Slashdot Mirror
45nm Opteron Performance, Power Efficiency Tested
An anonymous reader writes "Now that Intel has unleashed its next-generation Core i7 processors, all eyes are turned to AMD and its incoming wave of 45nm CPUs. To get a feel for AMD's future competitiveness, The Tech Report has taken a pair of 2.7GHz 45nm Opterons (with 75W power envelopes) and put them through the paces against Intel Xeons and older, 65nm Opterons in an extensive suite of performance and power efficiency tests — from Cinema 4D and SPECjbb to computational fluid dynamics and a custom XML handling benchmark. The verdict: AMD's new 45nm quad-core design is a notable improvement over the 65nm iteration, and it proves to be a remarkably power-efficient competitor to Intel's Xeons. However, 45nm AMD chips likely don't have what it takes to best Intel's Core i7 and future Nehalem-based Xeons."
No marketing talk in those names.
Not sure you would call it expensive, but the OPTERON chips per definition are only server chips. The Opteron 23xx series (45nm shanghai) is dual processor, while the 83xx series is quad processor.
The end user equivalent is the PHENOM series.
Note that this is a technical difference, not marketing talk. The Opterons use Socket F, while the Phenoms (single processor only) use the AM2+ socket. Different pin count, different number of interconnect ports (for connecting to other processory).
45nm Phenoms are IIRC supposed to appear soonish ;) Opterons start being available now - I pick up a new server on friday.
You should read up on the difference between dual CORE and dual PROCESSOR.
In fact, both current Opterons as well as current Phenoms are quad core systems. The Phenoms single processor (1x4 cores), the Opterons dual to quad processor (2x4 cores to 4x4 cores).
Hardly any end user system (i.e. non-server) today uses more than one processor. Dual core to quad core is normal now. But always on one processor.
The difference in sockets actually is for that - the AM2(+) socket lackss the HT bus for inter-processor communication, while the Socket F has separate lanes for the processors to talk to each other.
The main problem with Intel right now is that intel has no really nice solution at all in the multi processor side - they simply (again) do not scale from the memory side, thanks to a lack of a NUMA architecture (that they change now and coming).
I cant resist, on account of all the people who look at the Core i7 benches and think its all over; considering that the best "review" of a Nehalem EP (the dual socket variant) is a couple of guys who have a single screenshot for spec_fp, I'd say the battle's too early to call. All we've seen are single socket Nehalems-- & thats not been AMD's strong suit for some time.
Even considering that Intel's single socket game has been largely better for a while, there are some key areas AMD systems perform better. HPC, render farms, some web serving, virtualization... for all these places where people need a lot of cpus, AMD is has stayed in the runnings or maintained a lead (depends a lot on just what you're running). Unfortunately the benchmarks usually published dont factor in these kinds of workloads much at all. Cinebench is the only benchmark in the review anywhere near the above. I think if we ran some VMWare benchmarks, things would look drastically different.
But the real quesiton here is Intel: Intel is just now doing the infrastructure AMD did in early spring `03: QPI to AMD's HT, similar onboard memory setups... and thusfar aside from some spec_fp numbers, we have no idea whatsoever how well their implementation is going to work. Once Intel releases Nehalem EP for testing, we'll have an idea.
Seven out of the top ten supercomputers in the latest top500 list have AMD in them, including the top two, so I don't really see the whole "AMD losing momentum and competitiveness.
Seven out of the top ten supercomputers have Power Architecture processors in then too, including the top two, but I'd say that Power Architecture has lost its momentum, wouldn't you?
PS. For those who don't know. Roadrunner uses PowerXCell 8i processors, which are Power Architecture. All Cray XT3/4/5 supercomputers uses PowerPC 440 based communication processors called SeaStar. BlueGene uses PPC 440/450 based custom CPUs. DS.
- Henrik
- when the Shadows descend -
Well to be pedantic:
I wouldn't be so quick to say that AMD has lost it all.
