AMD - Bulldozer Chitchat

[kalniel]

I guess that's the crock of gold - let's hope they can find the end of the rainbow I don't know how simple it would be to take flops that were programmed and coompiled for running on an x86 (well, x87 technically?) CPU and translate them to a vector compute unit. Still, if a math co-processor was good enough for us in 1980...

I think one of the purposes of GCN is to be much more able to do exactly that kind of thing, especially the x86-64 address space:
http://www.anandtech.com/show/4455/a...-for-compute/6

[Terbinator]

I was going to jump onto a Llano laptop with my student loan (currently cheap at Lenovo) But do you think it would be worth waiting for these to troll around in a few months?

[shaithis]

Depends.....if Llano is more then sufficient to perform the roll you want it to perform for the next 1-3 years.....why wait?

If you think you'll be pushing it and it may let you down in certain areas....wait.

[CAT-THE-FIFTH]

I was going to jump onto a Llano laptop with my student loan (currently cheap at Lenovo) But do you think it would be worth waiting for these to troll around in a few months?

I saw the deal on HUKD - was it something like £350 to £400 for an A8 based laptop?

[Terbinator]

I saw the deal on HUKD - was it something like £350 to £400 for an A8 based laptop?

Yeah, there is a lots of chopping and changing but what annoys me is that its nowhere as near as good as the pre-Christmas offer in terms of pricing/selection.I understand the majority of this rests on the HDD situation, but getting nobbed on the hard drive price and only being left with a measly 320GB is infuriating.

Saying that, you can squeeze a lot of power into a sub £500 price from a respected company - not often you can say that.

[Biscuit]

320GB is still decent for a laptop, especially if you have a desktop aswell. My laptop hard drive is only 120GB and i ust offload onto my NAS when its full

[Terbinator]

320GB is still decent for a laptop, especially if you have a desktop aswell. My laptop hard drive is only 120GB and i ust offload onto my NAS when its full

Yeah, i do agree to an extent.

It's just that i'd never specifically buy 320GB in a laptop if there was smaller, cheaper options, and given the current price - no option of an SSD is very annoying.

[scaryjim]

I didn't know that the FPU could only do two 128 bit AVX instructions! That would explain a lot.

OK, reading around more I'm utterly confused. I've found this evaluation at extreme tech saying its SSE FP performance is roughly the same as Thuban despite the 4 FPUs v 6, presumably benefitting from pushing 2 128bit instructions through each FPU to give 8 simultaneous flops but at lower IPC? They also show that double / float AVX performance is poor due to having to split 256bit instructions into 2x128 (perhaps it's the AVX bit of the FPU that's not working and it's having to push the two 128bit instructions consecutively, rather than concurrently?), whilst int AVX is much higher than an i7 2600k.

The more I read about bulldozer's architecture and performance, the less I understand it

I think one of the purposes of GCN is to be much more able to do exactly that kind of thing, especially the x86-64 address space

If it happens it'll be a huge step forward and give Intel a real kick in the flops. Now all they need to do is get the lightly threaded int performance sorted, and they'll be home and dry

[kalniel]

If it happens it'll be a huge step forward and give Intel a real kick in the flops. Now all they need to do is get the lightly threaded int performance sorted,

OK, keep it realistic...

[CAT-THE-FIFTH]

OK, reading around more I'm utterly confused. I've found this evaluation at extreme tech saying its SSE FP performance is roughly the same as Thuban despite the 4 FPUs v 6, presumably benefitting from pushing 2 128bit instructions through each FPU to give 8 simultaneous flops but at lower IPC? They also show that double / float AVX performance is poor due to having to split 256bit instructions into 2x128 (perhaps it's the AVX bit of the FPU that's not working and it's having to push the two 128bit instructions consecutively, rather than concurrently?), whilst int AVX is much higher than an i7 2600k.

The more I read about bulldozer's architecture and performance, the less I understand it

The article says that 256 bit AVX operations are split into two 128 bit AVX operations.It could be that for most FP operations it is only running a single 128 bit operation through each module, so you are not seeing the advantage in real life.

However,it looks like SSE performance is overall better than a Core i7 2600K and only AVX performance is bugged. However,not all applications use AVX(after all it was only really relevant after Windows 7 SP1) and there are useful gains if FMA and XOP are utilised too.

[scaryjim]

The more I read, the more despondent I get about bulldozer. When you can load it up heavily, it's one of the fastest processors available, but give it light loads and it barely gets out of the gate. Offloading flops to a "coprocessor" seems unlikely to help unless that coprocessor is inherently faster at processing a single operation, as the kinds of heavily parallelised workloads that we *know* make good GPGPU fodder are the kinds of workloads that BD is already competitive in compared to Intel's workstation / server chips.

Still, if they can get base clocks up in a respin, or at least improve turbo frequencies, and they can reduce cache latencies, we might see some of that decent int IPC coming to bear...

EDIT:

... It could be that for most FP operations it is only running a single 128 bit operation through each module ...

If I'm reading the SSE results right, BDs 4 FPUs are actually pushing the same throughput as Thuban's 6. To me that suggests that it is pushing 8 ops through the FPUs but at lower IPC - each BD 128bit FMAC is processing an instruction at 3/4 the speed of a Thuban FPU. of course, in lightly threaded loads (i.e. anything less than 7, in fact) Thuban is going to do float ops 33% faster than BD...

[Golden Dragoon]

I saw the deal on HUKD - was it something like £350 to £400 for an A8 based laptop?

My ex wanted a laptop at exactly the right time, got her to order an A8 based Lenovo direct from them, and 8gb of ram from ebuyer for £430

[CAT-THE-FIFTH]

If I'm reading the SSE results right, BDs 4 FPUs are actually pushing the same throughput as Thuban's 6. To me that suggests that it is pushing 8 ops through the FPUs but at lower IPC - each BD 128bit FMAC is processing an instruction at 3/4 the speed of a Thuban FPU. of course, in lightly threaded loads (i.e. anything less than 7, in fact) Thuban is going to do float ops 33% faster than BD...

You forget - if you are running only one operation through the FPU,performance will be better than a shared operation. However,the problem is whether a 2011 benchmark which is designed to test these kinds of things,will show similar scaling in applications. Its like with CineBench R10 and R11.5 - the Phenom II X6 did comparatively better when compared with contemporary Core i7 CPUs in the latter.

TBH,I am waiting to see reviews with newer versions of software to see if any real improvements will be seen.

If you look at the hotfix,games show the most improvement. However, it could indicate maybe the sharing is already happening on non-AVX FP operations,or maybe the improved scheduling means less conflicts for FP heavy calculations.

[CAT-THE-FIFTH]

The Bulldozer sucessor supposedly will have 4 memory channels:

http://www.xbitlabs.com/news/cpu/dis...evelopers.html

[Biscuit]

The Bulldozer sucessor supposedly will have 4 memory channels:

http://www.xbitlabs.com/news/cpu/dis...evelopers.html

It said that in one of your previous posts/links... is your brain full?

[CAT-THE-FIFTH]

It said that in one of your previous posts/links... is your brain full?

Thats what happens if you have a Moose as a helper!