AMD's Bobcat and Bulldozer

[Manicdan]

Well, yeah adding more cores means a reduction of frequency. Same basic idea of Hyperthreading, sacrificing performance in a single thread to potentially gain throughput benefits by doing work in a second thread.

ok lets start over, your first post says:

It at least doesn't hurt games as much as adding two cores in Thuban.

i believe you have the idea way to generalized for an adequate comparison.

2 cores were added because the power efficiency improvements AMD has made allows it to add so much more. when we first got a 125W cpu, it was 3ghz. soon we will have 3.5ghz at the same envelope. and btw we now have hexies at 3.2ghz, which is faster than the original fastest quads.

if you wanted something more than a quad, then your right, reduce mhz, add cores. but its a purchase choice, not a bios toggle that says if your maxed out with a quad or hex

however we can start to argue what the turbo does, turning 6 slower cores into 3 faster ones, which is relying still on the tdp being the same. what if instead of re-releasing the 1055T at 95W, they left it at 125W, but it would 4 turbo cores at 3.5ghz.

owners of the 1090T have full control over turbo, they can set the number of cores, and even what the max multi is for each core at turbo, and at stock. if one of them decided to do such a thing and convert their 3.2x6/3.6x3 into 3.2x6/3.5x4, and benchmarked it, i bet it would still fall into the same tdp range.

AMD did try to prevent apps from becoming slower with the hexies by adding turbo, but since it wasnt built from the ground up that way, they were a little limited. duel core apps like most games, should notice an increase

or we can also just argue that every Thuban clocks better than Deneb and there is no excuse if your cpu ends up being slower then deneb.

basically im trying to point out that such a generalized statement really dosnt make it easy to prove if true, since there are so many different ways one can solve such an issue.

[STaRGaZeR]

Fair enough, but that situation is becoming increasingly rare since most if not all the new games benefit from HT/multicores.

Lame excuse. HT hurts past, present and with 99% probability future games too, period. HT is good for certain things, and bad for others, this is one of them. Don't try to hide this fact. We still don't know how HT is done in SB, so I'll not comment on that.

You can hear all the bullcrap that X game supports Y number threads blablabla. The truth is that most games only load 1, 2 or maybe 3 threads to the max, the others are just for minor things that are done in a short period of time.

[terrace215]

One way to think of hyperthreading is that it is about as far down the "spectrum of sharing" that you see in a BD module as you can go. That is, imagine that the 2 integer halves of each BD module have even more of their resources shared.

As a (HT sharing-level) consequence, the second thread does not add as much additional performance.

But also as a (HT sharing-level) consequence, when not running a second thread, the first thread gets even more of the "module's" resources to itself, and performs better.

In the BD-level sharing, the non-shared resources are made less wide, to conserve power, and allow for more overall cores and throughput in a heavily loaded situation, at the cost of some single-thread performance.

So, really, the two concepts are about different levels of resource sharing per pair of threads.

In SB, the core (or "module", if you want) has been given more resources (e.g. load added to the store port, store added to the load port: meaning you can also do 2 loads or 2 stores instead of just 1 load & 1 store) which should particularly help HT performance.

Edit:

To expand on this, going the other direction from BD-level sharing, you arrive at 2 completely separate cores per "module".

So there are a number of knobs to turn:

1. How many resources are shared?
2. How wide/narrow are the shared resources?
3. How wide/narrow are the UNshared resources?
4. How many modules on the part?

And then you have largely power-related (some die size) constraints, and you want to optimize for various workloads.

So you can argue about what the best settings are for the knobs, but really, a BD "module" sits on the same design spectrum, just in between separate cores and an Intel HT core. Meaning, I suppose, that it is not some qualitatively different thing, so rather than "OMG Hypertheading sucks!11 BD modules rock!!!!!1", perhaps more refined arguments about the choice of design knob settings would make sense?

In this context, my new sig can be expressed as: Turning down knob 1 means that the constraints & optimization require knob 3 to be turned down, as well, unless you want to mess with 4, which you don't.

[accord99]

basically im trying to point out that such a generalized statement really dosnt make it easy to prove if true, since there are so many different ways one can solve such an issue.

I'm just referring to the stock factory situation. So while Thuban's extra cores doesn't hurt much in gaming performance, in reality neither does Hyperthreading. Looking at this review by IXBT, enabling Hyperthreading has essentially no impact on gaming performance:

http://ixbtlabs.com/articles3/cpu/ar...2009-5-p8.html

[Manicdan]

I'm just referring to the stock factory situation. So while Thuban's extra cores doesn't hurt much in gaming performance, in reality neither does Hyperthreading. Looking at this review by IXBT, enabling Hyperthreading has essentially no impact on gaming performance:

http://ixbtlabs.com/articles3/cpu/ar...2009-5-p8.html

charts were a little confusing at times, but i dont quite see how HT was tested, since they turned turbo on too.

