AMD to Intel: "Real men use real cores"

[twilyth]

According to an interview posted on April 6th, AMD's attitude toward Hyperthreading is that real cores are better.

Real men use real cores. We’ve got real cores across our products. Hyperthreading is basically designed to act like a core except that it only gives 10 to 15 percent performance bump for real applications workload. That’s because hyperthreading requires the core logic to maintain 2 pipelines: its normal pipeline and its hyperthreaded pipeline. A management overhead that doesn’t give you a clear throughput.

I have to agree. My assessment is:

Real cores = manmeat

Virtual cores = strap-on.

Comments?

[Serpentarius]

According to an interview posted on April 6th, AMD's attitude toward Hyperthreading is that real cores are better.



I have to agree. My assessment is:

Real cores = manmeat

Virtual cores = strap-on.

Comments?

+1

Intel's pricing is getting more and more agressive ....

[XS_Rich]

I have to agree. My assessment is:

Real cores = manmeat

Virtual cores = strap-on.

Comments?

Given that choice, I'd rather be facing a strap-on than manmeat...

[karateo]

Of course real cores are better!
If you can provide extra cores instead of HT at the same prices with intel, then you have sth new to say.

[twilyth]

Given that choice, I'd rather be facing a strap-on than manmeat...

I don't think it goes in your face, but I understand what you're saying.

[Sintel]

Except that one does not rule out the other

If a virtual core gives a 10-15% performance improvement for a 5% increase in transistor budget, then that is damn good spending of transistors...

That's the thing, of course a real core is better, but you have to double transistor count / die size for that extra core... And if you have lot's of threads that sparsely use cores (server-type workloads), then HT makes much more sense than having lots of cores with not much to do. It just isn't efficient use of available hardware.

Don't forget, nehalem and deneb have more or less the same transistor count, but there are vast performance differences depending on workload..

[jcool]

Seeing as HT boosts the WCG output of my i7's by around 40% as opposed to running with only 4 threads like ordinary quads, I have to disagree. The Core i7 ain't no octacore, and it isn't sold or advertised as one. Just see the HT function as a bonus - people that don't need it can turn it off and save some elec, the rest should be happy with the additional power.
Besides, my Phenom II 940 sucks so hard at crunching/OCing/anything but gaming it doesn't matter one bit if you use HT or not on the i7... it'll still wipe the floor with all of AMDs current lineup. I really hoped that 940 could do better..

[Mechromancer]

AMD (and I think VIA) is the only CPU microarch today that does not use some form of SMT. Intel, IBM, Sun all do 2 to 4 threads per core. I like to see REAL cores more than anybody, but Sintel is absolutely correct. The amount of extra transistors needed to make SMT work is tiny compared to the gain in performance.

It may be unknown to most (especially Intel folks) that if you disable both TurboBoost and Hyper Threading on a Core i7, you get the same or slightly less performance than a comparable Phenom II. Auto Overclocking and SMT do wonders in CPU reviews LOL. AMD can clock with Overdrive 3.0 profiles, but they need a little SMT lovin' to go along with that. AMD should have jumped on SMT after they saw P4-HT years and years ago.

[Throwed]

AMD (and I think VIA) is the only CPU microarch today that does not use some form of SMT. Intel, IBM, Sun all do 2 to 4 threads per core. I like to see REAL cores more than anybody, but Sintel is absolutely correct. The amount of extra transistors needed to make SMT work is tiny compared to the gain in performance.

It may be unknown to most (especially Intel folks) that if you disable both TurboBoost and Hyper Threading on a Core i7, you get the same or slightly less performance than a comparable Phenom II. Auto Overclocking and SMT do wonders in CPU reviews LOL. AMD can clock with Overdrive 3.0 profiles, but they need a little SMT lovin' to go along with that. AMD should have jumped on SMT after they saw P4-HT years and years ago.

Get your Intel loving butt outta here!.......Nah, just kidding. You make some excellent points. I'll be AMD until I die though.

