Slashdot Mirror
AMD, Intel, and NVIDIA Over the Next 10 Years
GhostX9 writes "Alan Dang from Tom's Hardware has just written a speculative op-ed on the future of AMD, Intel, and NVIDIA in the next decade. They talk about the strengths of AMD's combined GPU and CPU teams, Intel's experience with VLIW architectures, and NVIDIA's software lead in the GPU computing world." What do you think it will take to stay on top over the next ten years? Or, will we have a newcomer that usurps the throne and puts everyone else out of business?
I predict wrong predictions.
Damping absorbs vibrations. Dampening is caused by moisture.
With greater personal power, we won't have Microsoft dictating what 3D features we can have. With individuals become supercomputers, these three companies will be out of business. However, personal survivability and power will be sufficient that former employees will be fine.
What?
"I like to lick butts!" by MobileTatsu-NJG (#32700246) (Score:5, Informative)
All three will be marginalized by the ARM onslaught. Within 10 years, smartphone will be the personal computing device, AMD and Intel processors will power the cloud.
In short: make your time.
Mod me down, my New Earth Global Warmingist friends!
The use of futurism has been thoroughly discredited.
I predicted that years ago.
On the contrary, I think the CPU will go the way of the coprocessor. The humble Atom may be enough CPU power for most people these days, but you can never have enough GPU power... at least not until your po-- I mean, games, are photorealistic in real time.
A slashdotter who didn't build his own computer is like a Jedi who didn't build his own lightsaber.
Well... There's 2 ways of looking at it. Either the GPU and the CPU will be merged into one beast, or there will be further segregating of tasks. In terms of price, what's more efficient: Having 1 chip that can do everything (Picture a 128 core CPU, that has different cores optimized for different tasks. So 32 cores optimized for floating point processes, 32 for vector processes and 64 for "generic computing") or having multiple chips that are each fully optimized for their task. Actually, now that I think about it, I'd probably say both. Economy computers would be based off the "Generic" cpu, whereas performance computers and servers would have add-in modules that let you tailor the hardware more towards the task at hand. So the motherboard could get an additional 8 sockets (similar to DIMM sockets) that would let you plug in different modules. So if you need to do graphics heavy processing (video games, movie rendering, etc) you'd add 8 GPU modules to the motherboard. If you needed floating point capacity, you'd add 8 FPU modules... Etc... The advantage of doing it that way over the current PCIe method, is that you get to skip the southbridge (So these modules would have full speed access to system memory, hardware and each other). Of course, there are a lot of hurdles to implementing such a thing...
I am not an engineer, these are just thoughts that rolled off my head...
If a man isn't willing to take some risk for his opinions, either his opinions are no good or he's no good
You can always spot a sensationalist post when part of it predicts or asks who will go out of business. Or what thing will disappear.
For example, in his post, ScuttleMonkey asks this:
NNote, the post is a good one - Im not being critical. But change in the tech industry rarely result in big companies going out of business - if they do, it takes a long time. I think sun is the canonical example here. It took a long time for them to die - even after many, many missteps. Sun faded away not because of competition or some gaming changing technology, but simply because they made bad (or some would say awful) decisions. Same for Transmeta.
People have been predicting the death of this or that forever. As you might imaging, my favorite one is predicting Microsofts death. Thats being going on for a long, long time. The last I checked, we are still quite healthy.
Personally, I dont see Intel, AMD, or NVIDIA ding any time soon. Note, AMD came close this last year, but they have had several near death experiences over the years. (I worked there for several years...).
Intel, AMD and NVIDIA fundamental business is turning sand into money. This was a famous quote by Jerry Sanders the found of AMD. Im paraphrasing, but it was long the idea at AMD that it didnt matter what came out of the fabs as long as the fabs were busy. Even though AMD and NVIDIA no longer own fabs, this is still their business model (more or less).