[Manicdan]

I got bored here at work and put together a little table/chart of what i expect to see for core scaling with BD, relative to that of known Deneb and Thuban. keep in mind i did not try to compare IPC between any of the chips, just difference within itself should be noticed. the idea is relative efficiency between old and new stuff

Left to Right is # of threads running
Top to Bottom is the core # (in the second column) and their performance relative to a single core
and at the bottom is the sum of all cores, followed by a chart.

for BD
my assumption is that for every 80% core turned off, the other core will get about 5-10% stronger for IPC (extra L2), and then 15-20% turbo, so 25% total increase. and then as its using less than 4 threads i added 5% more turbo per drop

for Thuban i put in 15% gains for 3 or less threads since its close enough, some chips its up to 19%, others 12%

[accord99]

charts were a little confusing at times, but i dont quite see how HT was tested, since they turned turbo on too.

I compared the HT disabled 4 core scores with the HT enabled 8 core scores. I considered Turbo to be a small factor given that its a i7 950 which will only Turbo up at most 2 multipliers and the games are generally multi-threaded which lessens the chance of Turbo activating.

[Dimitriman]

I don't understand this whole discussion of pro-Intel's trying to derail BD modules and pro-AMD trying to derail Hyperthreading.

It all comes down to how many threads can be processed in a given die size and power envelope.

If SB can run 16 threads within 300mm2 @ 100W power consumption, while AMD can do exactly the same with 8 BD modules, they both look equally as attractive to me in a MARKETING point of view, which ignores all the very fine architectural differences.

Also, when it comes to performance per core:
For Intel, if they can do HT in 1 core without sacrificing die size, then WHY NOT?
If AMD can do 50% better than Hyperthreading by adding only 20% more die space, then WHY NOT?

Both look like very elegant solutions to me regardless in which way you look at it. What I'm really looking foward to seeing in the future is REVERSING this solution = Using more than 1 core to run 1 thread. THAT is an interesting and EFFICIENT method to boost your overall performance and utility.

[Calmatory]

What I'm really looking foward to seeing in the future is REVERSING this solution = Using more than 1 core to run 1 thread. THAT is an interesting and EFFICIENT method to boost your overall performance and utility.

[Dimitriman]

Insightful contribution.

PS:

here's other things i look forward to:

3rd generation fusion processors
100% integration of 64 bit OS
DDR4
Low power high performance graphics cards


You have any problems with my personal expectations and wishes for the industry?

[Manicdan]

Insightful contribution.

i think hes saying that if he had to think about how its possible, or how to code for it, it would be like hitting his head against a wall.

[madcho]

Reverse hyperthreading, it's stay a myth for now

[FlanK3r]

AMD confirms the AM3+ as a necessity for a Zambezi product:
http://www.planet3dnow.de/cgi-bin/ne...&id=1282840508



It looks like they had a difficult choice to make and they went with more performance which is ,in the end,a right thing to do.Present AM3 system owners can use their AM3 CPUs in the new boards so not everything is broken compatibility wise. Use the old chip with the new board when AMD launches it and then just slide in the new 8 core Bulldozer when it comes next year.Not the perfect solution for present owners ,but if it is for performance reasons then it's understandable .

good news, because i prefer better performance. Now i can waiitng at new boards

[Dimitriman]

i think hes saying that if he had to think about how its possible, or how to code for it, it would be like hitting his head against a wall.

Well, I would take back my stress if thats the case.

I hardly ever make any technical statements in this site as I am not any qualified engineer or specialist. So I understood he was mocking my opinion.

apologies to the readers for any drama.

Back on the topic, if reverse hyperthreading is so hard to achieve, were all the news that Intel is officially announcing its development just Fud?

[Chumbucket843]

i would not consider that reverse hyperthreading. anaphase is a thread than looks at memory access patterns from the real thread and prefetches data based off of what algorithm it uses. most applications have a high cache hit rate and most of the latency is actually waiting on cache.( L3 isnt that much faster than a good DDR3 burst read or write.)

[informal]

Apart from being extremely difficult to solve(the "Reverse HT"),there is an easier way of upping single thread performance - dynamically increasing clocks based on load levels,the road both intel and AMD are taking.Of course,there is a continuous work on the core level improvements with the new generation of chips,those help too but are not cheap and easy like clocks.

[JF-AMD]

Bulldozer and HT are both ways of looking at supporting larger numbers of threads. People are getting caught up in semantics and dicing/slicing the architectures instead of focusing on what is the performance, what is the power consumption and what is the price.

