AMD's Next-Gen Steamroller CPU Could Deliver Where Bulldozer Fell Short

MojoKid writes "Today at the Hot Chips Symposium, AMD CTO Mark Papermaster is taking the wraps off the company's upcoming CPU core, codenamed Steamroller. Steamroller is the third iteration of AMD's Bulldozer architecture and an extremely important part for AMD. Bulldozer, which launched just over a year ago, was a disappointment. The company's second-generation Bulldozer implementation, codenamed Piledriver, offered a number of key changes and was incorporated into the Trinity APU family that debuted last spring. Steamroller is the first refresh of Bulldozer's underlying architecture and may finally deliver the sort of performance and efficiency AMD was aiming for when it built Bulldozer in the first place. Enhancements to Fetch and Decode architecture have been made, as well as increased scheduler efficiency and cache load latency, which combined could bring a claimed 15 percent performance-per-watt performance gain. AMD expects to ship Steamroller sometime in 2013 but wouldn't offer timing detail beyond that."

Comments

AMD has cool code names.

[Anonymous Coward]

They all sound like sexual positions.

Re:AMD has cool code names.

[Johann Lau]

And Intel ones don't? Who are you kidding?

Aladdin
Bad Axe
Bad Axe 2
Batman
Batman's Revenge
Big Laurel
Black Pine (a cute name for anal sex I guess)
Black Rapids (I don't want to know)
Bonetrail
Caneland
Cougar Canyon
Glidewell
Tanglewood (sounds bi to me)
Warm Springs

and last, but never least, the

Windmill (also known as "Helicopter Dick")

Re:AMD has cool code names.

[dbIII]

In some races they are just about alone on the track. An AMD based server with 64 cores and 128GB of memory will set you back $9000. with Intel you can now get 80 cores for about ten times that, or 40 cores for about five times that.
For some tasks when you can get 640 slightly slower cores (the ten core Intel chips have a lower clock than the ones with less cores) for the same price as 80 it's pretty easy to see which way to go. If anything is massively parallel you can forget about Intel at this point.

Re:AMD has cool code names.

[gweihir]

Indeed. I think AMD is actually far ahead of Intel (again, think e.g. integrated memory controller, for quite a few server-loads Intel was vastly behind for a time due to that). The speed increases of CPUs have become slower and slower and mater less and less. The trick for AMD will be to survive intact until Intel gives up and gets a next-gen architecture of their own. By then AMD will have ironed out the kinks and they will be on an equal footing again. When looking at their relative sizes and cash-reserves, it is impressive that AMD can compete at all. But the bottom-line is that in almost all cases (exception: You need a small number of CPUs with high power because your software is stupid, and cost of the CPUs is not an issue) you get significantly better value for the money from AMD.

They need to innovate

[Gothmolly]

Things like hitting the 1GHz mark first, and making a workable 64bit chip that also speaks x86 only get you so far. AMD needs to come up with something cool, else they're doomed to play catch-up.

Re:They need to innovate

[ArhcAngel]

Well they definitely need to step up with their current offerings but I will forever be grateful for their 64 bit x86 extensions. If not for that we'd be stuck with Itanium desktops...*SHUDDER*...

Re:They need to innovate

[tlhIngan]

CPUs don't really drive software development that much. Or else we would have migrated off x86 years ago. If intel can get the same/similar performance without a paradigm shift in development methodology, developers won't bother.

The migration from x86 has already started, actually - the architecture they're moving to is ARM. (After all, there are more ARM-based SoCs shipped than x86 CPUs - every PC includes one or more ARM cores doing something).

But on a more user level - tablets, smartphones are becoming the computing platforms of the day, all running ARM processors. Regardless of whether they run iOS or Android. Developers have embraced it and cranking out tons of apps and games and other stuff for this. It's so scary that Intel's investing a lot of money bringing Android to x86 because the writing's on the wall (when more phones and tablets ship than PCs...)

But x86 won't die - it has a raw performance advantage that ARM has yet to reach, so for computation-heavy operations like databases, it'll be the heavy lifter. Perhaps serving an entire array of ARM frontend webservers.

