WOF: What CES products are you most excited about?

[Wagan8r]

What about NVIDIA's Project Denver? I think this was the biggest announcement made at CES this year. It could possibly turn the CPU market on its head! I'm rather shocked to not see it in the list.

 

[Archean]

It could possibly turn the CPU market on its head!

I am sorry, but if I say this is just so generalist sort of a statement without grasping the underlying facts, I wouldn't be too far from reality.

I am not a fanboy of any one processor supplier, as I've used CPUs from AMD/Intel and graphics cards from nVidia and AMD for that matter, so I couldn't but agree when I read this:

"You want to come and party in our kitchen and rattle the pots and pans? I've got Sandy Bridge. Bring it on," Intel spokesman Dave Salvator said at the Consumer Electronics Show in Las Vegas.

Just consider these facts:

- People often forgets that Windows CE (not the most efficient of OS ever built by the way) ran on ARM platform for years.

- WP7 is running on ARM (so the news is not entirely new here either)

Now to the fact that Windows OS (PC) is to be ported to ARM (based) SoCs with its very complex kernel and overall performance demands of it:

- Salvator rightfully pointed out in the same interview that people tend to focus on ARM's meager power requirements; but please do remember to achieve such performance so one can be competitive with AMD/Intel offerings here they will need to rework the whole architecture, which will result in higher transistor count, larger die, hence, the end result will be increase in power utilization, and there is no escaping that.

- They will surely need to deliver multicore chips which must have good out-of-order execution engine, also I absolutely have no idea about e.g. branch prediction efficiency of ARM's In-Order (which limits performance in itself) architecture; without these they simply have no chance of competing with AMD/Intel's desktop processors.

I think what nVidia may eventually come out with is an reasonably efficient low end SoC for netbooks / mobile platforms / tablets / cellphones etc. Which should bring exceptionally good quality graphics to these markets forcing others to improve their own offerings as well. Which I think is an excellent development in itself.

Note: By the way MS itself holds license to design its own ARM processors; which shouldn't be ignored, as at some point MS may itself jump in the this whole game.

 

[dividebyzero]

I am sorry, but if I say this is just so generalist sort of a statement without grasping the underlying facts, I wouldn't be too far from reality....
Just consider these facts:

- They will surely need to deliver multicore chips which must have good out-of-order execution engine, also I absolutely have no idea about e.g. branch prediction efficiency of ARM's In-Order (which limits performance in itself) architecture; without these they simply have no chance of competing with AMD/Intel's desktop processors.

CPU's require OoO primarily because they are limited in tthe number of simultaneous computational threads they can deploy and the memory resources available.

GPU properties lead to a very different processor architecture from traditional CPUs. CPUs devote a lot of resources (primarily chip area) to make single streams of instructions run fast, including caching to hide memory latency and complex instruction-stream processing (pipelining, out-of-order execution and speculative execution). GPUs, on the other hand, use the chip area for hundreds of individual processing elements that execute a single instruction stream on many data elements simultaneously. Memory latency is hidden by very fast context switching; when a memory fetch is issued while processing one subset of data elements, that subset is set aside in favor of another subset that is not waiting on a memory reference.

Now bear in mind that the Project Denver outline is to marry (using GF110 as example) 512 stream processors (essentially a very basic CPU core with linear functionality)-CUDA cores in nvidiaspeak- with it's inherently faster (wider bus/greater bandwidth/lower latency) memory resouces, to what will in all likelihood be a 64-bit ARM based architecture- presumeably-as is presently the case- to allocate resources where program code branches becomes more divergent than would be suited by the CUDA cores SIMD nature.

"You want to come and party in our kitchen and rattle the pots and pans? I've got Sandy Bridge. Bring it on," Intel spokesman Dave Salvator said at the Consumer Electronics Show in Las Vegas.