[ownage]

According to an interview posted on April 6th, AMD's attitude toward Hyperthreading is that real cores are better.

I have to agree. My assessment is:

Real cores = manmeat

Virtual cores = strap-on.

Comments?

Depends on how you look at it.

Intel has manly CPU's at this moment if you look at the overall quality/performance. AMD however are more like lesbians. Trying to be manly but fail horribly. That doesn't mean you can't have fun with AMD, lesbians are fun.

[flippin_waffles]

Except that one does not rule out the other

If a virtual core gives a 10-15% performance improvement for a 5% increase in transistor budget, then that is damn good spending of transistors...

That's the thing, of course a real core is better, but you have to double transistor count / die size for that extra core... And if you have lot's of threads that sparsely use cores (server-type workloads), then HT makes much more sense than having lots of cores with not much to do. It just isn't efficient use of available hardware.

Don't forget, nehalem and deneb have more or less the same transistor count, but there are vast performance differences depending on workload..

One word. Cache. The core logic of Nehalem is about 40% larger than Shanghai. But Shanghai's cache makes up most of the difference in die size and transistor budget.

[jcool]

It may be unknown to most (especially Intel folks) that if you disable both TurboBoost and Hyper Threading on a Core i7, you get the same or slightly less performance than a comparable Phenom II. Auto Overclocking and SMT do wonders in CPU reviews LOL. AMD can clock with Overdrive 3.0 profiles, but they need a little SMT lovin' to go along with that. AMD should have jumped on SMT after they saw P4-HT years and years ago.

I have to disagree, even w/o HT and Turbo an i7 beats the PII clock for clock in just about every app out there. Add in the fact that the i7 OCs way higher using "standard" means of cooling (no ln2/liquid helium ) which makes it very hard for AMD to compete right now.

Note that I am not an Intel fanboy by any means (used to love my AMDs..), but I am a professional system builder and have to determine and use what's best for the people that want a pre-build high end setup. And right now, there is no way around the Core i7. For Budget/Gaming maybe, but not high end...

[alucasa]

Yep, Intel Core i7 is definitely a beast that AMD cannot touch at the moment.

[Particle]

When you consider that half of a die is for cache and a good deal more for the IMC, it becomes clear just how tiny a budget an actual core is. We're talking sub-100M. I think I saw somewhere that an actual "core" of the Phenom II is in the neighborhood of 25-30mm^2. Compare that to the overall die size of 260mm^2, and it doesn't look that bad anymore. Maybe 330mm^2 for a six-core chip if the caches stay the same.

Isn't AMD planning to implement HT by ~2011 anyway? The point may be moot as it's just a bonus if implemented well (ie, not like a netburst chip where performance often decreases).

One word. Cache. The core logic of Nehalem is about 40% larger than Shanghai. But Shanghai's cache makes up most of the difference in die size and transistor budget.

That doesn't jive. The i7 has 9.25MB of cache meaning a minimum of 465.6M transistors go to that. Deneb has 8.5MB of cache meaning a minimum of 427.8M transistors go to that. The rest of the chip is core logic, control logic, IMC, etc. While your statement may be true (I don't know), the supporting reason doesn't add up since Deneb has less cache.

[flippin_waffles]

Particle, just check out die shots of the two. Intel's cache is clearly more dense, and core logic much larger. Their cores have a much higher transistor count than Shanghai cores. Which should make a 6 core Istanbul a significant size smaller than the 6 core Nehalem, what ever that's called. Or maybe they can't get one out on 45nm, i don't know.

[Mechromancer]

Get your Intel loving butt outta here!.......Nah, just kidding. You make some excellent points. I'll be AMD until I die though.

Even in jest, you have effectively blasphemed by accusing me of being an Intel-lover! Take a good long look at my SIG. I dropped $2500 on a Phenom II rig when I could have easily built a Core i7 system well within that budget. I am zealous in my lust for GREEN POWER (not the tree hugging kind) .

