Thoughts on "8 Core" Bulldozer and "4 Core Sandy Bridge"

[Arachnotronic]

So I was thinking...it seems to be a concern that it takes 8 AMD BD cores to match/beat 4 SNB cores, but my question is, why is this a big deal? In GPU world, it takes 1536 AMD "Cores" to match 480 NVIDIA "Cores", but nobody cares because when we choose our GPUs, we care about the bottom line performance of the part, not the per-core performance.

So why can't we see the CPU in a similar fashion? Software is moving towards being, in general, "multi-threaded", so we should see increasing execution cores in the same way as increasing cache, widening/shortening pipeline length, improving branch predictors, etc. -- just one aspect of the final product.

What do you all think?

 

[Gigantopithecus]

I don't care about core counts and clock speeds. I care about real-world performance, power consumption, and platform cost.

 

[podspi]

First, I think you're correct. Final price, performance, and power usage should be all that matters.

That having been said, I think Bulldozer's performance vs SB is going to be a toss up in many applications. For heavily threaded integer and non-AVX FPU workloads, I believe 8-core BD will be competitive with the 6-core SB-E. For single threaded integer workloads, I think SB will be faster than Bulldozer, and for single threaded AVX workloads it'll be a toss-up.

Because Bulldozer will be slower than SB in some things, and in others faster, AMD decided to keep the price relatively close to the 2600K.

The above of course is just pure speculation on my part Please don't quote me for anything except congratulations if I'm right, or mockery if I'm wrong!

 

[smakme7757]

Before we see what Bulldozer is capable of it's all just guess work, but as a consumer i would still buy an AMD product even though Intel can get the same performance with less cores. In the end it doesn't matter if the performance is delivered by the product.

Of corse the underlying technology is exciting and interesting, but i really don't think it matters how the companies achieve their performance, be it Ghz, cores or some other crazy way of moving the bits!

 

[GammaLaser]

So I was thinking...it seems to be a concern that it takes 8 AMD BD cores to match/beat 4 SNB cores, but my question is, why is this a big deal? In GPU world, it takes 1536 AMD "Cores" to match 480 NVIDIA "Cores", but nobody cares because when we choose our GPUs, we care about the bottom line performance of the part, not the per-core performance.

So why can't we see the CPU in a similar fashion? Software is moving towards being, in general, "multi-threaded", so we should see increasing execution cores in the same way as increasing cache, widening/shortening pipeline length, improving branch predictors, etc. -- just one aspect of the final product.

What do you all think?

It's because we haven't reached a point where CPU workloads are "massively parallel," and it will probably always be the case that some "fat" cores are needed to help in single-threaded situations. Many apps written for CPUs are not like GPGPU apps where performance can scale linearly with core counts numbering in the hundreds or thousands. Things like synchronization/locks/mutexes and cache coherency can become a big problem. There's also Amdahl's Law--single thread performance can become a limiting factor even for multithreaded apps anywhere the computation has to be serialized.

 

[richierich1212]

Intel's philosophy: Pack bigger cores onto a CPU
AMD's philosophy: Pack more smaller cores onto a CPU

 

[drizek]

An 8GHz CPU will always be twice as fast as a 4GHz one. A dual core 4GHz CPU will be somewhere between 0-100% faster than a single core.

So for AMD to compete, they need to use a dual core 6GHz CPU that would be at somewhere between 75% and 150% of the performance of an 8GHz single core. Then consumers who use single threaded apps will choose the Intel and the ones who use multithreaded ones will be better off with AMD.

Which basically explains the design of Bulldozer. Lots of shared cache, many small cores, and high clock speeds. The execution of Bulldozer... well that's another matter.

 

[nyker96]

I dont think design matters, it's just the end result. they can use 1000 cores if they like, at the end I just care a) performance b) power c) OCablility. that's it. whoever wins on all 3 wins my money.

 

[Gigantopithecus]

An 8GHz CPU will always be twice as fast as a 4GHz one. A dual core 4GHz CPU will be somewhere between 0-100% faster than a single core.

Only if the architectures are the same. As I said, clock speed is irrelevant compared to real world performance. 1.8GHz single core vs. 3.6GHz single core.

