Comment by rkalla | Hacker News Reader

Comment by rkalla | Hacker News Reader

Parent Full thread

XBit did a good calling out what AMD is doing here that is different (and in its infancy, disappointing with traditional benchmarks).

AMD has decided to step back from the brink of maximum-speed per core, to maximum-cores-per-chip.

If you look at the Anand benchmarks: http://www.anandtech.com/show/4955/the-bulldozer-review-amd-...

you'll see in very multi-threaded applications (like compression or encoding) the AMD CPU actually keeps pace with the 2600k; in other applications where it is more important to have faster single threads of execution, the Intel 2600k pulls ahead by a drastic amount though.

These are the early days, but if AMD gets this down to 22nm next year and puts 16 cores on a desktop PC -- each core may be only running at 4Ghz, but in a scientific computing or server environment where you have 16, 4Ghz cores as opposed to 4, 6Ghz cores on some equivalent Intel can make a big difference.

The new Bulldozer platform looks to be a strategic company gamble on all fronts (more cores, less speed per core) that could win AMD big in the server space (which is exploding) and in scientific computing environments.

I'm not saying the numbers aren't disappointing, I was hoping to see some big ones in there to. I'm only pointing out that this isn't quite apples to apples, and when it is (e.g. straight multithreaded benchmarking) the comparisons look OK.

All that said, I'm sure Intel has something waiting in the wings (e.g. 2800k) to shit all over this party and beyond that we have tri-gate IvyBridge CPUs next year that will hopefully bring a much-needed* leap to CPU performance.

* "needed" as defined by "my impatient ass likes fast things.

Ok, but based on the die size and the transistor count we could have had 12 Thuban cores in the same amount of space.

It really feels like if we are going to go with more cores, why not make them simpler? Why not dust off the old Athlon K7 at 37 million transistors. Roughly 54 of those fit into the 2 billion transistors making up the 8 core bulldozer. Bulldozer has an Itanic feel to it, no that is too harsh, it has a P4 feel to it.

I really feel like desktops are done, GPUs have this biggest effect on desktop perf right now. It would seem to focus on the fusion A series parts and the server. As a server side dev, I would rather have more cores (more cache actually). I don't need memory coherency (that should be mitigated by the hypervisor).

First, power. Second, the old K7's are not as fast as you remember.

I don't think power is it. Your saying the K7 generates more heat per computation (more inefficient use of transistors) than the newer cores? I was about to conflate power usage with computational work, not necessarily the same thing... The power will drop using the new process because the voltage is much much lower (1v vs 2v)

I remember the K7 doing just over 1 GOP (giga operation per second), so that would put the whole cluster in 50 GOP range (if clocked at only 1Ghz like the Thunderbird).

I wasn't being directly literal, I will allow these mythical engineers that work for me using all the 2 billion transistors to refine things from previous designs.

It is true the power usage of a K7 per op will drop if you redesign the core to use 32nm. That's a key detail though- you can't just drop a 130nm design onto a 32nm mask.

There are many hitches with this idea though. K7 cores have limited power management capability; modern cores have sophisticated frequency scaling, and shut down sectors that are idle. K7 cores are 32-bit. K7 cores do not support most modern extended instruction sets. 50 active K7 cores on one die will generate way more than 125W, causing cooling problems. In short, there are a lot of issues with dated processors that we don't think about, because we take the advances for granted.

This is not to say your idea does not have merit. If it was I, I would probably first investigate using a Bobcat core. The Bobcat core has low power usage for a x86 core (good for packing a large number of cores on one die), and considering a single Bulldozer 'module' (core) has 200M transistors, the Bobcat probably has ?100M?, allowing you to fit ?20? cores with 2 threads each.

P.S. If you are about to brush off the extended instruction sets bit, consider who a 50-core die would interest most. My guess is the scientific community would be a large part of the target markets, and I believe they make great use of extended instruction sets.

sit, you are exactly right and they are going that route with the Opteron's. The 16-core Opterons were announced a while ago, I believe there is GPU infrastructure in the desktop chips (similar to what Intel is doing with the always-includes-a-GPU-now CPUs).