Not overly unusual to hear a PR guy bigging up their company I would have thought. Dirk Meyer and Jen Hsun Huang are both adept at equally bombastic rhetoric when someone shoves a recording device in front of their faces. CES is, after all a PR driven event and Intel have product to push.

 

[Andruku]

ET3D said:
Personally I've been waiting for the AMD Fusion CPU's (or APU's, as AMD likes to call them) for a while, and I'm still waiting for real arrival and reviews. Hopefully soon.
.

Same here

 

[Archean]

CPU's require OoO primarily because they are limited in tthe number of simultaneous computational threads they can deploy and the memory resources available.

Yes the limitations are there, but OOO/OOE are there to help overcome these:

out-of-order execution (OoOE or OOE) is a paradigm used in most high-performance microprocessors to make use of instruction cycles that would otherwise be wasted by a certain type of costly delay. In this paradigm, a processor executes instructions in an order governed by the availability of input data, rather than by their original order in a program. In doing so, the processor can avoid being idle while data is retrieved for the next instruction in a program, processing instead the next instructions which is able to run immediately.

The benefit of OoO processing grows as the instruction pipeline deepens and the speed difference between main memory (or cache memory) and the processor widens.

Now bear in mind that the Project Denver outline is to marry (using GF110 as example) 512 stream processors (essentially a very basic CPU core with linear functionality)-CUDA cores in nvidiaspeak- with it's inherently faster (wider bus/greater bandwidth/lower latency) memory resouces, to what will in all likelihood be a 64-bit ARM based architecture- presumeably-as is presently the case- to allocate resources where program code branches becomes more divergent than would be suited by the CUDA cores SIMD nature.

Here in lies a major bone of contention, you will need to re-program/recompile/optimize everything for the new architecture; and I don’t see such a huge x86 base moving everything to this platform any time soon. What is the incentive outside the graphic intensive applications to do this? Absolutely none I’d say. Business world doesn’t move everything to new architectures or OS on the whims of people like us who jump on everything new and exciting. Just to add little support to this argument, we have dozens of computers still running Windows XP on P4s/Pentium Ds at work, and they are still doing what is needed pretty efficiently; hell our dispatch clerk is still using the good old 4L printer from the ancient times. . Another example, I used to impart advice with regard to IT issues to a large company here, and it is still running about 10,000+ PCs on Windows XP (P4s/Conroe based), using SAP etc. along with many other custom made solutions, I don't see such entities even consider such radical change, let alone jumping the ship at all.

Also consider this fact, Intel was able to improve graphic performance upward of 2x with Sandy Bridge, have nVidia or for that matter AMD ever able to have done that with any of their newer generation graphic cards? Considering the implications of Moore’s Law, and significant resources and technological insight Intel and AMD are in far better position to outrun nVidia here as long as they are able to double CPU/graphic performance of on-die solutions. Additionally I haven’t had time to read all the architectural improvements in Sandy Bridge, but what early little bits I read, I remember I roughly calculated that L3 cache latency was down by 38% (either due to improved pre-fetch unit / improved L3 Latency). I’d love to know improvements in GPUs at similar pace, I doubt it will happen on such consistent basis though.

And I haven’t yet touched the issue of power utilization of such higher end graphic cores, where nVidia will need to radically redesign their offering to reduce thermal output to be competitive at the least. Therefore, in the short term, nVidia won’t change anything here, but do they have enough resources to develop this into something competitive in desktop/notebook/server segments in the longer run is another issue.

However, as I said in my previous post in the mobile arena the picture will change because of more competition, hence resulting in superior hardware and better overall experience for end users; and whether we like it or not the world is heading to the mobile future.

 

[dividebyzero]

Yes the limitations are there, but OOO/OOE are there to help overcome these:

out-of-order execution (OoOE or OOE) is a paradigm used...[/]

If existing CPU architecture is able to utilise OoO to it's fullest why are we moving toward 6, 8, 10 and 12 core consumer CPU's ? Why are 2P and 4P workstations (24+ cores) de rigueur, with the next generation promising to more than double that number in short order (see Knight's Corner)
All I see in the quote above is an explanation of how out of order execution works-which I'm reasonably familiar with. If present instructions sets are sufficient to keep cores "working" then future 8, 10 and 12 core CPU's aren't required. If they aren't required, why are they being developed?

