22 nm Larrabee

It's not a representation of cumulative sales for the last 15 years. It's a representation of the hardware people had in Q1 2011. And the most interesting observation is that a 2005 ultra-low-end IGP is leading the chart...

It's going to take years before the majority of systems will be even capable of running OpenCL reliably at all. And during that time, CPUs won't stop evolving. AVX, AVX2 and eventually AVX-1024 all increase performance/Watt beyond Moore's Law, while GPUs are investing an ever increasing percentage of transistors for programmability and efficiency at ever more complex tasks. So nothing is turning this convergence around.

Put more bluntly; the CPU will continue to be the most reliable source of generic computing power. AMD doesn't stand a chance if it decides to attempt to take a different route. Computing and programmability convergence, Intel's process advantage, and unification of functionality and resource, all favor Intel's move toward a homogeneous architecture.

I'm not forgetting anything. This thread is mainly about Larrabee, desktop CPUs and GPUs, and close derivatives. The ultra-mobile market has different design goals and spending more silicon to achieve the lowest power consumption makes sense for the time being.

You mean like this: Runescape gameplay on intel gma 950 no lag?

Intel's current IGPs have become a lot faster since the GMA 950. So they will suffice for many years of casual games to come (and adequately run today's popular hardcore games too).

So really the value of having a more powerful IGP is very limited. Hardcore gamers continue to buy discrete graphics cards so they don't want a Fusion chip, and since future applications will have higher CPU demands too it's not wise to sacrifice CPU performance. Even some games already favor a fast quad-core CPU.

How do APUs help people who upgrade their GPU frequently?

What huge disparity? Tesla offers ~1 TFLOP theoretical peak performance, but only half that in practice (complex workloads are even worse). It also consumes 200+ Watt and takes 3 billion transistors. Haswell on the other hand is likely to offer 500 GFLOPS at 100 Watt and 1 billion transistors. Granted, NVIDIA will have a new architecture by the 2013 timeframe as well, but it will cost transistors and power to increase efficiency and programmability so I seriously doubt there will be a "huge" disparity. Last but not least, AVX-1024 will give the CPU another performance/Watt advantage to further close the gap, if there will even be a gap left at all...

 

ARM has a very, very long way to go to become part of a competitive HPC architecture. It absolutely won't taken Intel longer to implement AVX-1024 than it would take ARM and its partners to gain a foothold in the HPC market.

Why? Intel already has quite successful power-efficient 6-core CPUs at 32 nm. An 8-core Haswell capable of delivering 1 TFLOP at the same TDP doesn't seem much of a challenge at 22 nm + FinFET. That would still be a very small chip by today's GPU standards.

 

Why are matrix operations an NVIDIA weakness but not for AMD? And what will it cost to reach parity with CPUs?

 

I never said x86 is the center of the universe. I just think Intel is very close to producing a homogeneous power-efficient high-throughput CPU.

Why would we come up with new workloads only after CPU-GPU unification? I agree it will allow new workloads which are a complex mix of ILP / TLP / DLP, but I can't really imagine anything beyond that...

Do you have any supporting arguments? Or are you basing things solely on sentiment?

GPUs used to have a huge advantage in raw throughput due to exploiting TLP and DLP, while legacy CPUs only exploited ILP. Today's situation is very different. CPUs now feature multiple cores and multiple wide SIMD units, soon to be extended with FMA support. And AVX-1024 gives it the power efficiency of in-order processing. This leaves GPUs with no unique advantages.

Yes, supporting a scalar ISA costs some throughput density, but any APU still requires CPU cores so you have to compare homogeneous architectures against their entire heterogeneous counterpart. Note also that GCN will feature its own scalar ISA...

Why would the thermal and power limits of the CPU socket be any different than those of a graphics card?

What do you mean it did squat? Software renderers which don't use SSE are way slower.

 

A third option would be to increase ILP so they need fewer warps in flight and each of them gets a larger cut of the register file. Simply adopting GF104's superscalar issue could do the trick.

 

Throughput will continue to increase, even with a fully homogeneous architecture. What you're claiming is that won't suffice, and we'll get workloads which will require heterogeneous dedicated hardware again. Could you give me an example of a task which requires much more throughput than graphics but less programmability, and would be worth the dedicated silicon?

Compared to SSE, AVX2 increases the throughput fourfold, adds non-destructive instructions, and features gather. That's way more than "just a few more flops", and then some.

That's sentiment, not fact. GPUs are losing this advantage too. If Sandy Bridge had FMA support and no IGP, it would be less than 200 mm² and deliver 435 GFLOPS. GF116 is 238 mm² and delivers 691 GFLOPS.

That's merely a 33% advantage in computing density. But let's not forget that a CPU is still a CPU. The GPU is worthless on its own. And the convergence continues...

What's not to understand? You said CPUs are constrained by the thermal and power limits of the CPU socket. So I'm asking, what would keep them from increasing these thermal and power limits?

IBM's POWER7 has a 200 Watt TDP, and they fit two of these on a board. And GIGABYTE claims its 24-phase motherboard can deliver up to 1500 Watt to a single socket. Yes it's doubtful that can be sustained, but at least it shows there should be plenty of headroom for high-power CPUs, if there's a market for it. Once homogeneous CPUs replace IGPs and low-end GPUs, they can slowly scale things up when software starts to take advantage of the benefits of software rendering.

Ivy Bridge doesn't have AVX2 nor AVX-1024. IGPs will stay around as long as these haven't been implemented. Software rendering being slow is not a cosmic constant. They're currently limited to using 100 GFLOPS and emulating gather takes 3 uops per element. With four times higher throughput per core, non-destructive instructions, hardware gather, and more cores, software rendering is about to take a quantum leap.

 

That depends on the total memory access latency per warp. CUDA features prefetching, which also benefits from having fewer warps competing for cache space. So increasing ILP with superscalar issue should help in several ways.