I do have to confess, I have an Intel processor inside of my case. It is an IOP332 processor on my Areca ARC1210. It powers my SSDs inadequately too. I lust after IOP348 boards now...but that is as far into the BLUE ocean I will swim. Blast! If only AMD made IO processors, my system would be pure...

LOL

[Throwed]

Even in jest, you have effectively blasphemed by accusing me of being an Intel-lover! Take a good long look at my SIG. I dropped $2500 on a Phenom II rig when I could have easily built a Core i7 system well within that budget. I am zealous in my lust for GREEN POWER (not the tree hugging kind) .

I do have to confess, I have an Intel processor inside of my case. It is an IOP332 processor on my Areca ARC1210. It powers my SSDs inadequately too. I lust after IOP348 boards now...but that is as far into the BLUE ocean I will swim. Blast! If only AMD made IO processors, my system would be pure...

LOL

I noticed your sig a long time ago and gathered you were Pro-AMD from past comments. I had a Intel Q6600, but sold it a while back for funds for my Phenom II 940 system. That was my 1st and last Intel CPU. BTW, Amen to the I/O comment, but what can ya' do?

[xVeinx]

Hyperthreading is ultimately little more than using the core more efficiently. Granted, you have to add the core logic and design this in, but in certain applications it makes sense. Constantly feeding the core with threads to run makes sense rather than wasting cycles. On the other hand, you do waste efficiency in other areas, hence the performance loss in other programs. As previously mentioned, server applications and highly multi-threaded applications can benefit from this; applications requiring raw power are better off with it disabled. Some form of hyperthreading could be good for AMD, but it's a little early to tell if they will take the time to invest in it. They already beat Intel with that one already (sort of; hyperthreading v1 did have its uses, but wasn't implemented nearly as well as v2)-with more cores in fact. For the moment, they think that they can do this again. What would help them would be getting the IPC/integer values up alongside the increased cores, as well as increased efficiency in branch prediction, making the performance gain more substantial. More cores is better, but if Intel can do just enough work with hyperthreading, then they win in terms of die price, and therefore margins. AMD doesn't want to play the ever-increasing-die-size game again...

[demonkevy666]

Seeing as HT boosts the WCG output of my i7's by around 40% as opposed to running with only 4 threads like ordinary quads, I have to disagree. The Core i7 ain't no octacore, and it isn't sold or advertised as one. Just see the HT function as a bonus - people that don't need it can turn it off and save some elec, the rest should be happy with the additional power.
Besides, my Phenom II 940 sucks so hard at crunching/OCing/anything but gaming it doesn't matter one bit if you use HT or not on the i7... it'll still wipe the floor with all of AMDs current lineup. I really hoped that 940 could do better..

chickpatty has a big thread on TPU of this

AMD (and I think VIA) is the only CPU microarch today that does not use some form of SMT. Intel, IBM, Sun all do 2 to 4 threads per core. I like to see REAL cores more than anybody, but Sintel is absolutely correct. The amount of extra transistors needed to make SMT work is tiny compared to the gain in performance.

It may be unknown to most (especially Intel folks) that if you disable both TurboBoost and Hyper Threading on a Core i7, you get the same or slightly less performance than a comparable Phenom II. Auto Overclocking and SMT do wonders in CPU reviews LOL. AMD can clock with Overdrive 3.0 profiles, but they need a little SMT lovin' to go along with that. AMD should have jumped on SMT after they saw P4-HT years and years ago.

Hyper threading of any kind need more IPC on the chip

When you consider that half of a die is for cache and a good deal more for the IMC, it becomes clear just how tiny a budget an actual core is. We're talking sub-100M. I think I saw somewhere that an actual "core" of the Phenom II is in the neighborhood of 25-30mm^2. Compare that to the overall die size of 260mm^2, and it doesn't look that bad anymore. Maybe 330mm^2 for a six-core chip if the caches stay the same.

