Ars: AMD Fusion won't kill the GPU, but what if it kills the CPU?

[richardginn]

http://arstechnica.com/business/news...t-kill-gpu.ars

[James]

Interesting take on the future of CPU/GPU dynamics and which will be more important in the future.

[taggart6]

7 years ago I sat in a Nvidia company presentation where the Nvida rep (one of the R&D leads) made a comment that with how large GPUs are, it would only be a matter of time before CPU+GPU ship was released.


Kinda funny, IMHO, if AMD/ATI beats them to it.

[Nater]

Well, AMD beat them to it, not ATi. That's a big difference, especially considering that nVidia has no x86 license.

The thing is, AMD is at least a generation behind nVidia when it comes to GPU compute and have been for sometime. Not just hardware, but the complete ecosystem.

Oh, and not to mention that Fusion chips still aren't out and the 'real' Fusion (with a serious, full power x86 unit) has been delayed into the middle of 2011.

[ImaNihilist]

We've been talking about this for years. Reality is that it's cheaper and ends up making better business sense to keep them separate. Nvidia's IGP's have gotten really good. You can play DX 9 games with high settings on dirt cheap IGPs. Vendors can mix different IGP/dedicated solutions across their models easily, and cheaply.

In order for something like this to ever really take off it's going to have to be a lot better and a lot cheaper. You'll need a large player like Dell, HP or Apple to back this across a number of different models before they can reach any kind of scale that makes sense.

[richardginn]

We've been talking about this for years. Reality is that it's cheaper and ends up making better business sense to keep them separate. Nvidia's IGP's have gotten really good. You can play DX 9 games with high settings on dirt cheap IGPs. Vendors can mix different IGP/dedicated solutions across their models easily, and cheaply.

In order for something like this to ever really take off it's going to have to be a lot better and a lot cheaper. You'll need a large player like Dell, HP or Apple to back this across a number of different models before they can reach any kind of scale that makes sense.

I do not see the CPU going anywhere, buttt.....

You might see IGP on a desktop CPU powerful enough to replace a 75-100 buck video card in the next few years.

[Nater]

I think the best place for a product like Fusion are applications extremely sensitive to heat and power. Obviously that means laptops, but I think an architecture like this would be quite useful for HPC applications. There are, however, a couple of problems for AMD when it comes to using it's Llano-type processors in supercomputers.

1. They aren't even building versions with multiple HT links to my knowledge. I've never really heard of a serious supercomputer being built out of single socket nodes. When you're trying to cram in as many cores as space and power will allow, using multiprocessor nodes has an obvious advantage.

2. nVidia eats their lunch when it comes to using GPUs as HPC accelerators. Not only are nVidia's GPUs better performing from a DPFLOPS/watt basis, nVidia also actually has a software framework. AMD continually talks about this or that (Brook+, CTM, now OpenCL) but never really gets anything done.

Those two obstacles are going to stop someone like Cray, SGI, or IBM from using thousands upon thousands of these chips in HPC systems. Well, that and the fact that SGI has pretty much gone with intel, IBM is now split between POWER7 and intel, and it's very likely Cray will also go with intel once it's series of supercomputers developed from Red Storm (XT3, XT4, XT5, XT6) reach the end of their life.

[ImaNihilist]

I do not see the CPU going anywhere, buttt.....

You might see IGP on a desktop CPU powerful enough to replace a 75-100 buck video card in the next few years.

$75-$100 GPUs have already been replaced by IGPs integrated onto the motherboard, which offer OEMs a lot more flexibility in terms of product line.

[nukefault]

Surely you jest?

9400M gets smoked by $40-50ish 9500GT and HD4650. You can now grab a 9800GT, HD4850, HD5750, or (on recent Newegg coupon deals) an HD4870 or 9800GTX+ for $100 - all of which are in a completely different league than the 9400M.

Oh yeah, and the best CPU you can run it with is a C2Q.

That leaves us with the HD4290 as our high end IGP. It generally costs $30-40 more than mobos with the somewhat slower HD4200 (although you do get some other benefits from the upgrade) and even the 4290 is slower than 9400M.

It's better than 9400M in that you can use it with Phenom 2 instead of C2Q, but it's slower than a current gen entry level discrete card.

On what planet is either 9400M or HD4290 even close to a replacement for any of the cards I listed?

[nukefault]

Unless of course you're referring to rumors that mainstream Fusion will be carrying HD5650-equivalent graphics on-die... But of course that's not a mobo GPU, and even then a 5650 doesn't replace a 4850 or a 5750 or a 9800GT.

Intel and AMD are both betting against you (note that Sandy brings graphics onto the CPU die as well).

[Nater]

Unless he means intel's integrated "GPUs". Those are the most common graphics adapters out there.

If he means stuff like nVidia's integrated solutions or AMD's, I'd like to see evidence.

[nukefault]

How are Intel GPUs possibly a replacement for an HD 5450, much less an HD 4850? That makes even less sense.