Here in lies a major bone of contention, you will need to re-program/recompile/optimize everything for the new architecture; and I don’t see such a huge x86 base moving everything to this platform any time soon.

True...to a degree. Firstly I think the same argument was used with the 64 bit extension for x86. It's been a long haul but I think people are starting to see the light.
People as a whole are resistant (and sometimes fear) change. A simple look at some of the TS forum thread postings make that painfully obvious.

What is the incentive outside the graphic intensive applications to do this? Absolutely none I’d say.

Given that nvidia is in the graphics business I'd say that you're posing a loaded question......but I'll give it a shot anyway (bearing in mind that software USUALLY follows hardware)
Mathematica
HALMD
Then of course you have the usually familiar SETI@Home, F@H, BOINC distributed computing initiatives

I would think that with 1. A fully fledged OS in place, 2. Nvidia's record of software development, 3. Probable incentives (SDK's etc) offered by Nvidia, MS and every probable vendor offering ARM-based products-Apple excluded, that either porting existing software, or making newer revisions of that software Win8 ARM compatible shouldn't be a big hurdle.
Question: Do you see all (or the majority of) software manufacturers turning their back on the ultra-portable market and leaving MS and Nvidia affiliates to have 100% of the software market. Would you also expect MS to have Win8 support ARM processors but not MS's own software stack (MS Office for example) ?

Business world doesn’t move everything to new architectures or OS on the whims of people like us who jump on everything new and exciting. Just to add little support to this argument, we have dozens of computers still running Windows XP on P4s/Pentium Ds at work, and they are still doing what is needed pretty efficiently; hell our dispatch clerk is still using the good old 4L printer from the ancient times.

Just as well the tech world isn't totally reliant on business intransigence. Although I would hazard a guess that small/medium business is not the main aim of either the software or hardware- hardly surprising since businesses generally upgrade only when they need/have to and typically businesses are the last to migrate to a new OS. If business were the prime motivator we wouldn't have 64-bit computing, Core2Duo/Quad, discrete graphics, SATA interfaces, DisplayPort, DVI USB 3.0, Sandy Bridge, Phenom II, Eyefinity/nvidia Surround, 1080p + monitors, tablet/notebook PC's, smartphones, motherboard tweaking options etc, etc...and we would all be using single core CPU's hooked up to 865PE chipset boards viewing the efforts of our labour on CRT 1024x768 monitors. Thrilling stuff.

Also consider this fact, Intel was able to improve graphic performance upward of 2x with Sandy Bridge, have nVidia or for that matter AMD ever able to have done that with any of their newer generation graphic cards?

This would be the same Sandy Bridge that uses Nvidia graphics IP then? ( a major contibuting factor in Intel settling with Nvidia for $1.5bn. Sandy Bridge, Ivy Bridge and Haswell will all use the Nvidia IP.
I suppose I shouldn't answer a question with a question though....so how about G80 ?

Considering the implications of Moore’s Law, and significant resources and technological insight Intel and AMD are in far better position to outrun nVidia here as long as they are able to double CPU/graphic performance of on-die solutions.?

That's a big "if". At present Intel is at 32nm. GPU's are at 40nm. Maybe we could revisit this thread when GPU's are on 28nm later in the year.
Firstly I'd say that the discussion isn't about "graphics performance", it's about computational ability. On-die GPU performance is presently in the "OK for casual gamer" category- you won't be doing any 1080p gaming of note with Sandy Bridge, nor Ivy Bridge, nor Llano. The first of these isn't even DX11 capable, and those that follow most likely will have the feature as a semi-useless bullet point in the specs.
If Intel have "technological insight" then it might pay to see that their next processor strategy is a 50-core GPGPU. If your process technology has a finite end in the short-medium term then parallel computing is the only alternative.

