AMD - Piledriver chitchat

[CAT-THE-FIFTH]

It seems one of the websites looked at Crysis3 again. It seems HT causes problems in the game and the part they tested was the most intensive part of the game which has loads of vegetation which has a highly multi-threaded load. It seems during other parts of the game a Core i7 with HT switched off tends to be faster,but in the most intensive part is no faster than the AMD CPUs.

In that case,it makes me think that a FX6300 probably seems a better value AMD CPU than an FX8350. It will be easier to overclock with a cheaper motherboard,and will match a FX8350 in the more lightly threaded parts and still have Core i5 level performance in the most intensive parts.

[kompukare]

Or the FX4300. Both of which should overclock better than a FX8300 since they use less power and a high clock will compensate for their lower IPC (vs i5/i7). For extreme overclocks 'dark silicon' must come into it too, which begs the question: which cores/modules are disable on these. On the 2 modules FX4300 if the 'dark silicon' modules are diagonally to each other (as opposed to opposite), will that make a few C° difference?

Anyway, whatever is going on with HT the i3 looks poor value. Still Intel must be happy because their actual costs (die sizes) have gone a lot while their prices have not. Just sticking with i5's (because they're all listed on one wiki page):
i5 Lynnfield (45 nm) = 296mm²
i5 Clarkdale (32 nm) = 81m²
i5 Sandy Bridge (dual-core, 32nm) = 131mm²
i5 Sandy Bridge (quad-core, 32nm) = 216mm²
i5 Ivy Bridge (dual-core, 22 nm) = 118mm²
i5 Ivy Bridge (quad-core, 22 nm) = 160mm²

Of course, some of those are really what's considered i3 now (2C/4T). Wonder if anyone has graphed die size, transistor count, RRP and some IPC measure (even passmark) and plotted a trend. Certainly the days of big increases or better prices have almost ended.

[CAT-THE-FIFTH]

The problem with the FX4300 is that AMD decided to deactivate half the L3 cache and then price it at around £90!!

Edit!!

http://blogs.amd.com/play/2013/03/05...ool-is-coming/

??

[watercooled]

You're also going back to 4 cores with the 4300, and Crysis 3 seems to scale fairly well beyond that.

IPC isn't that much of an important metric by itself, and is fairly unreliable when comparing between architectures. Much like clock speed, it's a number and is only really a useful metric within a microarchitecture. Also, IPC doesn't necessarily scale well at all with clock speed, for example memory bandwidth might start to become a bottleneck and reduce the benefit of throwing clock speed at it. There are a *lot* of factors that dictate shipping clock speed, and lower shipping clock speed doesn't necessarily imply a better uArch like some trolls like to think. I know you weren't implying any of that, but it's a good opportunity for me to rant.

When comparing die size/transistor count, you also have to factor in things like integrated GPU. Clarkdale had most of the uncore on a separate 45nm die, which is why the die size is so small.

[kompukare]

Or for that matter, just throwing transistors at the problem doesn't always help. That's why a trend plot would be hard to do. I know the days of a doubling in CPU performance every 2-3 years at the same price has come to a stop - and not just at the budget end of the market.

Recently both Intel and AMD have been throwing lots of transistors at their iGPUs (where transistors do help), but while AMD is 'willing' (ok, forced) to offer more features (core count, unlocked multipliers, vm support etc.), Intel is so comfortable that probably half or more of the CPUs they sell have gone through the fuse blowing machine...

I mean Intel have the biggest pockets, the best fabs and yields yet they only have two or three dies for S1155. Their good yields mean that pretty much every Celeron or Pentium they sell is probably a chip which would have yielded an i3 but the fuse blowers have turned off half the features.

We'll see what Hasswell brings (BLCK overclocking is meant to return) but I still haven't found a value replacement for my Pentium E5300@3.6GHz. Of course, the main reason being that it's 'good enough' and better idle power is about the only thing I'd like.

[CAT-THE-FIFTH]

The FX6300 has got to be one of the best value CPUs out there. On top of this it seems AMD is refreshing the Vishera CPUs,with an improved version in June it seems according to rumours.

They you have CPUs like the FM2 Athlon II X4 CPUs like the Athlon II X4 750K for around £60 to £65 which can be overclocked a decent amount. They seem to hit around 4.5GHZ from the stock 3.4GHZ or thereabouts.

Look at how the Core i3 CPUs are falling behind in more modern games,and a dual core without HT is really not that great for games now??

Look at the overclockable Intel CPUs?? They are hitting £180 to £190 already,and unlike the rest of their range are nicely going up in price. Some say its down to exchange rates,but it really seems to be only the socket 1155 K series CPUs which are showing this trend.

