumers.
--fetch daddy's blue fright wig, i must be handsome when i release my rage
Actually even if all cores are on the same silicon, inter-core communication is not as good as it could be. This artical has some interesting information on the topic:c ore-dtr-analysis_12.html
http://www.xbitlabs.com/articles/cpu/display/dual
Anybody who encodes audio and video will love quad core. Xvid 1.2+ supports parallel encoding, and sweet jesus does it do it quickly on the proper machine. I can encode a 90 minute film to 1400MB Xvid w/multi-pass encoding in about 45 minutes on a stock E6600. On my stock AMD64 3000+, the same operation takes about 150 minutes.