Additionally I haven’t had time to read all the architectural improvements in Sandy Bridge, but what early little bits I read, I remember I roughly calculated that L3 cache latency was down by 38% (either due to improved pre-fetch unit / improved L3 Latency). I’d love to know improvements in GPUs at similar pace, I doubt it will happen on such consistent basis though.

Apples and oranges. CPU's are dependant upon cache latency- GPU's aren't for the simple reason that the GPU has a MUCH higher degree of simultaneous multi-threaded performance. The reason that GPU's don't have an L3 cache.

And I haven’t yet touched the issue of power utilization of such higher end graphic cores, where nVidia will need to radically redesign their offering to reduce thermal output to be competitive at the least..

You realise that to increase the on-die graphics performance on Ivy Bridge, Llano derivatives etc. requires adding more shader pipelines? Adding shaders and/or shader frequency requires more power. Again, quite how this has morphed into a graphics discussion is beyond me. iirc GF108, 106 and 104 suffer no real drawbacks in either thermal or power usage characteristics if you take relative performance/watt into account. or are we comparing Fermi with HD2000/3000 ? Seems like comparing a Lambo LP640 against a Toyota Prius using fuel economy as the only metric.
For arguments sake, lets take into account that Fermi (and it's derivatives, AND Northern Islands) were originally destined for TSMC's ill-fated 32nm process. Porting the architecture to 40nm made the designs of both AMD and Nvidia somewhat more power hungry than they would have been otherwise. Note that both GTX570/580 and HD6950/6970 are rated at ~200w or more, both designs feature power limiters.

 

[Archean]

As you rightfully pointed out later in your post that software follows the hardware evolution, hence, the more threaded an application becomes the more benefit you get from extra cores. Therefore, multicore CPUs or even GPUs can be a very handy workhorse indeed, and these will continue to evolve resulting in greater number of CPU cores or CUDA cores in the respective hardware.

True...to a degree. Firstly I think the same argument was used with the 64 bit extension for x86. It's been a long haul but I think people are starting to see the light.

Yes there are significant improvements in the situation from few years ago but again, the pace is rather slow, probably that is why MS is still planning another x86 Windows OS, which doesn’t make sense to me …..

I must admit I forgot about Folding@home and Seti so thanks for bringing up this important factor.

IGP Performance: I agree unless IGPs are able to reach performance level of upward of 70$ graphic cards or ideally around 100$ (at the given time) they won’t be much good for gaming but for most of the other things they are more than enough.

I somewhere read (a rumor I'd say) that Intel is planning to put 1GB on-die dedicated memory ......... have you read or heard any such thing?

Considering the 40nm issues (also I remember nVidia struggled with GT200 GPUs 65/55nm path) and their reliance on outside foundries I wouldn’t rule out such problems in the future hampering their progress (read reliability as supplier). Intel’s full node + custom design concept worked perfectly so far + manufacturing capabilities gives them some advantage here as well, hence, when they move to 22nm later this year they will be ahead in this area.

I never doubted that MS will not port everything in their portfolio to ARM, and Office will be the first thing which will get ported any way IMHO.

This would be the same Sandy Bridge that uses Nvidia graphics IP then? ( a major contibuting factor in Intel settling with Nvidia for $1.5bn. Sandy Bridge, Ivy Bridge and Haswell will all use the Nvidia IP.

I suspect that is a reason why Intel never really killed Larrabee, they want to come up with something which can successfully compete, and save them all these licensing issues/cost with nVidia. However, the catch is, if (again I’d say HUGE if) Intel (or AMD) can come up x86 based high performance GPGPU in the given time window, this whole ARM+nVidia debate may turn out to be storm in a teacup. But if we do have a 3rd relatively competitive/successful player offering economical solutions in the market we the users will be the biggest winners.

