AMD - Piledriver chitchat

[DanceswithUnix]

Is there no room for anything in the closest slot, or does it only apply to RAM with heatspreaders?

Distance could certainly be an engineering concern; longer PCB traces would introduce greater inductance and capacitance. Even the speed of light has to be factored into some design aspects. Engineers also have to try to keep traces on the same clock the same length, this is why you see squiggly traces to make up extra length.

Sadly I think it is simpler than that. I notice the uATX motherboards that look OK to me are 9.6" x 9.6". The cheaper ones not as deep, like 9.6" by 8" or 9.6" by 7.2". They save the cost of a couple of memory sockets, some VRM regulator circuitry, and multi layer circuit board is quite cost per square inch for a low end motherboard.

[CAT-THE-FIFTH]

Trinity retail CPUs will be released on the 2nd of October:
http://semiaccurate.com/2012/09/24/a...nd-gen-apus-2/

It also appears FM2 will be longer lived too.

[Biscuit]

Trinity retail CPUs will be released on the 2nd of October:
http://semiaccurate.com/2012/09/24/a...nd-gen-apus-2/

It also appears FM2 will be longer lived too.

Link didnt work for me mate

[CAT-THE-FIFTH]

It seems fine for me??

[watercooled]

Sadly I think it is simpler than that. I notice the uATX motherboards that look OK to me are 9.6" x 9.6". The cheaper ones not as deep, like 9.6" by 8" or 9.6" by 7.2". They save the cost of a couple of memory sockets, some VRM regulator circuitry, and multi layer circuit board is quite cost per square inch for a low end motherboard.

The slimmer ones are ideal for some applications though, like more cramped cases, but yeah you're probably right.

[DanceswithUnix]

The slimmer ones are ideal for some applications though, like more cramped cases, but yeah you're probably right.

I think a slimmer motherboard would be just what I am looking for as the motherboard that just died in my Silverstone SG02 is full uATX sized and was really hard to get in there.
Was looking at AM3+ motherboards to replace the old AM2 one that failed, and there really doesn't seem to be much choice these days. Came to the conclusion the only one I liked was the again 9.6x9.6 sized Asus M5A78L-M/USB3, the only one that was 140W cpu rated and came with 4 memory slots. It won't have a CPU that power hungry, I won't be overclocking, and at least to start with it will have 8GB of ram on 2 sticks and no intention to upgrade. Thing is I hate buying a board knowing that from day one I have maxxed it out, and a 95W rated motherboard in that case strikes me as too likely to overheat. That might be what happened to the old board.

If FM2 is just around the corner then I will wait. Just hope this one isn't another paper launch, I could do with getting that machine up and running again.

[shaithis]

I've had 2 Asus M5A78L-M/USB3 boards for a while now and cannot fault them really. They were even running the first shipping BIOS until a couple of weeks ago when I dropped BD chips into them and finally flashed them for the first time. I run at least one of them 24/7 as well.

[scaryjim]

... an FM2 motherboard should be upgradable to steamroller and I'd be gobsmacked if the Mainstream steamroller isn't a lot faster, both CPU and GPU than Trinity. I'm getting much better steamroller vibes than I had for any AMD product since the Athlon 64.

Steamroller looks like it fixes a core problem with Bulldozer that Piledriver won't, which is a narrow front end. AMD's simulations obviously suggested that a narrow front end would keep enough instructions in flight, but it didn't. If they can get that right and finally give the modules a chance to breathe, we could see something pretty good. As to the GPU, the move from VLIW4 to GCN appears to offer about 15% efficiency increase (interestingly, the same as VLIW5 -> VLIW4) - at which point I guess the question is how many shaders they'll cram into Kaveri...

Sadly I think it is simpler than that. I notice the uATX motherboards that look OK to me are 9.6" x 9.6". The cheaper ones not as deep, like 9.6" by 8" or 9.6" by 7.2". They save the cost of a couple of memory sockets, some VRM regulator circuitry, and multi layer circuit board is quite cost per square inch for a low end motherboard.

I'm sure that's part of it, but my A55 board is the 9.6" square, and the socket's still jammed up to the RAM slots *shrug* There must be a specification or technical reason for it. Of course, once you've got that, dropping the second 2 ram slots and slicing off an inch of board is a great way to save money. I've noticed that recent entry level boards also seem to be dropping expansion slots, to the point that there are A55 boards out there with just one PCIe x16 and one PCI slot (double spaced, of course!) - cost cutting taken to the extreme...

It also appears FM2 will be longer lived too.

I think it was pretty much accepted that Kaveri was going to be on FM2, although that does mean they'll be keeping dual-channel memory for yet another generation. Wonder how they're going to generate the necessary memory bandwidth for the GPU on Kaveri? stacked RAM maybe?

[DanceswithUnix]

I've had 2 Asus M5A78L-M/USB3 boards for a while now and cannot fault them really. They were even running the first shipping BIOS until a couple of weeks ago when I dropped BD chips into them and finally flashed them for the first time. I run at least one of them 24/7 as well.

