AMD Discloses Architecture Details of High-Performance “Carrizo” System-on-Chip | Hacker News Reader

AMD Discloses Architecture Details of High-Performance “Carrizo” System-on-Chip

amd.com

52 pointsmatt4225 comments

The AnandTech article is probably better: http://www.anandtech.com/show/8995/amd-at-isscc-2015-carrizo...

Calling Carrizo "high-performance" is interesting because AMD reduced peak performance to increase power efficiency. Below ~65W Carrizo's performance should be higher than Kaveri (although still lower than Intel) but at higher power budgets performance would be worse.

I will be interested to watch how the hUMA stuff works out. Having 100% overlap in memory space between the GPU and the CPU can be useful. Intel tried to go there with Larrabee. Having a collection of specialized processors in the same cache coherent space allows for some interesting data flow architectures to be tried out.

Do they have a common MMU? Have they released information on how that works in practice?

I guess this must be a solved problem since there have been many systems with similar memory architectures, but I wouldn't be surprised if it's quite a challenge.

Cache coherency can also have interesting security properties.

Honestly, while reading, I couldn't help myself thinking about the "bullshit font" from one of yesterday's (or was it the day before) submissions.

prestigious - wealth of - new, advanced - substantial - true ... bleah.

Impressive what they manage to do on 28nm compared to Intel.

If you forced both companies to use the same die size then AMD compares favorably to Intel. But on the leading edge of performance, AMD lags Intel by quite a bit (Intel was selling 14nm processors at the end of 2014, and AMD isn't planning to release them until 2016). It forces AMD to compete in other ways, because they just don't have it when it comes to raw speed and efficiency. Putting CPU and GPU on the same chip is nice, but I long for the days when AMD and Intel were duking it out and high end users actually had to debate which was better.

I long for the days when AMD had Intel thoroughly beat. That brought us the Core2Duo.

Intel is not exactly lacking in competition these days, either. The ARM/mobile threat is pushing really hard against Atom.

Intel is definitely making a play for the low-powered market (including wearables and embedded) and experiencing an uphill battle, but they're still pretty much uncontested everywhere else. I seem to recall some discussion of ARM trying to move into servers but I haven't heard anything about it since.

I always have this bad feeling when it comes to Atom, like they are intentionally crippling it and are not exactly giving it a proper shot. They are so far ahead on the raw manufacturing that the Atom should have no problem stomping the competition outright, yet it's barely scraping by.

Early atoms where unquestionable intentionally crippled. During the netbook era, by leaving out stuff like out of order execution, SSE functionality, no x86_64 bit extensions and forcing vendors to bundle crappy intel GPUs and limiting systems to 2 gigs of ram and XP only. I'm shocked there was never a lawsuit over this.

Its hard to say how much that continues today, but one gets the distinct impression that atom is still held back much like the other low end intel processors (pentium, celeron) with stuff like AES-NI and AVX disabled.

Seems to be part of a ploy to preserve high end margins and fend off ARM on the lowend with "just barely good enough" chips.

I miss those days as well.

Silver lining is that the processor market is mature/saturated, so Intel still has some fierce competition: itself. Nearly everybody buying an Intel desktop/laptop CPU already has an Intel desktop/laptop CPU. So Intel still has a lot of incentive to improve - they need to beat the performance of the Intel CPU you already bought in 2011 or 2012 or 2013 or whenever.

I wonder why they don't double down on Open Source. Their architectures, since the ATi merge, have been very creative, providing a lot of heterogeneous computing power but it seemed it was unusable unless you were using their Windows drivers.

In the videogame space, optimized drivers represented a competitive advantage with some drivers detecting a title and substituting a particular kernel. This edge is somehow lost when going OSS.

If only AMD was at least on the most cutting edge non-Intel process. But they can't even seem to be able to do that.

Imagine what they could do with Samsung's money and process (Samsung's 10nm process is almost ready - http://www.tweaktown.com/news/43761/samsung-stuns-world-show...) if they were acquired.

It's been difficult to take claims of "high performance" from AMD seriously for a while :)

They may not manufacture the highest-performing cores, but they have consistently provided the best performance for the price. Who cares who's fastest? The consoles went with them because they were fast enough and better priced.

AMD's problem for a long time has been power consumption. They have had the performance per dollar crown, but their stuff runs hot. Only place I ever see them is gamer market and some servers.

Add that the performance per dollar lead is only if your workload is atypically multi-threaded. Intel has really taken a bite out of AMD's already tenuous market with the "Pentium" variants.

This somewhat reminds me of the Netburst era, when AMD's Athlons were in the lead while Intel decided to chase after high clock frequencies with much higher power consumption as a result. Then Intel went back to the P6 to produce Core, and rapidly overtook AMD.

I still hope AMD will make another sudden leap like that, to keep the x86 market more competitive, but they seem more intent on keeping to the path started by Bulldozer...

But you're commenting on a press release about how this architecture's main benefit is gains in performance-per-watt. AMD has a reputation for using more power right now because they slowed down to focus on unifying architectures. I think ARM has proved this strategy works in the long term. Everybody knows power efficiency is king right now.

If electricity were free, you'd have a great point. TCO is something that needs to be considered, and Intel's chips are very competitive.

Intel makes 160 wH chips too, but you would probably compare something like a 95 wH Core i7 with a 125 wH FX. That's about $42/year in the USA, but that's assuming running at full power 24/7. Real cost would be a lot lower.

> they have consistently provided the best performance for the price

Need a citation on that.

For low thread count or FP heavy applications, i.e. the vast majority of cases, Intel's Pentium G series has been the best bang for the buck for several years now. AMD is only cost effective for applications with lots of threads doing integer work which isn't terribly common.

You could look at Passmark:

http://m.cpubenchmark.net/cpu_value_available.html

Low-end Xeons are on top, but the first you might want are the FX-6300 and 8320, followed by the Pentium Gs. Since you are not comparing equal cores, the cheaper Pentiums do make sense in a lot of cases, but AMD does often have cheaper motherboards. Actual performance is going to depend on use case. I feel like the console situation speaks for itself. AMD is still putting out some good stuff, especially for an undedog.

Reply on news.ycombinator.com