Server Memory - what's the diff?

[scaryjim]

Partly for my own interest, and partly for a mate at work who's speccing a new HPC box, what's the real world difference between the various options for server memory? ECC vs non-Ecc I can just about handle (i.e. ECC handles errors in memory), but what do buffered / registered add to & take away from RAM?

EDIT: argh, new forums bug! Apparently the link from Hardware > Memory points to the corsair forum! I'll get a friendly admin to move this thread...

[Agent]

This is simplified, but it gives you a 'buffer' between the RAM and what's controlling the RAM (normally the memory controller).

Once the buffer becomes full, it then passes the information along. It allows for more control on what's done with the RAM and when. The advantage? much finer electrical current control, which in turn can cut down issues with overloading them.

There is really no situation you'll need them in a normal home system, unless the motherboard specially requires it (very unlikely, but HP have done some weird things in the past.)

The downside is that they are slower as the buffer has to fill up before it can do anything.

Think of it this way: If you have a pond at the bottom of your garden you need to fill, you might use a hose pipe to spray the water into the pond, or you might fill lots of buckets and manually empty them in the pond. The latter offers tighter control over the water supply to the pond, but takes longer as you need to fill each bucket before you empty it.

In short: Rarely if ever used in home systems, but useful for servers / mission critical machines.

Another point to consider is that buffering comes from the days of SIMMS (and older). Electrical current is *much* more tightly controlled these days in modern systems. Buffering is far less of an issue to overcome electrical stress in a modern system.

[jim]

Sadly no friendly admins were available, but the village idiot has sorted it

[scaryjim]

Thanks snoots, I'd rather have an obliging and helpful idiot than nothing

Agent: nice explanation, thanks! This is for a GPGPU build based on SNB-E Xeons, so high throughput but I'd guess most of the burden will be on the GPUs rather than the subsystem. Sounds like buffered would be the wrong way to go for this, and unbuffered ECC would be a safer bet then. Thanks for the input

[Butuz]

Out of interest what GPUs has he chosen?

Butuz

[scaryjim]

He'd gone for 7970s until I threw a spanner in the works and started chatting to him about the GTX680 I suspect he'll stick with the 7970s though, simply because they're more of a known quantity and he'd already started prepping for AMDs developement environment.

[Butuz]

Yes it depends if he's going to use CUDA at all really. Also Nvidia's 600 series are MUCH weaker than AMD's consumer units on FP64 compute so again depends if he is using FP64 or FP32.

Butuz

[Agent]

I wonder if they'll separate the Geforce and Testla in terms of major hardware differences in the future?

[Butuz]

Yes is the short answer to that one.

Theyre losing money on Quadro / Tesla because Geforce has previously been far too good at compute. I think the 600 series is just the start, the ranges will diverge even further from now on. AMD will probably follow suit over the next few years.

Butuz

[Agent]

Can't really blame them. We have a few machines at work, and like you say, GeForce is just a no brainier for most things. I don't even know if we have any (new) Quadro cards since our last update.

[scaryjim]

Thanks Butuz, very useful extra info there Something I wouldn't have known, since GPGPU isn't even vaguely my area of expertise

[Funkstar]

Theyre losing money on Quadro / Tesla because Geforce has previously been far too good at compute. I think the 600 series is just the start, the ranges will diverge even further from now on. AMD will probably follow suit over the next few years.

I don't really know anything about this area, but you've got me curious. How will they differentiate them and make the Quadro/Tesla cards more attractive? I assumed the pro cards were much better at GPGPU tasks than the consumer cards, but obviously I was mistaken.

[Butuz]

Quadro / Tesla cards will remain the same. They are already good.

Geforce cards on the other hand will get worse @ compute in order to make the Quadro / Tesla cards look more attractive to the compute masses.

For example in double precision math (what many scientists would be after) the GTX580's FP64 performance was capped at 1/8th of its single precision FP32 performance FP64 tasks would run 1/8th slower than FP32 tasks. On the GTX680 FP64 performance is capped at 1/24th of its FP32 performance. So even through FP32 performance is good as it's such a powerful card, its FP64 performance is abysmal, 25% slower than even a GTX560ti.

So you can see advanced compute consumers who want double precision maths would have previously used a GTX480 or 580 as they were powerful, won't be using a 680. I they want better than a 580 they'll have to go Quadro / Tesla.

So now that Nvidia have got many many people from small labs to supercomputers addicted to using CUDA and OpenCL on £400 Geforce class cards - BOOOOM now if you want to upgrade you gotta buy a Tesla now that'll be £2000+ please.

Clever.

Butuz

[Funkstar]

Got you. Thanks for clarifying

[Agent]

The other thing to consider is that the Quadro / Tesla range come with much more tested drivers with support to back it up. You'll get issues fixed at the driver level, if they are needed, much faster in our experience if you're a user of one of those ranges. It's in their best interests to make dam sure that their 'professional' partners have stable drivers.

[scaryjim]

Presumably Quadro/Tesla will support ECC memory too, which (afaik) consumer GeForce don't. Again a serious consideration if you're doing a lot of scientific calculation: you don't want your entire simulation thrown out by a 1 bit data error...