Isn't AMD planning to implement HT by ~2011 anyway? The point may be moot as it's just a bonus if implemented well (ie, not like a netburst chip where performance often decreases).



That doesn't jive. The i7 has 9.25MB of cache meaning a minimum of 465.6M transistors go to that. Deneb has 8.5MB of cache meaning a minimum of 427.8M transistors go to that. The rest of the chip is core logic, control logic, IMC, etc. While your statement may be true (I don't know), the supporting reason doesn't add up since Deneb has less cache.

I usually wonder why no one noticed this slight cache advantage even though it's minimal

Now remeber people L1 cache is set mount of data you can only fit so much into while Adding IPC.

right now your taking about a 32Kbyte L1cache that is 8 way vs a 64Kbyte L1 cache that is 2 way.

IPC's are different.

looking a bulldozer it seems as if AMD thinks it's got plenty of data in the core but not enough int' and FPU, per 64Kbytes,
they're putting two of each in the 64 Kbytes of cache.
From the looks of rumors it looks like 1 core is going to be 128Kbyes of L1 cache and data. ( but those are rumors/specutlaion)

Mt in Intel is being used to load parts of the core the aren't getting used much.
to me it's like saying there just isn't enough data to go around.

even though they only need 5% core space for 15% improvement with SMT.
scaling comes in to play eventually. Istanbul is adding 2 more cores and cache for 30% more with out increasing power.
and believe I7 uses 1% more power for 2% more performance over core 2 duo.
AMD was already on smaller core Brisbane vs conroe the core for conroe is very spread a part
Intel cores are very spread out compared to AMD cores.


126mm2

145mm2

cache density is on Intel side.
so I'd go with bigger L1 and more cores.
it scales better I'll stick to more's law.

[Mechromancer]

I noticed your sig a long time ago and gathered you were Pro-AMD from past comments. I had a Intel Q6600, but sold it a while back for funds for my Phenom II 940 system. That was my 1st and last Intel CPU. BTW, Amen to the I/O comment, but what can ya' do?

I was just trying to be funny, ie. knight/zealous crusader on horseback and green DAAMIT sign. LOL...I wasn't terribly serious, except for that damn I/O comment! I hate this IOP332 so much. I NEED MORE POWER!

Back to the topic of discussion: Bulldozer looks to be an over-the-top, brute force approach to AMD's fundamental problems. More IPC, more INT, more FP.

Bulldozer =

I can't wait for David Kanter to get his hands on solid Bulldozer specifications to do a writeup. In my other Bulldozer Patent post, the patents above my technical knowledge, atm, so I need it translated into forum-speak (a base, less-smart, form of English).

[jcool]

I hate this IOP332 so much. I NEED MORE POWER!

You mean something like this?

[Mechromancer]

jcool, are you implying that you know how to OVERCLOCK an Adaptec 5405 RAID controller?! That is sick! Just sick! I WANT IT NOW! Or am I overreacting and you're just curing the overheating problem those cards tend to have. Either way you're my hero for the day.

You also need to buy 4xVertex SSDs and bench for us!

[Anemone]

Intel says to AMD:

"REAL companies make profits..."

[YukonTrooper]

Sounds like a child crying. Until there are more than four physical cores at retail, AMD will only sound ridiculous making comments like this. I honestly don't get it. Intel has four cores with the added benefit of HT, and AMD has just the four cores, yet AMD is trying to claim some sort of superiority? Give me a break.

[villa1n]

but he does have a good point here.

Now, Intel will rave about the performance of QPI but it’s only if you buy their top end parts. If you buy their mid-range parts, the QPI speed drops down, and if you buy their lower end parts the QPI speeds drop down even more.

Meaning that if you have an application that rely on high I/O and high memory throughput but doesn’t need a lot of compute power, like a Web server, a file server or network infrastructure - which are the real backbone of today’s data centers - you would have to buy the fastest Nehalem processor to get the fastest QPI! Instead, we offer the same HyperTransport speed on all of our Opteron chips.