Those are the 3 that matter to customers.

[Chumbucket843]

that's already been done on itanium. it has 64 clockspeeds. the gain in efficiency was ~25%.

the 4 things that will bring single thread perf up is faster and smarter caches, faster and smarter branch prediction, faster main memory and increased clockspeed. note that two of those are speculative. it is likely that working on interconnects and parallelism will gain more performance.

[madcho]

Bulldozer and HT are both ways of looking at supporting larger numbers of threads. People are getting caught up in semantics and dicing/slicing the architectures instead of focusing on what is the performance, what is the power consumption and what is the price.

Those are the 3 that matter to customers.

Why slicing the architectures ? We look for know how it works, if performance it's gonna be great or not.
Ok we can't really know how is the process.

For moment AMD don't give real performances numbers, intel has done it with SB, preview already on anantech ...

We are customers, if performances power consumtion and price are not good, we won't buy it.

For moment we are slicing architectures, to get an idea, what's next on AMD side that's all.

If AMD need wait before doing some previews, or some real bench to the world, we don't have any choice, we just wait, and try to know, what is best our next upgrade.

[Dimitriman]

Why slicing the architectures ? We look for know how it works, if performance it's gonna be great or not.
Ok we can't really know how is the process.

For moment AMD don't give real performances numbers, intel has done it with SB, preview already on anantech ...

We are customers, if performances power consumtion and price are not good, we won't buy it.

For moment we are slicing architectures, to get an idea, what's next on AMD side that's all.

If AMD need wait before doing some previews, or some real bench to the world, we don't have any choice, we just wait, and try to know, what is best our next upgrade.

The problem with your logic is that you're forgetting Bulldozer will not fully compete against these SB shown in Anand, they will actually have to take on an even faster and better equipped processor and platform. The previewed SB is meant to take on Llano.

So in order avoid giving intel info on the competition of their unreleased SB chips, AMD is keeping quiet.

IMO they could and should give some benchies on Ontario and Bobcat, as well as Llano. I don't see any reason to hold those back as they are sooner to be released and their competition is already layed out.

[-Boris-]

Low power high performance graphics cards


You have any problems with my personal expectations and wishes for the industry?

There are low power graphics cards released every year that is a lot faster than it's predecessor.
But it's always possible to gain lots of performance while using more power. So if they can double the performance by doubling the heat output of that low power card, they will. You can still buy these low power cards, it's up to you.

What if (then) they come up with an idea that gives us twice the performance of GTX480 while only using 90 watts? They will use the same technology to make a card six times faster while consuming 270 watts. And you will still complain that they should make this monster in the same 90W enveloupe.

[Chumbucket843]

Bulldozer and HT are both ways of looking at supporting larger numbers of threads. People are getting caught up in semantics and dicing/slicing the architectures instead of focusing on what is the performance, what is the power consumption and what is the price.

Those are the 3 that matter to customers.

true, but looking deeper will provide insight to why it is faster. people are curious about what your company makes! all we do is buy a peace of etched silicon and it just magically runs faster?

another potential reason for these discussions is to predict what these companies will do next. hardware geeks think its cool to be an insider for tech products. what needs to improve? performance? duh. how does one do that? use a dual rail self resetting domino bypass multiplexer. it gets technical and nerds love it!

fyi, i buy hardware that i think is interesting and cool. performance matters but i'd really like to know what's going on in there, maybe a story about the design too. there are many failure and success stories behind chips. it's pretty cool.

[Dimitriman]

There are low power graphics cards released every year that is a lot faster than it's predecessor.
But it's always possible to gain lots of performance while using more power. So if they can double the performance by doubling the heat output of that low power card, they will. You can still buy these low power cards, it's up to you.

What if (then) they come up with an idea that gives us twice the performance of GTX480 while only using 90 watts? They will use the same technology to make a card six times faster while consuming 270 watts. And you will still complain that they should make this monster in the same 90W enveloupe.

Quite true

[Mats]

The problem with your logic is that you're forgetting Bulldozer will not fully compete against these SB shown in Anand, they will actually have to take on an even faster and better equipped processor and platform. The previewed SB is meant to take on Llano.

I find that hard to believe. That midrange 2400 they tested is very close to the 980X, that's very fast (although with HT, which it won't have, but no turbo). While I do think BD will be fast, I can't imagine a Llano being close to the 980X, which is what you indirectly say.

Remember that the LGA1366 i7 950 will be replaced by a LGA1155 i7 2600, according to the article.
I don't think we'll see any nice price i7 920 successors for the LGA 2011.

[mAJORD]

Somethings not right with that chart ^^^^^

These Perf Numbers suggest an i7 2600k will smash an i7 880..