AMD Sued Over Allegedly Misleading Bulldozer Core Count

An anonymous reader writes: A class action suit accuses AMD of misleading buyers about the number of cores in its Bulldozer-based CPUs. The complaint claims that the chips effectively had only four cores, while AMD claims there are eight. According to Ars: "AMD's multi-core Bulldozer chips use a unique design that combines the functions of what would normally be two discrete cores into a single package, which the company calls a module. Each module is identified as two separate cores in Windows, but the cores share a single floating point unit and instruction and execution resources. This is different from Intel's cores, which feature independent FPUs. The suit claims that Bulldozer's design means its cores cannot work independently, and as a result, cannot perform eight instructions simultaneously and independently. This, the claim continues, results in performance degradation, and average consumers in the market for a CPU lack the technical expertise to understand the design of AMD's processors and trust the company to give accurate specifications regarding its CPUs."

Comments

Re:Pretty Laughable

[Ramze]

I agree. Also, if I understand correctly, that shared FPU can be used either by one CPU core as a 256 bit FPU, or by both simultaneously as 2 independent 128 bit FPUs.

So, shared, but not likely to be a bottleneck.

Also, since when do cores have to include all the "extras" ? I recall when 486's had a math co-processor and there were no mmx instructions or other such multimedia or physics sets. This guy is going to have a really tough time explaining how exactly AMD's architecture doesn't provide exactly the number of cores listed -- even if the architecture has its limitations due to sharing resources.

Re:So AMD called their Hyperthreading a CPU core?

[gbnewby]

Concur. The design of the bulldozer modules was abundantly clear. The fact that the "cores" share an FPU was clearly disclosed and part of any diagram of the parts. The shared FPUs played a huge role in assessing bulldozer-based CPUs for high performance computing workloads. For HPC, the usual benchmark is HPL (a.k.a., high performance LINPACK), which is a measure of double precision floating point performance for a particular matrix operation called DGEMM. The fact that the FPU was doing double-duty for two "cores" on a module meant that the peak theoretical performance was limited by the number of FPUs in a CPU, not the number of cores or modules or anything else.

As others have noted, hyperthreading via Intel can have exactly the same impact: the threads share various components, including the FPU.

Another aspect that can have a major impact on performance is the number of memory channels, and how things like cache coherency is handled. Among other things, AMD's hyper-transport exhibits different scalability characteristics depending on the number of sockets. In a four or eight-socket configuration, latency due to cache coherency operations can have a big impact on performance.
    - gbn

Re:i5, same thing?

[fyngyrz]

Mod parent up. It is correct. Some i5's do have 2 cores. The original AC post near the top is also correct, which I notice the mods have knocked to -1 in the usual frenzy of "I disagree, although I am uninformed about the matter."

For a couple decades, "core" meant CPU. 4004, 1802, 8080, 6800, 6809, Z80, 68000... etc.

Then it meant CPU+MMU.

What's funny is that "core" now seems to mean CPU+MMU+FPU (which is great... I love me an FPU per core.) But some (often aimed at mobile or tablets or phones) also have GPUs.

So how long before we won't accept "core" except as a label for CPU+MMU+FPU+GPU?

Then it might be "Nah, that thing has no NPU (neural processing unit), that's not a "real core": CPU+MMU+FPU+GPU+NPU

Then perhaps it'll be "Nah, that thing has no QPU (quantum processing unit)", that's not a "real core": CPU+MMU+ FPU+GPU+NPU+QPU

Then... well, I have no idea. But I do think it'll be something. It's always something. :)