AMD Starts Shipping First Bulldozer CPU

MrSeb writes "After an awfully long wait, AMD has finally begun shipment of its Bulldozer-based Interlagos (Opteron 6200) server-oriented CPU. If you believe AMD's PR bots, it is the world's first 16-core x86 processor. Unfortunately, and possibly because of reports that AMD is struggling to clock its Bulldozer cores to speeds that are competitive with Intel's Core i7, there's no word of the 8-core desktop-targeted Zambezi CPU. If AMD doesn't move quickly, Intel's Sandy Bridge-E will beat Zambezi to market and AMD will lose any edge that it might have."

Comments

Sandy Bridge-E

[Anonymous Coward]

If AMD doesn't watch out their mainline $200 processor will be made obsolete by Intel's $1000 EXTREME CPUs!

Re:Sandy Bridge-E

[CajunArson]

You may laugh but think of it this way... if that $1000 Gulftown CPU from March of 2010 can still beat an 8 core Bulldozer that comes out 19 months later, then you would only have to realize a marginal benefit of about $1.75 per day to make it economically worth your while to have bought the "overpriced" Gulftown chip. (that includes the cost from Intel motherboards that tend to be more expensive and the extra RAM for a triple-channel configuration). Nevermind the fact that 6-core chips have been sold for $600 for some time as well. I can think of a bunch of professional applications that can easily show a $1.75 / day benefit from the extra cores. Maybe not for playing games, but for a lot of real applications.

  Bulldozer should beat the consumer-level SB chips at perfectly threaded integer benchmarks, but it remains a very open question if it will be able to beat the almost 2 year old Gulftowns at the same tasks, and it is an almost foregone conclusion that it won't beat the 6 core SB-E chips at those tasks. Factor into account the 315 mm^2 die size of EVERY Bulldozer (not just the 8 core ones, but the cheap 4 core ones too since AMD only has 1 die design) and the immaturity of AMD's 32nm process and things could be expensive for AMD on the desktop. That's why it makes sense to ship the server chips first where AMD has some hope of getting higher ASPs.

Re:Sandy Bridge-E

[sgt scrub]

the immaturity of AMD's 32nm process and things could be expensive for AMD on the desktop.

That is true, and as you point out, for the desktop. Machines in a data center are cooled so the number of cores is a better measure of functionality. If you build machines that run multiple VM's, which is usually the case, that cheaper 6200 will not only outperform Intel's Gulftown and more likely be preferred when adding more machines to the data center even over the SB-E chips. If AMD can get a better footing in the "cloud" infrastructure they might make enough to move to a die size smaller than 32nm, which is REALLY what they must to do.

Re:Sandy Bridge-E

[bhcompy]

That $200 is still a fuzzy number given the motherboard prices. You can get a very good AMD chipset(880/890) for $100 and have all the latest features(USB3, SATA3, etc) while being forward compatible for quite a while(manufacturer dependent, but Bulldozer(AM3+) is compatible with AM3 chipsets with BIOS updates). With Intel, you're still paying more for the equivalent and next year you'll need to get another motherboard to upgrade that processor because of constant socket changes.

Re:Sandy Bridge-E

[V!NCENT]

Talk to a server farm when Intel is putting money in the development of Coreboot and we'll talk.

For now AMD is superior, because a server reboot requires about 1/100th of the time that it takes an Intel CPU farm to get back up due to horrible BIOSes. The more motherboards you have, the longer it takes due to serialized bootup. Ouch... Massive ouch...

Downtime versus marginable CPU speed... And less cores...

Re:Sandy Bridge-E

[Sloppy]

Is the 6200 16 physical cores or just 8 dualthreaded ones?

Part of what makes Bulldozer interesting, is that there is no quick straight-forward soundbite answer to that question. It's 16 less-than-complete cores, or 8 more-than-merely-dual-threaded cores.

I look forward to the day when some profiler nerd figures out that most cost-effective ratio is to make a chip with 3 memory busses, 17 integer units and 11 floating point units, with 23 sets of registers. /proc/cpuinfo will say it's a 23-core machine so that'll be the soundbite answer, and then some math twit will whine that it's really only an 11 core machine. And everyone will be both wrong and right simultaneously.

The price/performance ratio.

[Chrisq]

What will be interesting is the price/performance ratio compared to the Intel chips. This chip will be typically used in server farms, and this will be at least as important as the raw power - though obviously there is an overhead in running more servers. AMD has usually been ahead of Intel, and it still is on most mid-range and low-end chips, but it has started to fall behind at the high end.

What is the FPU performance of these things...?

[serviscope_minor]

Does anyone know the FPU performance of these things?

So comparing a 16 "core" 'dozer to a 12 core magny-cours:

The number of parallel integer (and memory addressing) threads has gone up from 12 to 16.

The number of FPUs has dropped from 12 to 8.

The new FPUs are now twice as wide with the AVX instructions.

So, two threads share one wider FPU now. If it's hard keeping an FPU full, then this should make better use of the hardware. It seems that if your code does well for parallel, scalar FPU work already, then there may be a performance drop.

If you have trouble filling the FPU for scalar work, then this should give better utilisation of less hardware. There's a possible performance increase if your utilisation is currently under 67%. Since the two core units can feed the FPU independently, there is a little latency hiding now. This could help even if there are two completely independent processes using the FPU at the same time.

I suppsose the reasoning is that there is often fine-grained parallelism to be had, and the problem of fine grained parallelism and keeping the FPU full are often independent. So AVX will improve performance there.

So, it seems that the peak FPU performance has increased in the ratio of 16/12.

The actual performance could be all over the place. It will be interesting to see.

The other thing is that these are now single chips with 8 bulldozer units on and 16ish cores. Perhaps AMD will go and make more MCMs like before, giving 32ish cores per socket :)

It is all about the die size

[sgt scrub]

Unfortunately, and possibly because of reports that AMD is struggling to clock its Bulldozer cores to speeds that are competitive with Intel's Core i7, there's no word of the 8-core desktop-targeted Zambezi CPU.

If you increase the clock on the CPU you have to cool it. Reducing the die reduces the amount of cooling that needs to be done. AMD is not able to shrink their die. Yet.