I think its interesting how a couple of posters have talked about ARM - remember, AMD and NVIDIA can jump on the ARM bandwagon at any time. Intel already is an ARM licensee. Like AMD, they are in the business of turning sand into money - they can and will change their manufacturing mix to maintain profitability.
I also dont see the GPU going away either. GPUs are freakishly good at what they do. By good - I mean better than anything else. Intel flubbed it badly with Larabee. A general purpose core simply isnt going to do what very carefully designed silicon can do. This has been proven time and time again.
Domain specific silicon will always be cheaper, better performing and more power efficient in most areas than a general purpose gizmo. Note, this doesnt mean I dislike general purpose gizmos (like processors) - I believe that the best system designs have a mix of both - suited to the purpose at hand.
-Foredecker
Jibe!
If he is getting paid well, he doesn't give two shits what kind of clap you are doing.
Nerd rage is the funniest rage.
I disagree. Floating-point coprocessors basically just added some FP instructions to a regular single-threaded CPU. There was no parallelism; they just removed the need to do slow floating-point calculations using integer math.
However, GPUs, while they mainly do floating-point calculations, are essentially vector processors, and do calculations in parallel. They can easily benefit from increased size and parallelism: the more parallel processing capability a GPU has, the more realistic it can make graphical applications (i.e. games). And with all the GPGPU applications coming about (where you use GPUs to perform general-purpose (i.e., not graphics) calculations), there's no end to the amount of parallel computational power that can be used. The only limits are cost and energy.
So if someone tried to fold the GPU into the processor, just how much capability would they put there? And what if it's not enough? Intel has already tried to do this, and it hasn't killed the GPU at all. Not everyone plays bleeding-edge 3D games; a lot of people just want a low-powered computer for surfing the web, and maybe looking at Google Earth. An Intel CPU with a built-in low-power GPU works fine for that, but it won't be very useful for playing Crysis unless you think 5 fps is good. People who want to play photo-realistic games, however, are going to want more power than that. And oil exploration companies and protein-folding researchers are going to want even more.
GPUs aren't going anywhere, any time soon. Lots of systems already have eliminated them in favor of integrated solutions, but these aren't systems you're going to play the latest games on. For those markets, NVIDIA is still doing just fine.
At some point, the GPU goes on the CPU chip, and gets faster as a result.
Maybe.
GPUs need enormous bandwidth to memory, and can usefully use several different types of memory with separate data paths. The frame buffer, texture memory, geometry memory, and program memory are all being accessed by different parts of the GPU. Making all that traffic go through the CPU's path to memory, which is already the bottleneck with current CPUs, doesn't help performance.
A single chip solution improves CPU to GPU bandwidth, but that's not usually the worst bottleneck.
What actually determines the solution turns out to be issues like how many pins you can effectively have on a chip.
I am still waiting for OpenCL to get traction. All this CUDA and StreamSDK stuff is tied to a particular company's hardware. I think there is a need for a free software implementation of OpenCL with different backends (NVidia-GPU, AMD-GPU, x86-CPU). Software developers will have great difficulties to support GPUs as long as there is no hardware-independent standard.
Yes - it takes about two years (or more) to go from a white board to first silicon. Until I worked at Microsoft, I worked at hardware and silicon companies. But remember, the competition to Intel, AMD and NVIDIA will be other silicon companies - not software companies. The new compitetion will have the same constraints. This is also a small industry - its very difficult to do someting both major and new in secret. When I was at AMD, we knew about Transmeta's plans when they were still in stealth mode. It wasn't because of anything nefarious - the community is small and leaky. -Foredecker
Jibe!
You just said people will never want photorealistic rendered porn. You have got to be the worst predictor ever.
Have you ever even met a person?
What I find interesting is the overall lack of game-changing progress when it comes to non-3d-or-hd-video-related tasks. In March 2000, i.e. ten years ago, top of the line CPU would be a Pentium III coppermine, potentially topping out around 1 Ghz. I could put Windows XP on one of those (with enough RAM) and do most office / browsing tasks about as fast as I could with today's top of the line CPU. Heck, it would probably handle Win7 okay. Contrast the period 2000-2010 with the period 1990-2000. In 1990 you would be looking at a 25mhz 486DX.