 

[Phynaz]

Increase single core performance and all applications and situations benefit.

Increase the number of cores and only applications that can use those additional cores benefit.

Increase the number of cores and multi-tasking may improve, depending upon circumstances.

 

[pantsaregood]

Multicore is just another dying marketing ploy.

Clock speed was used for marketing for an extremely long time. Throughout the 1990s and early 2000s, everyone was in a clock speed race.

L2 cache was a major selling point in the mid-90s.

Multicore was a huge selling point in 2005.

Now it isn't anything special. Just something expected. No one cares about L2 cache anymore. Most people realize that clock speeds aren't all-important. Multicore is becoming no different. It is a perfectly legitimate method of increasing performance, but it is no longer a huge selling point. The consumer doesn't care (or really, need to know) how many cores a CPU has, what its clock speed is, or how much cache it has. The consumer simply cares how it performs.

 

[drizek]

Well, obviously, I'm assuming perfect and consistent scaling.

That was a pretty entertaining comparison though...

 

[Edrick]

Multicore is just another dying marketing ploy.

Clock speed was used for marketing for an extremely long time. Throughout the 1990s and early 2000s, everyone was in a clock speed race.

L2 cache was a major selling point in the mid-90s.

Multicore was a huge selling point in 2005.

Now it isn't anything special. Just something expected. No one cares about L2 cache anymore. Most people realize that clock speeds aren't all-important. Multicore is becoming no different. It is a perfectly legitimate method of increasing performance, but it is no longer a huge selling point. The consumer doesn't care (or really, need to know) how many cores a CPU has, what its clock speed is, or how much cache it has. The consumer simply cares how it performs.

Exactly. And now extensions like AVX, FMA, etc. are important selling points.

 

[Skurge]

1.8GHz single core vs. 3.6GHz single core.

Well, I didn't realise this thing was that slow.

BD might be faster for some situations and SB might be faster for others. Just like AMD cards are faster for some games and nVidia cards are faster for others.

 

[exar333]

Intel's philosophy: Pack bigger cores onto a CPU
AMD's philosophy: Pack more smaller cores onto a CPU

True.

It is like a college football team trying to play an NFL team. They can compete, but they need an extra 4-5 players to do it.

BD is great, but it kind of like AMD is throwing in the towel for performance and saying they can only complete if they have more cores to do it with. That really does limit them in a lot of applications. Performance remains to be seen, but I really don't think most average users will care about BD initially anyway. They don't need a lot of cores to run their browsers and iTunes, just a quick processor and enough memory to run the applications.

 

[Nemesis 1]

Well, I didn't realise this thing was that slow.

BD might be faster for some situations and SB might be faster for others. Just like AMD cards are faster for some games and nVidia cards are faster for others.

That review would be really interesting today . If you compared a P4D Single core against single core Hammer. Using todays bench marks you would find out that the programmers were lazy good for nothings. Intel being first with dual core followed closely by x2 made the programmers do multi thread. Those old cores run on benchies today the outcome would be much differant until X2 comes along .

 

[sm625]

what makes you say that? A celeron 450 is still faster than a Pentium D 840. I know because I tested both a couple weeks ago to see how much better it is. 2.4 GHz is the level of OC it takes a celeron 4xx to beat a Pentium D 840 at 3.0GHz.

 

[Nemesis 1]

So do same test with AMD hammer . single core and tell us all about it. I have an old P4C 3.2 here running . I would be happy to compare against AMD hammer on todays multi thread benchies.

 

[Tuna-Fish]

So do same test with AMD hammer . single core and tell us all about it. I have an old P4C 3.2 here running . I would be happy to compare against AMD hammer on todays multi thread benchies.

I got a A64 3500+ serving as a second machine, pick any bench you want to.

 

[alyarb]

So why can't we see the CPU in a similar fashion?

We do. BD cores are the first of their kind in x86 and represent a big departure for AMD not only with respect to K10, but the most dramatic divergence is really how AMD has chosen to differentiate from intel with respect to SMT concepts. Historically AMD does not achieve anything like the xtor density of intel, and yet a BD core is quite large compared to an SNB core. The tradeoff is that on paper a BD core should be much more flexible than hyperthreading as far as wringing out every last drop of threadability out of a process. Truly, a BD core should actually appear and behave like a pair of traditional cores. Intel can't exactly say that about their architecture. By contrast hyperthreading is super efficient for some workloads because of the small area required to implement it, but on other workloads it has no affect at all so you just have a bunch of transistors taking up power and area and not contributing to overall throughput in any way.