 

[dividebyzero]

IGP Performance: I agree unless IGPs are able to reach performance level of upward of 70$ graphic cards or ideally around 100$ (at the given time) they won’t be much good for gaming but for most of the other things they are more than enough.

If you remove gaming from the equation then all you need (and are left with) is hardware decode, multi-monitor support and HD playback, which is what the HD3000 already provides. In this respect it doesn't really matter if the GPU is on-die or soldered to the board. On-die integration with the CPU is always going to lag behind a standalone solution due to size constraints of adding larger shader blocks (and a dedicated tessellator if DX11 is required).

I somewhere read (a rumor I'd say) that Intel is planning to put 1GB on-die dedicated memory ......... have you read or heard any such thing?.

Stacked memory. Intel have been toying with it for around four years I think. Recently Charlie D repackaged the story with the emphasis on Ivy Bridge. The principle drawbacks are increased die size, a big power draw/thermal envelope penalty and low speed suitable for a high end mobile or desktop only part. As CD mentions, bandwidth wont be a problem on a 512-bit bus, even if the GPU architecture/graphics drivers themselves are less than ideal

Considering the 40nm issues (also I remember nVidia struggled with GT200 GPUs 65/55nm path) and their reliance on outside foundries I wouldn’t rule out such problems in the future hampering their progress (read reliability as supplier)

GT200, GF100 and GF110 to an extent are more a product of Jen Hsun's adherence to a large monolithic die stategy and Nvidia's vision battling the laws of Physics. AMD's GPU's are made by the same foundry (TSMC) on the same process (40nm) as Nvidia's big chips. The difference lies in feature set and what markets the resultant GPU's sell in. AMD are almost exclusively gaming/HTPC orientated. Nvidia adds compute (GPGPU) into the mix- 72-bit ECC memory, wider memory bus widths, high double precision requirement (largely disabled in gaming cards) etc.

Intel’s full node + custom design concept worked perfectly so far + manufacturing capabilities gives them some advantage here as well, hence, when they move to 22nm later this year they will be ahead in this area.

I don't remember arguing against Intel's ability to execute. What I do see is the processor moving towards a greater degree of parallelization as an alternative to increased core speed. Out of Order Execution is parallelization by another name, and the more threads/cores the processor has at it's disposal the greater the data throughput- hence Intels use of AVX now and their longer range plans for multi-core/multi-pipeline processors in future.

I suspect that is a reason why Intel never really killed Larrabee, they want to come up with something which can successfully compete, and save them all these licensing issues/cost with nVidia..

The reason I think Larrabee is still alive is that the idea behind the hardware is sound. What Intel lacks is GPGPU experience in hardware and drivers. There's a reason that there are only two main player is the discrete GPU market, and only one player in workstation/compute driver arena.

However, the catch is, if (again I’d say HUGE if) Intel (or AMD) can come up x86 based high performance GPGPU in the given time window, this whole ARM+nVidia debate may turn out to be storm in a teacup...

See above re Intel and Nvidia. Nvidia are by accounts are looking at a 2013 timeframe using the Maxwell GPU (20nm process, ~14-16 GFlops of double precision per watt target performance) as the GPU arch in question- which both looks significantly faster than the Intel solution and light years ahead in drivers and compiler -remember that Nvidia supports both CUDA and OpenCL.
Which leaves us with AMD
At this point in time AMD has shown approximately 0% interest in GPGPU. Their sole contribution to the space seems to be the OpenCL Stream SDK- so basically if software devs aren't doing the work then it's not happening. The rest comes down to Intels x86-64 GPGPU versus Nvidia/Microsoft et al. 64-bit ARM with integrated stream processing, and much as the x86 CPU and ARM exist side-by-side now I really don't think it will be much different in future. Heavily threaded relatively simple pipelines have already shown that they don't always require the bloat that x86 CPU's use.

several readers have asked for elaboration around the assertion that ARM CPUs could disrupt X86:

The threat is definitely there. ARM architectures have been on a steady path to disrupt x86 CPUs for over a decade. Starting at the low-end with cell phones, ARM-based CPUs have somewhat quietly moved up-market into smartphones, and now netbooks. As they continue to improve in line with Moore’s Law, they are settling into 2 GHz while remaining far lower power and lower cost than even Atom. The tipping point is likely to begin in earnest during 2010.”