Thanks, that's good to know. This is my home server, runs my dns email etc so is on 24/7. TBH an FM1 motherboard would be fine, but I don't want to buy parts for this machine that will be hard to find replacements for in just a few months. So longevity card leaves AM3+ or FM2. Of those, FM2 should be more energy efficient which means a quieter system. Right now my HTPC has been re-purposed, so I don't have to choose just yet.

[DanceswithUnix]

Interesting GF roadmap here. Wasn't AMD will be trying to get Steamroller out in late 2013, or is that all in my head? That would give them a choice of 20nm though.

http://www.brightsideofnews.com/news...mobility-.aspx
They seem to have mixed nodes and half-nodes in that slide, but still interesting though.

[scaryjim]

I'm interested as to when AMD will move production onto GF's half-node processes. It looks like GF is currently offering 28nm, and presumably at some point AMD designs will end up taping out on that (Kaveri perhaps?), followed by 20nm and 14nm (which I think again are half-nodes?). Since AMD APUs are all currently on 32nm, and Kaveri is slated for 28nm next year, it'd be a big jump for them to suddenly go 14nm? Interesting that 32nm doesn't even feature on that slide - presumably it's targeted at mobile/ARM customers? Which presumably means GF are only running 32nm for AMD?

Lots of very confusing business decisions being made there...

EDIT: OK, can anyone explain to me in laymans terms the difference between half-node and full-node? I'm not sure I'm getting it *sigh*

[kalniel]

EDIT: OK, can anyone explain to me in laymans terms the difference between half-node and full-node? I'm not sure I'm getting it *sigh*

Half-node is just where you have a small enough change that your existing designs work. Full-node requires you both shrink the process and redesign the circuitry/transistors design, either to better package it or to overcome/take advantage of new physico-electric characteristics of the new process.

[scaryjim]

Ah, OK So it's to do with the size of the change, rather than the actual size of the node?

[kalniel]

Ah, OK So it's to do with the size of the change, rather than the actual size of the node?

Yeah, size of the work required I guess.

[watercooled]

It's not really as clean-cut as it once was; a while back, some fabs like TSMC would offer the 'standard' ITRS nodes followed by half-nodes when available, offering lower power consumption/smaller die size/etc. The half-nodes could generally be made without significant changes to the fab or circuit design. Take some older GPUs as an example, you had 130>110>90>80>55 but from then on 40>28, skipping the ITRS nodes, essentially following their own 0.7x path instead of the ITRS> 0.9x ITRS path.

I'm not entirely sure why they chose that path, it could be a cost thing; developing two nodes in the time the likes of Intel develop one is probably fairly hard to say the least, especially if most of your customers only want one of them, and leapfrogging Intel instead of just catching up then getting overtaken again is probably better for marketing. I'm just speculating though.

The node size isn't everything though, for example 32nm SOI is said to offer similar power/performance to 28nm bulk, but SOI is possibly more expensive. Intel were apparently considering to switch to SOI for 22nm but considered it too expensive so chose FinFETs, or '3D transistors', instead (as reviews have proved, it's not some epic improvement - it's simply a necessity of the shrink). This is partly why, despite being behind on node size, AMD's process has kept close to Intel in terms of power/performance - Intel use bulk CMOS, AMD use SOI.

Transistor density can also vary regardless of node. For example, low power processes tend to be less dense, a sort of power/die size tradeoff.

[DanceswithUnix]

I'm interested as to when AMD will move production onto GF's half-node processes. It looks like GF is currently offering 28nm, and presumably at some point AMD designs will end up taping out on that (Kaveri perhaps?), followed by 20nm and 14nm (which I think again are half-nodes?). Since AMD APUs are all currently on 32nm, and Kaveri is slated for 28nm next year, it'd be a big jump for them to suddenly go 14nm? Interesting that 32nm doesn't even feature on that slide - presumably it's targeted at mobile/ARM customers? Which presumably means GF are only running 32nm for AMD?

Lots of very confusing business decisions being made there...

EDIT: OK, can anyone explain to me in laymans terms the difference between half-node and full-node? I'm not sure I'm getting it *sigh*

I think you had it, hence I said I thought the roadmap was confused (for the purposes of marketing, so not surprising they are carefully missing stuff out).

32nm is a full node, it was hard for them to get there and I don't think it would have looked so impressive on that chart. 28nm is a "half node", a finesse of the full node and ISTR they only offer a subset of design rules on it hence no AMD CPUs on it.
To make it worth doing, traditionally fabs always try and halve the size of silicon required to do a chip so if the area goes down by 2 the feature (node) size drops by square root of 2, then round it to a nearby number that they think they can get away with.

sqrt( 2 ) is about 1.4, so starting with 65nm node I believe it goes:
65/1.4 = 46, we got 45nm node.
45/1.4 = 32, we got 32nm node.
32/1.4 = 23, Intel went for 22nm. Looks like GF are really rounding down to ret and get 20nm.
20/1.4 = 14. Believe it when I see it, these things never seem to go smoothly