Maybe for the average office worker that just does word and excel and connects to a data base or two. But there will always be a market for workstation class desktops and gaming rigs.
I’m not sure I totally follow your post, but there are a few misconceptions. The first is Intel as a manufacturing company primarily. While that is a big strength, TSMC etc. sells fab space, something Intel has never done until recently. I know a senior professor who spent a sabbatical in Rio Ranchos. I asked him about how the fab people saw themselves, and if they felt good about doing such great manufacturing work. He said that they felt like second class citizens compared to design. When I was in design my colleagues felt that way, but it was interesting fab guys did too.
I assume you are not saying that Itanic wiped out other big server chips. That chip has been a disaster from the start. I wonder how many billions of bucks Intel has lost on it.
While desktops will survive as a niche market, the rise of the cloud is good for big chips, not bad. While the consumer is going to buy stuff with relatively low powered processors for cloud use, the machines in the cloud are going to need powerful processors, not ARM cores.
Depends on how much you overclock. You might get a part that just squeaked by, and has suffered some speed loss, or you might get a part that is way faster than spec. And speed tests don’t have 100% coverage. You might exercise a path not exercised by the test.
As for reliability, the important thing is to have less than N chips fail during the first Y years of life. That may translate into 80% not failing in Z years - but I’ve never seen really old parts get returned. Processors are very close to the most reliable part of a computer system. Fans and power supplies and disks almost always go first.
Most cloud tasks are trivially split over multiple cores. And for server farms electricity and heat are critical. ARM is aiming for the server market with lower power chips than Intel, you can thus fit more cores in the same space for less power.
Qualcomm has a 24 Core ARM CPU aimed at this market.
Here’s another one with 48 cores from Cavium:
Qualcomm is already working on an ARM server CPU with 64 cores.
http://www.fudzilla.com/news/processors/37939-qualcomm-has-64-core-arm-server-processor
This was a mantra I remember from a long time ago. People closer to the heart would probably have a better idea.
Indeed. Itanic was a very sad disaster. I feel sad because I had a lot of hope for it. I liked the ISA. For a while we even have a pretty big Itanic based SGI box. I actually still have a die set in plastic in my brief case. Makes for a good conversation piece. Itanic seems to be the cause of lots of disasters. The wiping out of the Alpha high amongst them. Whether Alpha had the legs to keep ahead, hard to say, but it was a lovely ISA, and the manner in which it was sold down the river something of a crime.
Time will tell. Not that long ago I would have agreed. For some things this will remain true. One group I work with does some pretty heavy lifting (many hundreds of cores for tens to hundreds of hours for a run). But a huge amount of other stuff is remarkably lightweight. I see many cloud cores that are mostly idle. Web servers that see little traffic, simple databases holding simple datasets - or more often metadata for big flat data. And power continues to drive thinking. I can see a big role for lightweight low power cloud boxes. A big part of the future of the cloud is much more about data than compute. (And as a long term HPC weenie, that causes me some sadness.) This year’s big thing is containers, things like Docker. This is driving a move away from VMs, and I can see a need for lightweight, low power, boxes that host simple Docker images.
Cloud is going to evolve for quite a while yet. I very much doubt a lot of people have really sat down and realised just how big the changes are going to be.
ISTM that you (MEM) missed the rest of Voyager’s sentence. And did again in his/Francis’ later posts.
You the end user can’t overclock a chip. You can only overclock a system.
What I think **Voyager **meant, as restated by others as well, is that the chip has to test in a factory tester at 2.75MHz to be able to run reliably at 2.5Mhz on a real board in a real system under real workloads with real cooling.
Whether you can bump the system clock and/or voltage & retain reliability is unknowable. They *can *statistically predict the shape of the overclock/overvolt failure curve for hundreds of chips in hundreds of systems.
What they *can’t *predict is how your individual chip will respond in your individual system. There’s probably some headroom. But there may not be. It may be you’ve got a CPU chip that could in fact go 20% faster, all the way to 3.0 MHz. But that same board has a weak bus controller chip that is already on the verge of failing at rated speed and can’t take even 1% over. And if that’s the case, your system will not successfully overclock at all.
Got it, thanks.
Speaking of MS, I’ve been wondering how they would remain relevant with tablet/phone encroachment and I think their continuum/display dock device is a pretty good strategy. (You connect a phone to the device and it drives a normal monitor and keyboard running phone apps in a desktop-like mode).
A substantial amount of business computing use these days is either office or web-based apps. Web based apps don’t require much client computing power and MS owns office so they can make sure it runs well on phone/tablet.
One other thing to consider is if your part is the fastest they sell. I assumed it was, because otherwise someone would buy the faster part and not overclock a slower one. In this case your chances of the overclocking working is higher, because a faster part will be sold at a slower bin. If there are faster parts sold, any faster part would get sold at the higher price higher bin, and your chances of the overclocking working are much lower.
You might get lucky in that yield at the higher bin would be greater than demand, so fast parts get sold as slower ones, but I wouldn’t bet on it.