Two top questions have been (1) can ARM have a place in HPC, and (2) can it really disrupt X86? The short answers: yes and yes.

And a second opinion

 

[Archean]

If you remove gaming from the equation then all you need (and are left with) is hardware decode, multi-monitor support and HD playback, which is what the HD3000 already provides.

I ignored it for the desktop arena (where lies interest of most of the enthusiasts), however 'casual' gamers would be ok with Sandy Bridge for now (talking about notebooks here); Ivy Bridge may improve the picture further, as the resolutions aren't that higher. If somebody wants a gaming notebook, they in all probability will be power users hence will go for the right choices from the start.

I do see is the processor moving towards a greater degree of parallelization as an alternative to increased core speed

No question about it, no matter which route the future may take us to, parallelism is here to stay.

With regard to LRB: I have reasons to believe Intel is still improving the hardware, and bring the software side along with it; as along side nVidia, Intel has had reasonably sound software side performance of the things until they hit the stumbling block with LRB.

“As the battle heats up, ARM will likely have the most appeal with customers who value low power and low cost, as opposed to raw horsepower.” Which make perfect sense for any such emerging architecture from nVidia (or others).

I think probably MS and nVidia may have the end goal of being able to provide ‘total experience’ for their customers from hardware to software, in certain segments of the market.

However, when I raised the point about large x86 software base which in all probability won't be moving to any other architecture in a hurry, I was thinking about what happened with IA64. As the current shift to x64 base is being 'facilitated' by the 'backward' compatibility of all existing CPUs (and OS as well san some limitations), which won't be the case with ARM, and herein lies the biggest stumbling block for nVidia about its plans.

 

[dividebyzero]

That presupposes that Nvidia is planning on moving into the x86 space, no?

I don't see Project Denver as a series of desktop SKU's nor high end mobile - at least not for the foreseeable future.
It is obviously being aimed at ARM + small shader block SoC (Basic OS function with native apps, HD playback, hardware accelerated video encode/decode, HTML5) in the ultra portable/ULV space, and Multi-ARM + complex shader block for server/HPC (open source OS enviroment)

Nvidia's momentum is geared towards the tablet/ULV market (Tegra), pro graphics (Quadro) and HPC/Server (Tesla). Both Tegra and Tesla operate in an enviroment where x86 is a non-factor -that will not change- MS and Google are effectively making the statement that x86 is too unwieldy for the ultra-portable market, while x86 has never been a factor in server/HPC/Medical Imaging/Research/Analytical etc....which leaves Nvidia's cash cow, Quadro- which will likely remain the premiere professional graphics enviroment brand for some considerable time (OptiX/SceniX/CompleX/pro-graphics apps/plug-ins). Considering Nvidia has 90% of the market I don't think they'll mind too much being tied to x86-64 there.

Again, I don't think it's either RISC or x86 in the future- it's more a case of using the right ISA for the right enviroment/hardware fit. Intel naturally are going to push x86 everywhere they can- it's the foundation that Intel is built on after all. But can you honestly see an Intel (or AMD for that matter) having a x86 CPU+GPU with a 200mV power draw ?

In the matter of what most people would consider the conventional computing space (consumer and business/OEM desktops/laptop) I think you're right to say that inertia (and a degree of prejudice/apprehension in all likelihood) will keep x86 pre-eminent. It's also a case of the best tool for the job - x86 still has a lot of gas left in the tank (it's just not a catchall solution and never has been). But whereas that market used to be the only option for personal computing, that is longer the case.