Phone as a Terminal
The best solution would not be to run apps on the phone at all. It would be to get always on bandwidth from a PC at home to your phone that was fast enough to do remote desktop at a speed where you couldn't tell that you were working remotely. Once we have that kind of bandwidth, the phones are basically done. The phone as a terminal. With this configuration, you get:
* Massive upgradeability on the phone since to make your phone faster, you just upgrade the PC in your home.
* Far greater battery life, as once the phone is a good terminal, adding more processing power to the PC will add power, but since that part is plugged into the wall, it won't drain your battery at all.
* Losing your phone does not effect any of your data.
* Replacing your phone is simpler.
* You can access the same application from a desktop, TV, or the phone, and there is no reason the interface cannot change for each.
* Better utilization of processing power, since people will end up with a home server anyway, for running their home media servers, security systems, home automation, etc...
* Cheaper. It will always be more expensive to build these things smaller, so putting it in a PC makes it cheaper.
* Faster to market. It takes time to shrink electronics.
* Possible functionality that is impossible on the phone. We are getting to the point where we may be limited by physics on how small transistor can become. This means that the amount of processing power that would be supplied to the phone as a terminal may be impossible to have in a handheld device.
Your post is based on several assumptions that make no sense to me as a student of human nature, and an engineer.
1. 1080p is current technology. Even if we assume we will not have hologram visual output within the near future, there will still be some new technology that the powers that be will sell to the masses. It may be an incremental improvement, but it will still be enough to drive the markets.
1a. As long as it's new and shiny, there will always be someone to buy it.
2. Consoles use GPUs and CPUs the same as PCs do. There is a longer update cycle in place, but whenever each cycle ticks they adopt all the new technology that has been developed during the lifetime of a console. As such, it makes sense for the console makers to encourage such development.
3. Intel would have to shut down all of their operations to let nVidia claim the workstation market. Like it or not, Intel still makes pretty hefty CPUs, owns the workstation market, and has more disposable cash, and a bigger engineering staff than any other chip maker. The embedded market has even more competition for its crown, so I will not go there. The supercomputer market, while good for satisfying the nerd bragging rights quota, is not know for being an amazing source of profit.
4. The AMD vs Intel battle for the mid-range market is actually something I can see coming to pass. I would not be too surprised if this market gets a third player as the line between computation devices becomes blurred.
5. ARM is not the only company in the world that can make a low power chips. Worst case, ARM has a few years of dominance before the other guys catch up. Also, as the article pointed out, integrated CPU/GPU has several obvious advantages over discrete CPU + discrete GPU.
In all, while I am not ready to make my own predictions, yours could use a bit more analysis and tweaking.
What are you talking about?
The current version of OSX can run apps as far back as 2001. Apple does not provide any official way of running apps older than that.
The current 64 bit versions of Windows can run apps as far back as 1993. Microsoft provides an official way of running even older apps (XP Mode). 32 bit Windows can often run 16 bit apps without the emulator.
Microsoft has lots to fault them for, but their record on backwards compatibility is WAY better than Apple's.
Sun's VLIW architecture (MAJC) was more interesting than Intel's. The point of VLIW is the same as that of RISC; take more stuff that isn't directly connected to executing instructions off the CPU and make the compiler do it. EPIC missed the point and tried to do VLIW + a load of extra stuff on the chip. Sun did proper VLIW and took it to its logical conclusion with a JIT doing absolutely everything (branch prediction, instruction scheduling, even dynamic partitioning for threads). Unfortunately, it came from an era when Sun was still in the Everything Should Use Java mindset. Something like MAJC with something like LLVM could be insanely fast, but LLVM is still a few years away from being ready for that kind of use and no one is developing successors to MAJC so it probably won't happen for a long time, if at all.
I am TheRaven on Soylent News