Ima stated that mainboard IGPs are replacing $75-100 PCIe cards. I can't think of any context in which that hold true. The pinnacles of mobo IGP are the aforementioned HD 4290 and 9400M, which are the equivalent last-gen entry-level $30ish PCIe cards at best.

Even Sandy Bridge's impressive graphics (per Anandtech's preview) are only pushing into the $50 space. And, of course, the clear future of IGPs is in CPUs a la Sandy and Fusion. Mobo IGPs will be a distant memory a year from now.

[Nater]

Yeah, I really don't know. I don't pay attention at all to OEM desktop computers, I pay a glancing bit to workstations but they always have some form of discrete card.

Then you have laptops, which are an entirely different ballgame. The differences between on-board and discrete are pretty slim there.

[ImaNihilist]

Ima stated that mainboard IGPs are replacing $75-100 PCIe cards. I can't think of any context in which that hold true.

Notebooks (and netbooks). You know, the largest, fastest growing segment of mainstream computing.

Most gamers don't play at 1920x1200 with all settings set to ULTRA and 8xAA/16xAF. Less than 20% play at 1080p or higher.

IGPs can now run most games at 720p resolutions with medium settings and 2xAA/4xAF. 60% of gamers play at resolutions 1280x1024 and below. NVIDIA's IGPs are absolutely KILLING it in the mobile market right now. A few years ago playing a game on an IGP was kind of a joke. Now it's completely legitimate.

You can play Portal on 13"-15" notebook, at its native resolution, with medium/high settings and basic AA/AF, at 45+FPS. That's remarkable, and a HUGE leap from where we were just a few years ago with garbage like the Intel GMA 950. That's probably around the time that AMD thought this Fusion gimmick was going to be a big win—back around the time the Intel 915G/945G chipset was the most popular thing since... maybe since ever. Seriously, the 945G may be one of Intel's time best-selling platform chipsets. That thing is everywhere.

The 9400M has changed the entire landscape for IGPs and mobile graphics.

I'm a bit biased though, since I'm not even really interested in the big-gun cards that play games at 4K resolutions with full AA/AF. I don't have a 4K screen, and neither do most people. $500 for a video card? Meh. I'm interested in when we'll have $500 boxes with IGPs that can play Portal 2 at 1080p. And really... we are like a year, maybe two, from that. NVIDIA and Intel will own that space. No one is going to want this AMD Fusion garbage.

[nukefault]

Seems like a pretty early dismissal given that

1. AMD's approach is highly similar to Intel's with Sandy (is nobody going to want Sandy either?) Intel just beat them to it with a heat start on process tech and engineer power. The fact that Intel is following AMD's footsteps (in realizing where tech is going) should be good evidence that Fusion is actually a good idea.

2. There are massive technical benefits to having the IGP on-die rather than on the chipset. Far lower latency to both CPU and memory, L3 cache access as per Sandy and probably Fusion, no need for an ultra-wide system bus to manage the IGP, being manufactured on a state-of-the-art process rather than on one 2-3 generations old as chipsets tend to be, saving power via advanced gating and other technologies, more efficient thermals via advanced manufacturing process and having a single hot area that needs cooling, etc.

The efficient thermals are proven with Sandy's combination of an IGP that absolutely destroys the 9400M and a 95W TDP (remembering that Intel's TDP estimates are a theoretical worst-case scenario and realistic usage even under max load tends to be substantially lower).

Given all that, and given your own arguments about the importance of IGP performance, I don't see how you can possibly trash on Fusion. It's an amazing way to increase IGP performance, it'll reduce chipset and bus complexity and lead to smaller and more efficient form factors, and it's one more step toward mainstream SoCs and ITX-sized computers with ATX-sized performance.

Even if you're skeptical of AMD for some reason, are you really going to say nobody will be interested in Sandy either? It's better than Core2+9400M in every way you can name, and even if nVidia somehow gets a bus license from Intel, their chipset IGPs will be at a huge technical disadvantage being made on older manufacturing processes and not having L3 cache access or ultra-low latency to CPU and memory.

I completely agree though about $500 video cards being ridiculous. I'll never own another desktop or another laptop with discrete graphics (my last laptop had a 2600XT). Heat, battery life, and price all heavily favor IGPs, and with both Intel and AMD putting IGPs on-die you can bet that's the way the market is going to go.

So why pick on Fusion when Sandy is so similar?

You might be disappointed with Llano being a K10 derivative (no, it's not a K8 regardless of what Fuad quotes in that conference call... AMD was referring to the fact that K10 is a K8 derivative, not that they're actually going to use the K8 core - in all likelihood it'll be a process shrunk quad core version of Thuban with the usual tweaks) but you already said you're not a desktop guy. Ontario might be too low end for either of us (although there's a chance I'll get it if Sandy is too expensive and/or it rivals Sandy's IGP and the P320's CPU) but the concept is gold and I think Bulldozer+Fusion will probably be a solid price/performance candidate in the laptop space.