I imagine when we see the 16-core AMD's next year on 22nm, we'll have 24-core Opterons or something equivalent that is a boon to the hosting industry.

Not saying Intel won't be right there with their own solution, but going the "more cores" route should be interesting to watch and see if it pays off.

Honestly is sounds exactly like the strategy Sun pursued with UltraSPARC systems; however in my experience, time and time again commodity hardware with higher per-core clock speeds beat the pants off of Sun hardware except for very narrow HPC tasks (supposedly, I couldn't even get decent speeds for the distributed computing projects available on the platform).

It also seems to rely on software to be written specifically to take advantage of multiple local threads. This may or may not happen in a timely fashion for most projects (and again, chances are fewer cores with a higher clock speed will still outperform).

Time will tell though I guess!

Agreed -- it really seems to be a long term bet. They see the growth of hosted apps and cloud environments and I am sure that is making up a large margin of their sales (and have been for years).

I think they are banking on dominating the multi-tenant server market with these designs and if consumer PCs follow suite with some convoluted virtualization mechanism in the years to come, that might leave them sitting pretty with a lot of disparate, fast cores ready to execute the work thrown at them.

If not, and quad-core is all we ever really utilize, then Intel is going to dominate that arena as they have for the last decade.

in very multi-threaded applications (like compression or encoding) the AMD CPU actually keeps pace with the 2600k

What is the difference in price between these CPUs?

The 2600k is about $75 more expensive. The 2500k, which is closer to the FX8150 most of the time, is about $25 cheaper.

2600: $299: http://www.newegg.com/Product/Product.aspx?Item=N82E16819115...

FX-8150: $279: http://www.newegg.com/Product/Product.aspx?Item=N82E16819103...

I just picked up an i5 2500k from Microcenter for $179. A great processor at a good price.

http://www.microcenter.com/single_product_results.phtml?prod...

you'll see in very multi-threaded applications (like compression or encoding) the AMD CPU actually keeps pace with the 2600k;

Er, no it doesn't. In integer limited operations, such as winzip it keeps up. In anything using floating point, its 20% slower at least. At AAC encoding its nearly twice as slow!

I think you'd be right if AMD could halve their power consumption to at the very least match Intel.

No doubt they have some tightening to do, but they are always a generation behind Intel when it comes to fab; when they get the new architecture onto 22nm I think things should even out, but then Intel will be jumping to something much smaller at that point in a tick-tock cycle anyway.

Cat and mouse.

What kind of leap? The jump to an i52400 has been enormous for me. It's hard to imagine another similar one over the core 2 duo I was on before.

Do you think AMD are going to be ready for memristor based memory?

Last I read, the tri-gate IvyBridge is almost entirely focused on power consumption, with a very small incremental boost in performance.

Do you have a link handy?

I think you are right, I recall reading something similar that we were going to get huge power gains (like 30% improvement) for equivalent performance. That being said I expected Intel to fill up the high end with an appropriately overlocked CPU with the typical 95/130 TDP heat payload.

If AMD doesn't push them to do it though, I suppose they have absolutely no reason to out-do themselves.

Anandtech did a decent piece on this last month [1].

To paraphrase their description: every other year Intel keep their manufacturing process the same, but release a new microprocessor architecture (Sandybridge was one of those years), but in the other years they introduce a new manufacturing process (e.g. the new tri-gate 22nm transistors which will be found in the Ivy Bridge).

[1] http://www.anandtech.com/show/4830/intels-ivy-bridge-archite...

Right-o, that is the "tick-tock" terminology that I believe Anand popularized.

They keep tech the same and shrink die (32->22) then in the tock, they keep the die size the same and improve tech. Then the next tick, take all that new tech and shrink it again and on and on.

Seems to have worked out well for them.

SandyBridge (as you pointed out) was the "tick" (32nm) with the 22nm "tock" IvyBridge planned for next year.

I think the wildcard in this "tock" is the tri-gate transistor stuff. It'll either be incremental and we'll see our standard 15% hop in performance, or it will unlock a lot of potential on the high end for power savings, over clocking and growth which could be very exciting.

OR, we are all blindsided by ARM chips from Qualcom/NVIDIA/Samsung and start working on our phones :)

Reply on news.ycombinator.com