It all boils down to how AMD and intel have totally committed to two SMT concepts that are so distinct from one another such that the industry will be literally entrenched, head-to-toe, in this SMT dichotomy. It will become totally improper, by the end of the year, to use a term (like "core") to make these broad direct numerical comparisons, just as it is improper to compare nvidia shaders to radeon shaders.

All of you seem to be on the right track though as far as taking the easy way out, and likening a single BD mod to a pair of cores. This does make the chip sound bigger than its SNB counterpart and if a core-to-core arms race is coming, AMD has a big semantic advantage.

 

[Nemesis 1]

I got a A64 3500+ serving as a second machine, pick any bench you want to.

I will pick an old one just for fun than you pick one

PCmark 05 I held the world record with mine at a given GHZ . i even scoreded the gold trophy at XS in That special I think it HWbot what ever. I go look see . Ha they removed me from there list .Big ass babies. Off course at that time I was running the xtpe

 

[Tuna-Fish]

I will pick an old one just for fun than you pick one

PCmark 05 I held the world record with mine at a given GHZ . i even scoreded the gold trophy at XS in That special I think it HWbot what ever. I go look see . Ha they removed me from there list .Big ass babies. Off course at that time I was running the xtpe

I'm having trouble installing it -- apparently it requires windows media encoder 9, which has been discontinued. How did you install it?

 

[386DX]

First, I think you're correct. Final price, performance, and power usage should be all that matters.

That having been said, I think Bulldozer's performance vs SB is going to be a toss up in many applications. For heavily threaded integer and non-AVX FPU workloads, I believe 8-core BD will be competitive with the 6-core SB-E. For single threaded integer workloads, I think SB will be faster than Bulldozer, and for single threaded AVX workloads it'll be a toss-up.

Because Bulldozer will be slower than SB in some things, and in others faster, AMD decided to keep the price relatively close to the 2600K.

The above of course is just pure speculation on my part Please don't quote me for anything except congratulations if I'm right, or mockery if I'm wrong!

I think you're greatly over-estimating the performance of BD imo. a 6-core SB-E is going to kill it at virtually everything.

IMO BD is just AMD bringing the market back to when the Phenom II X6 was introduced:

4-core i7 vs 6-core X6:
Intel >> AMD @ Single (light) thread
Intel < AMD @ Multi-thread

4-core SB vs 6-core X6:
Intel >>> AMD @ Single (light) thread
Intel > AMD @ Multi-thread

4-core SB vs 8-core BD (my prediction)
Intel >> AMD @ Single (light) thread
Intel < AMD @ Multi-thread

6-core SB-E vs 8-core BD (my prediction)
Intel >> AMD @ Single (light) thread
Intel > AMD @ Multi-thread

* >> means significantly greater, > means slightly greater

 

[ultimatebob]

I don't care about core counts and clock speeds. I care about real-world performance, power consumption, and platform cost.

Personally, I care about what I can buy NOW... not what I can probably buy three months from now.

 

[aka1nas]

So I was thinking...it seems to be a concern that it takes 8 AMD BD cores to match/beat 4 SNB cores, but my question is, why is this a big deal? In GPU world, it takes 1536 AMD "Cores" to match 480 NVIDIA "Cores", but nobody cares because when we choose our GPUs, we care about the bottom line performance of the part, not the per-core performance.

So why can't we see the CPU in a similar fashion? Software is moving towards being, in general, "multi-threaded", so we should see increasing execution cores in the same way as increasing cache, widening/shortening pipeline length, improving branch predictors, etc. -- just one aspect of the final product.

What do you all think?

Because 3d rendering is an embarrassingly parallel problem and has fairly good scaling regardless of which the the two GPU architectural design choices you mentioned. General software that you run on your CPU is unfortunately still not as friendly to multi-threaded execution and there are much more pronounced limits to multi-core scaling.