When is AMD going to release real 8 core chips?

[Onceler]

The current CPUs are really just quad core as they have only 1 FPU to each pair of "cores"

 

[DeathReborn]

They were still CPU cores before FPU's were added in, so the current CPU's are historically accurate as having 8 cores.

 

[ShintaiDK]

They were still CPU cores before FPU's were added in, so the current CPU's are historically accurate as having 8 cores.

No....Its also far more than the FPU thats shared.

If you think desktop wise, 8 cores doesnt make much sense. So you could say..not anytime soon.

 

[Olikan]

They were still CPU cores before FPU's were added in, so the current CPU's are historically accurate as having 8 cores.

yeah...if bulldozer's performance was like an 8 core, this thread would not exist

 

[eternalone]

Never

 

[BenchPress]

The current CPUs are really just quad core as they have only 1 FPU to each pair of "cores"

Actually they have two 128-bit SIMD floating-point units per module (two integer cores). But yeah, it's a whole lot less than Intel who has two 256-bit units per core. AMD does have FMA support already, which doubles the theoretical throughput, but in practice it's closer to 50%.

AMD will get in deep trouble when Intel launches Haswell, which will have AVX2 support with two 256-bit units each featuring FMA.

AMD is betting the farm on its HSA technology, i.e. using the GPU for all the high throughput computing. But this is not being welcomed by the majority of developers. It takes a big effort (read: money) to support HSA and there's no guarantee of getting good results due to the heterogenous computing overhead. AVX2 on the other hand is trivial to support because compilers can easily auto-vectorize any code loop with independent iterations.

Manju Hegde claimed they still invest in both the CPU and GPU but I guess we'll have to wait and see how long it will take them to widen the SIMD units and support AVX2.

 

[nehalem256]

http://en.wikipedia.org/wiki/Opteron#Opteron_.2845_nm_SOI.29

8-core – Magny-Cours MCM (6124-6140)
Released March 29, 2010

About 2 year ago.

 

[Charles Kozierok]

Bulldozer "modules" are tricky. They don't have as much hardware as two real cores, but they have more than a single core with HT. Whether they are closer to two full cores or a single core depends entirely on what you're doing with them.

But the real problem is that they are still too slow.

 

[Anarchist420]

It doesn't matter when they do, because they'll just have terrible cache design. Intel will probably be the first to release a true 8 core CPU although there will be no 8 core Haswell if I'm not mistaken. I kind of think it's silly for intel to put so much into the iGPU rather than just use double the number of cores and possibly with some texture units while just emulating the color and depth units. With DDR4, they'll have enough bandwidth. Then the number of cores to dedicate to graphics could be configured by driver.

 

[Arachnotronic]

Xeon E5 2687W is your chip, then!

 

[ShintaiDK]

It doesn't matter when they do, because they'll just have terrible cache design. Intel will probably be the first to release a true 8 core CPU although there will be no 8 core Haswell if I'm not mistaken. I kind of think it's silly for intel to put so much into the iGPU rather than just use double the number of cores and possibly with some texture units while just emulating the color and depth units. With DDR4, they'll have enough bandwidth. Then the number of cores to dedicate to graphics could be configured by driver.

There is a 10core/20threads "true cores" Xeon EX.

But if we look serverside too, then MCM designs are 8cores too. For desktop there is none however.

 

[Phynaz]

So, number of FPU's is what determines how many cores a CPU has?

 

[Edrick]

Intel will probably be the first to release a true 8 core CPU although there will be no 8 core Haswell if I'm not mistaken.

8 and 10 core CPUs already exist from Intel. And Haswell will have an 8 core versions.

With DDR4, they'll have enough bandwidth

Huh? Source?

 

[Chiropteran]

So my old 486SX was actually a "zero core" since it didn't have an FPU?

 

[Edrick]

So my old 486SX was actually a "zero core" since it didn't have an FPU?

By todays standards?

 

[BenchPress]

So, number of FPU's is what determines how many cores a CPU has?

No, there is no strict definition. Using the same logic as GPU manufacturers (counting each SIMD lane separately), Haswell will be a 64-core chip...

To make matter worse there's the same confusion about threads. Hyper-Threading gives us twice the number of logical cores (from an application's point of view there's no difference between a quad-core with HT and an 8-core). But GPUs claim to run thousands of threads while not so long ago they were really only capable of running one shader at a time. Their definition of threads is better described as hardware managed fibers, and strands running in lock-step.