Not likely. Overclocking is predominantly used (by consumers) for cost savings. The top-tier products are expensive; a tier or two down and the products are much less expensive.
I would say that the main thing that overclockers take advantage of is the margin between the worst-case conditions and the conditions that a particular user can guarantee. GPUs are rated to work in a 95 C oven; if you can guarantee better conditions than that, then you have some overclocking margin.
The next thing is the tradeoff between voltage vs. power and lifetime. Bump the voltage, TDP goes beyond the manufacturer rating. Chip lifetime also goes down. The overclocker probably doesn’t care.
Statistical variation is probably the least concern. Sure, a few extreme overclockers will look out for the best samples, but the majority are happy with the improvement they get from the above items, which benefit even chips on the edge of a speed bin.
There’s also a reliability tradeoff. I don’t know what Intel’s standards are, but suppose it’s one error in 10[sup]20[/sup] cycles. An overclocker might be happy with one error in 10[sup]17[/sup] cycles instead.
I was of the impression (and it is only that) that the overclockers were mostly those seeking the bleeding edge. Gamers mostly. I have come across some HPC users, but again, bleeding edge, rather than cost savings. Everything about overclocking adds cost. The motherboards are a premium price, the cooling systems can get just plain silly. Top bin parts are a premium too. Interesting that Intel first marketed Skylake with an overclockable version.
Gamers probably are more tolerant of errors, in that it isn’t mission critical. They probably get peeved at system crashes or lock-ups - but regard that as part of the game. Reduced lifetime similarly.
I’d argue that the real bleeding edge types–and they certainly exist–are largely not gamers, or at least gaming is not their goal. These people are mainly interested in overclocking as an end in itself; to get the highest 3DMark score, for instance. They frequently only care about having enough stability to finish a single benchmark run. They’re a bit like the computer equivalent of drag racing. Some of them go so far as to use liquid nitrogen or other exotic coolants.
On the other hand, there’s a big segment of gamers just looking to get a bit more value for money. They’re probably using basic air cooling, and although they might spend a bit more on a motherboard, power supply, heatsink, etc, they don’t go crazy. They’re happy if they get 10-20% extra performance out of their hardware, which is usually doable without too much effort. Since they’re looking for value, they go for second or third tier products, which tend to have greater !/$ to start with.
I’m personally in the latter category. Sure, I can afford the high end products. But if I can get the equivalent of a $1000 CPU with a $300 CPU and a few hours of my time, it’s worth it to me. I’m building my own system anyway because I have particular needs that commercial vendors don’t meet, and at that point overclocking the HW a tad is a fairly trivial thing.
Some want to buy a mid-range chip and see if they can get it close to the performance of a high-range chip. Others want to buy a high-range chip and push it as high as it’ll go.
You mention Skylake and you’re quite right; Intel seemed to aim at people who were not at all price-sensitive and wanted the highest performance possible. On the other hand, the i5-4690k is quite popular with overclockers even though it’s a mere i5 and a mere 4000. I presume it’s for consumers who, like me, want high performance but are only ready to drop half a month’s rent on an upgrade.
Also, it’s more expensive to overclock a CPU than a GPU. With a GPU, no need for a premium motherboard or silly overcooling. Although that might be coming soon enough; GPUs are starting to get watercooling.
On the topic of desktop disappearing: I can see a time when a desktop is used in much the same conditions as a generator: When you’re off the grid. Although I would think that in those circumstances, you’d use a laptop.
It really depends on how quickly connections get faster, bandwidth gets cheaper and whether player inputs can be sent to the server, game logic processed, graphics of about 4k/60fps processed, encrypted, sent to the client and decrypted with a lower latency than the maximum the player will put up with. If that can be done, sure, I can see desktops being a niche. It would be a lot more efficient for players to share a common pool of graphics processing power than each buy their own GPU.
Adding a bit on overclocking & value for money: The i5-4690k costs about 55% as much as an i7-6700k but will very likely give you far more than 55% of its gaming performance when both are overclocked. With an overclocked i5-4690k, it will be many years before one’s gaming is CPU-bottlenecked.
I bought a medium-high GPU and went for a model of that GPU that was halfway between the lowest and highest prices, overclocked & overvolted it and now I’ve got a 970 that likely has about as much performance as a stock 980 which would have cost me 200USD more. I’m not the type who tries to establish world records with his computer (which still uses a Phenom II x4 955 from 5 years ago).
Actually, now that I think about it, I hadn’t taken the rest of the sentence as a limitation because I presumed that the factory tester likely presented a more trying environment than my system.
If my system is bottlenecked elsewhere, then indeed, even if the chip can do more, that potential will not be taped.
It is however useful to know if a chip is likely to have more potential because that allows me to identify where there is likely to be untapped potential. It also allows me to identify the most probable bottlenecks in my system and address those first.
Since I do not have the expertise and materiel to test all components and subcomponents of my individual system, I have to be satisfied with probability assessments.