[watercooled]

I know it's not necessarily that important, but I prefer to buy a CPU knowing it is as good as it was intended to be, not something which started out good but has had features uselessly disabled* to fit a price point. I know it's a way to upsell some people to a more expensive CPU, but I won't be caught paying through the nose to have features not disabled for a small gain. I do own a Celeron G540 as it was a fairly good fit for its purpose, I only really needed half-decent performance (i.e. not the Atom board I was trying to use for the same purpose), reasonable power consumption*2 and low cost. The fused features likely wouldn't have much impact in this case, and the rock bottom pricing means it doesn't bother me too much.

*Aside from cache, pretty much everything else which is fused off cannot be broken without the whole core being broken. Hyper-threading, virtualisation stuff, instructions sets e.g. AES-NI and so on, are not isolated groups of transistors which can break independently - either the core works, or it doesn't. There is no technical reason they are disabled.

*2 Veering off-topic a bit here, but something which might be useful to others. Originally, using Ubuntu server, power consumption was nice and low, but some time later I tested power consumption and found it had increased substantially (can't remember exact values). After some testing I discovered it was because of some kernel update which had disabled some low power states for the IGP as apparently it rarely caused some systems to lock up. This is one example of why I like the Debian model of testing for ages before release, locking versions/features, then only patching serious/security issues. I know each way has it's pros/cons, but here they are making power consumption/battery life worse for *everyone* using a given platform, to avoid a pretty rare problem which can be solved fairly easily (essentially the opposite of the fix I used).

Anyway, the fix is to change a boot parameter. In /etc/default/grub change the default
GRUB_CMDLINE_LINUX_DEFAULT="quiet"
to
GRUB_CMDLINE_LINUX_DEFAULT="quiet i915.i915_enable_rc6=1"

Then run update-grub and reboot, fixed.
If someone was having the lock-up problem, it would be as simple as doing the above but changing =1 to =0.

[CAT-THE-FIFTH]

I think the best value CPU I have seen in the last few years was my mates Athlon II X3. It unlocked to a Phenom II X4 and that was when they were £100 and the Athlon II X3 CPUs were around £60.

[DanceswithUnix]

Some Steamroller chitchat:

http://www.brightsideofnews.com/news...-unveiled.aspx

Kevari with 6 cores??

[watercooled]

It would make sense if they are planning to do away with the AM3 platform. Like we've been saying all along, I'm not sure where that leaves servers unless they're really counting on HSA in that area. Missing L3 cache is still concerning, but maybe architectural tweaks and more memory bandwidth mean it isn't worth the significant extra die space?

About GDDR5, I wonder how that will work? Lack of expandability wouldn't be great, although I assume there will also be DDR3/4 platforms available? Unless they're counting on motherboards to include soldered-on memory, maybe they'll include something on-package?

[DanceswithUnix]

Well that might be the "PCIe endpoint" support they are talking about. Wouldn't scale to huge numbers, but say you have a 4 socket system where three of the cpu sockets are PCIe slaves to one master socket, that might work? Edit to add: I'm thinking out loud here, with more out loud than thinking

As for GDDR5, AMD (and Nvidia for that matter) have been allowing that choice on graphics cards for years, I guess they are just unifying their designs into one memory controller layout.

[DanceswithUnix]

Less than a week until Richland?

http://www.fudzilla.com/home/item/30...-12th-of-march

[Terbinator]

Hyped for this.

I've been considering ditching my build for a while now as all i seem to do is play LoL. If these have got enough oomph i may take the plunge and grab one of those Streacom passive cases too - hopefully there will be a 6700 65W one.

[DanceswithUnix]

Some interesting digging from Charlie:

http://semiaccurate.com/2013/03/08/o...ow-power-apus/

So quad core Jaguar is showing about the OpenCL GPU grunt of the A6-3400M, and the dual core tablet variant coming in at half that.

[CAT-THE-FIFTH]

Some new CPU releases:

http://www.cpu-world.com/news_2013/2...pre-order.html

Some interesting digging from Charlie:

http://semiaccurate.com/2013/03/08/o...ow-power-apus/

So quad core Jaguar is showing about the OpenCL GPU grunt of the A6-3400M, and the dual core tablet variant coming in at half that.

Imagine a mini-ITX motherboard with the quad core Jaguar SOC??

[watercooled]

The VIA Nano Quad looked fairly interesting, but aside from the press releases there doesn't seem to be an awful lot of information about them. IIRC Nano performed significantly better than Atom, and had an encryption engine, but wasn't as power efficient. It's a shame they're not more widely available though.