Anyway, I digress. AMD defines a core as an independent scalar integer execution cluster, which has its own L1 cache and register file. They share an SIMD cluster and decoders though. For scalar integer code this can closely approach the performance of a true 8-core without Hyper-Threading (for a quad-module). The problem is in terms of SIMD performance (both floating-point and integer) it behaves as a quad-core with Hyper-Threading. They'll be able to keep up with a quad-core Haswell chip if they double the SIMD width.

 

[ShintaiDK]

As I said before, Bulldozer doesnt just share the FPU. But also the entire frontend with decoders and so on within its modules. Not to mention the L1 instruction cache.

Claiming Bulldozer is an 8 core is like saying 3770 is an 8 core too.

 

[formulav8]

BD is a 8 core, get over it already. And HT is not even close to BD. and blah blah blah

 

[BenchPress]

BD is a 8 core, get over it already.

I'm inclined to agree with that, but each core only has two ALUs while Intel has three per core. So it's not a very strong 8-core.

And HT is not even close to BD.

I disagree. The whole idea is the same: give the expensive front-end and SIMD cluster higher utilization by running two threads. Having 2 x 2 independent scalar integer ALUs or 3 shared ones is really just a detail and I believe they roughly perform the same for highly threaded software. Theoretically Bulldozer should even have a slight edge. When running one thread per core/module, Intel's architecture clearly wins.

So Bulldozer isn't horrible but they need technology like TSX to facilitate multi-threading, and they need AVX2 to keep up in the SIMD department. They can still compete against Intel, if they want to. I hope they grow a pair.

 

[Smartazz]

Even if Bulldozer is a true 8 core chip(and I think it is), its single threaded performance is horrendous. There's just way too many programs that need a very fast core. Even some multi core games require very fast cores, something Bulldozer does not have.

 

[Idontcare]

Core count, IPC, and clockspeed...all are figures of performance merit which on their own are meaningless when discussed in the absence of specifying the other two.

Add in factors of cost and power-consumption and the discussion loses even more relevance unless those two figures of merit are specified as well.

The issue with bulldozer is not core-count, or power consumption, or IPC...it is all of the above when combined into a singular product.

But that is where the great equalizer enters into the equation - price.

I include TCO when computing price though, for family and friends when it comes to recommending gear, and infortunately AMD does not. So while AMD does price their processors to be competitive with Intel on the shelf at Fry's, they don't price them to be competitive once you've factored in your TCO for electricity bills over the course of say 3-4 yrs that an individual will probably have the system.

But that is just me, and thankfully for AMD's sake they can count on most consumers not being like me, so pricing themselves to be competitive with Intel on the shelf is a wise move on their part, same as marketing their chips as "real" 8-cores and so on.

Can you blame them for trying? What kind of lame-ass company doesn't leverage their marketing division to highlight their strong points and minimize their weak points?

 

[pelov]

Why not call it what it is? A 4-module chip? You can't claim it's a true 8-core unless you say that Thuban is 6 super cores and the i7 2600K has 4 superawesomemagic cores. It's far simpler to refer to the BD architecture in CMT-based module count.

If you're going to refer to its performance and CMT tax then it performs somewhere between 6 and 8 theoretical cores due to ~-20% per filled module.

 

[Atreidin]

Anyone who doesn't like AMD will just define "core" to mean something that their processors don't fit. It's a nebulous definition anyway. Who cares?

 

[Charles Kozierok]

Who cares?

This is pretty much the bottom line: it doesn't matter. All that matters is performance and compatibility versus cost.

Arguing over cores and execution units is fine for hobbyists like us, but otherwise, it's not really important.

 

[frozentundra123456]

This is pretty much the bottom line: it doesn't matter. All that matters is performance and compatibility versus cost.

Arguing over cores and execution units is fine for hobbyists like us, but otherwise, it's not really important.

This is the main point to me. I view a CPU like a black box: All that really matters is the performance of the chip in the workloads you are doing. And by performance, I mean performance per dollar and per watt as well as raw processing power. Unfortunately for AMD, except in certain heavily threaded workloads, Intel wins on all 3 counts in the desktop space.