Re:They need to innovate

[iamhassi]

Ya know what? Nothing wrong with cheap and "good enough" the problem has been their new designs are cheap and shitty thanks to that lame "half core" they went for.

You take a good 85%+ of the people out there and a MOR AMD Deneb quad will frankly be twiddling its thumbs because it will blow through any jobs that they have, even gaming, even more so for Thuban. And their Brazos chips were fricking great, an APU designed for mobile video and basic tasks that got great battery life while often being cheaper than an Atom+ION setup.

I've sold many an Athlon II and Phenom II and the people are damned happy with them, they just blast through everything they want to do with plenty of cycles left over. I even put my money where my mouth is with regards to my family, me and the oldest are gaming on Thubans while the youngest took my Deneb, and they blow through any game we throw at 'em.

I see from TFA they've partially dropped the "half core" design but I can only hope that with Piledriver they'll drive a stake through it, as most of the people I've talked to Win 8 is a DO NOT WANT yet the half core scheduler bug is only fixed in Win 8. Meh, hopefully I'll still be able to get enough Thuban, Deneb, and Liano chips to get me through the whole BD/SR phase and the new Apple chip designer they hired will give us another Athlon64. One can hope after all.

This. I have a six-core 1055T. Bought it to overclock and it does hit 4ghz stable on air but guess what? I run it at stock 2.8ghz. Why? Because 99.9% of the time six cores at 2.8ghz is more than enough. Even games run perfectly. CPUs have finally reached the point where faster isn't better anymore, its power usage and heat output. Rather have it run cool using little power at stock then run it full blast all the time sucking watts and heating the room at 4ghz I'm not even using.

When I bought this intel didn't have anything close in price that performed as well. Sure I could have spent double and bought a faster intel chip, but why? What was the point of spending more on something I wouldn't use? Rather spend the $ on a ssd for real performance gains then extra ghz I'd never use. So I bought AMD and I'll probably do it again next year if the price is reasonable and the speed is "good enough"

Re:They need to innovate

[the_humeister]

This is just so weird. 20 years ago it was Alpha, MIPS, SPARC, PA-RISC, etc. that were the ones counted to do all the heavy lifting backend, HPC stuff. x86 was kind of a joke that everyone frowned upon but tolerated because it was cheap and did the job adequately for the price. Then x86 steamrolled through. Now no more Alpha or PA-RISC. MIPS is relagated to low-power applications (my router has one).

In Linux drivers, Intel is still king.

[BadgerRush]

AMD may be getting its shit together when in regards to chip design. but I'm still going Intel on my next PC because of their superior Linux drivers. At the moment I'm an unhappy owner of a laptop with a AMD graphics card that can't do anything because the drivers are useless. I'm looking forward to a new laptop with an Intel Ivy Bridge processor (I don't think I can wait Haskell).

AMD's in deep trouble with Steamroller

[PhrostyMcByte]

I think AMD's work here will provide some great evolutionary speedups that will be significant to many people. Unfortunately for them, at the same time AMD is bringing out these small "free lunch" general improvements, Intel will be bringing out Haswell -- which in addition to such evolutionary improvements has some really fantastic, significant new features that'll provide remarkable performance boosts.

• Integer AVX-256. For apps that can take advantage of it, this'll be a massive speed-up. Things like x264 and other video processing will take huge advantage of this.
• SIMD LUTs. One of the major optimization tricks programmers have been using for ages, look up tables, have been thus far out of reach to SIMD code without complex shuffle operations that usually aren't worth it.
• Transactional memory. This is not quite the easy BEGIN/COMMIT utopia people are hoping transactional memory will eventually get us, but it's a building block that'll enable some advanced concurrent algorithms that either aren't realistic now or are so complex that they're out of reach to most coders.

These are all pretty specialized features, yes, but they service some very high-profile benchmark areas: video processing and concurrency are always on the list, and AMD will get absolutely crushed when apps start taking advantage of it.

I'm a developer, a major optimization geek both micro- and macro-. I thrive playing with instruction latencies, execution units, and cache usage until my code eeks out as much performance as possible. Of course we'll never know until the CPUs are released for everyone to play with, but right now my money is on Intel.

AMD is in serious trouble here. I hope I'm wrong.