If there’s an 80% chance I can get more performance and I make a purchasing decision based on that but turn out to be in the 20% that can’t get more; Oh well, it was the best decision I could make based on imperfect information. I’ve regretted many decisions but never a decision that was the best I could make under the circumstances and turned out to be wrong.
When I ask about potentials and probabilities, it is to help establish the expected utility (N utils x %) of the different options. I also ask because these things are interesting in themselves.
I don’t buy it simply because local CPU / GPU compute is also getting faster and cheaper and the speed of light is a real limitation when you’re trying to do stuff like this. Sony’s PS Now service offers streaming of compressed PS3 games at 720p. They don’t look anywhere near as good as playing locally on a real PS3 and every reviewer has mentioned intermittent lag as a big issue. Meanwhile a $350 US PS4 gives you smooth 60 fps 1080p play. Or you can buy a second hand PS3 for $100 and used games for $20 each, still giving you a better experience than PS Now does. By the time you can stream 1080p reliably a console that can do smooth 4K at 60 fps will only cost $350.
Consumer internet doesn’t offer any guaranteed bandwidth levels, 4K streaming on netflix can work around that by buffering ahead. You can’t do that for streaming games from a cloud server and even if theres a hiccup only every 3-4 minutes thats enough to ruin your gaming experience.
Perf/Watt:
Perf/Watt is the major bottleneck? Does that mean that when it comes to big announcements of future generation of GPU, perf/Watt is the number we should pay attention to if we want to assess generational improvement rather than relying on CEO math?
Turning it up to 11:
While I understand that increasing perf/Watt is the long term way to increase performance over several generations, I’m surprised by how low power even the biggest GPUs are. What if someone wants more performance and doesn’t care about power consumption? Even two R9 290X will only gobble up 500 Watts*. Where’s the half kiloWatt or even full kiloWatt GPU?
Sure, a user could SLI/CX several GPUs or a manufacturer could stick two GPUs together but that’s severely limiting in terms of VRAM unless DX12 is used. Also, SLI/CX usually scale badly.
Now, you might say, who would want to foot the power bill for that?
Me, for example. The marginal cost of a kW/h for me is 4.1 US cents**. Other people have enough money that it wouldn’t make a big difference to them and others are ready to spend quite a bit on their favorite hobby. E.g.: A guy here bought a pair of Titan Xs which must have cost him 2000-2400 USD; I think he’d readily accept more performance even if his power bill went up.
Myself, if I didn’t care about Nvidia features and really wanted 4K/60fps gaming, I would have got a 1600Watts PSU and 4 R9 390s. It’s too bad that such a setup would need 8GB of VRAM 4 times instead of once.
**At the current exchange rates.
There’s currently no practical way to package a die that can dissipate 1KW without exceeding the maximum silicon operating temperature. The thermal resistance between the die and the package is too great. Maybe with a diamond package and chilled water.
Definitely. It is possible to struggle along as a company if you have low perf–you just target a lower segment. But if you have poor perf/watt, you have no hope.
What beowulff said is true, but there’s another factor. The primary determinant of perf/watt (for a fixed process node) is the architecture. The architecture sticks around for two or three years and is used on every product segment. They’re too expensive to not do that.
So while you and other enthusiasts might not care about power too much, for everyone else it’s of prime importance. You need beefier power regulators. More balls on the die to supply the power. A bigger system PSU. Auxiliary power connectors and associated circuitry like current sensing. Since all the power comes out as heat, you need a bigger and better cooling system. That’s extra weight, so the PCB might need to be thicker. The system as a whole needs better cooling. On mobile it affects battery life. It’s just a massive extra expense.
Since the same architecture is used for everyone and the enthusiasts don’t get their own, it has to be focused on perf/watt. Enthusiasts can use SLI and the like if they really don’t care about power (though if it’s working well, SLI doesn’t change the perf/watt factor: dual GPU is twice the perf and twice the watts).
There is another similar metric, perf/mm[sup]2[/sup] (these are square millimeters of silicon). It was more important in the past, when GPUs were a lot less powerful than they are. It might be more important in the future. It’s possible to imagine a GPU that has modestly lower perf/watt but excellent perf/mm[sup]2[/sup]. For instance, you could clock a GPU like a CPU, at 3-4 GHz, if you added lots of additional buffering and pipelining and such. You would pay some die cost for that but the clocks more than make up for it, perf-wise. That strategy doesn’t make sense now but there is a point on the manufacturing cost curve where it would.
I suspect you could do it with a high speed water jet impingement system. Just shoot a jet of water at the bare die at whatever structural limit it can withstand. Pretty sophisticated system, though.
Water jet seems to be the only technology that will get you 1kw/cm[sup]2[/sup].
The Cray X1 used a spray of flourinert to directly cool the die. But that is a much gentler mechanism than water jet. Just about everything gets insane if you are talking those pwer densities. Even getting the current on and off the chip is going to be a major difficulty.
(I really liked the X1 - a Mips with serious vector bolts ons. It got ridiculous performance.)