First of all, in the 4 and 8-socket market, AMD still has no competition. The Intel Xeon MP series is still using the outdated FSB technology. This series also requires expensive and power-consuming FB-DIMM modules instead of DDR2/DDR3. Nehalem-based Xeon MPs are not going to ship before Q4 2009. Etc.
Secondly, in the 1-socket and 2-socket market, and regarding the latest 45nm AMD Shanghai and Intel Nehalem, so far there are very few benchmarks comparing directly the 2 microarchitectures; most of the hardware review sites do the mistake of comparing Shanghai against the older Intel generation, or the older AMD generation against Nehalem. But from what I have seen, clock-for-clock, for most workloads, Shanghai and Nehalem are very close, +/-10% in terms of performance, and Shanghai seems to do this in the same or a slightly lower power envelope. Some workloads do exhibit a more significant performance difference, with either Shanghai or Nehalem pulling ahead of its competitor. Now comparing clock-for-clock isn't really what matters. What matters is dollar-for-dollar comparisons. But what is interesting is that AMD has priced the Shanghai Opterons 23xx to match very closing the Nehalem Xeon 55xx series at equivalent frequencies. This tends to indicate that AMD thinks that they offer a clock-for-clock value identical or better than Intel.
The only area where AMD will clearly be unable to compete in the 1 and 2-socket market is the very high end: 1-socket Shanghai processors will top out at 3.0 GHz, 2-socket processors will top out at 2.8 GHz, while Intel goes all the way up to 3.2 GHz. However these expensive processors represent a very small proportion of the market share (virtually nobody buys $1000+ processors), so it shouldn't be a huge factor regarding which processor manufacturer "wins" this 45nm battle. Intel will have the bragging rights, but that's about it.
Another last point I would like to mention is that AMD will be the only one to offer low-power 1-socket 45nm Shanghai for at least the entire first half of 2009: 55W and 75W ACP Opteron 13xx, and 95W TDP Phenom II. While Intel will only offer Core i7 and Xeon 35xx processors rated at 130W TDP (!). They are planning to release lower-power 45nm Nehalems only during the second half of 2009. I find it rather stunning for Intel to not care more about power consumption... especially for their Xeon 55xx line, the server market cares about energy efficiency. We all remember that extravagant power consumption and temperature was a major factor that caused the failure of the Pentium 4 Netburst microarchitecture...
i've often wondered if microsoft has a deal with intel to slow amd processors.
Yes, sort of.
It's called NUMA - Non Uniform Memory Architecture.
Up until recently Intel platforms had the memory controlled by the northbridge, with all CPUs and all cores having the same access to the memory.
Newest Intel platform and all 64bits AMD had the memory controller on the processor package. In a multi-socket configuration, each processor controls it's own chunk of memory, so for some range, the access will be faster because the processor is directly accessing it, and for other the latency will be increased because the processor has to ask its neighbour over HyperTransport / QuickPath.
To be able to function in a such configuration, an OS should pay some attention when scheduling process and threads to cores : it should be best that all threads from some process are all scheduled to cores having all direct access to the resources used by said process. (While at the same time scheduling two threads at a physical core and it's corresponding hyperthreading virtual core if there's a another physical core sitting idle)
Windows has always deeply sucked at this. Opensource OS, on the other hand, have much more work applied to them for that. (That's why they are much more popular on super computers).
This also introduces technical difficulties (like keeping the cache coherent). That's also why heavily multi-socketed (4 and up) motherboard won't be coming during the first year of Core i7's life. They probably have to fix all the fine details before that. As usual expect a change in socket format and a new iteration of Core i7 not quite exactly compatible with the previous one.
On AMD's side, currently sold Opteron are already adapted for 4 and more sockets configuration. (As explained by other /.ers, the 8000 series has a coherency protocol running on 3 HT interconnects, which should be enough to help on 4 and more sockets).
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]