I'm glad AMD is competitive again. Yes, Intel just surpassed them, but at twice the price. Plus, I suspect that competition pushes Intel into releasing better chips quicker.
I'm not really part of the "prosumer" market (what a word!), since I don't have a burning need for that much power, but it helps move prices lower in the middle and bottom of the performance range when the top end improves.
Same here. I haven't upgraded my home PC for almost 5 years now, even though I do some digging from time to time to see if there's something worth buying in a middle tier. So far, it was simply not worth upgrading, I hope things will change soon. Heck, even GPU bought back in 2013 is still doing well (I remember times when 1 year old GPU was considered obsolete, even in a middle tier).
> That depends very much on what you use your machine for.
In generally it actually doesn't. The AMD Ryzen 7 1800x is selling for less than half the price of an Intel i7-6900k ($420 Vs $1010), and in benchmarks the Ryzen system beats the i7 or practically matches the performance.
Getting the same performance for 40% of the cost is something that does not depend on your use case.
Getting the same performance for 40% of the cost is something that does not depend on your use case.
Sure, but in many cases the practical difference between an Intel i7-6900k and for example an Intel i3-7XXX (or a 4 year old i5/7) is basically negligible, and the i3 is less than half the price of the Ryzen
I just upgraded the gaming PC from an i7 2600 (ancient) to a Ryzen 7 1700. So far, video encoding is significantly faster with all 16 logical CPUs busy (but a GPU based program would be stupid fast). I'm having an issue where I get occasional lockups when gaming - updating the BIOS hasn't helped. I expect the game makers will need to ship updated code. One thing I found - the Ryzen is very sensitive to memory speed. Buy 2666 and don't overclock it much - if at all.
The last time AMD was competitive, they delivered really good value for money. Yet they were only available in the "enthusiast" market segment as Intel were able to give all OEM strong incentives to not buy AMD.
Is there any chance that they are not having all OEMs by the balls this time too?
it's not even that, but in the OEM and server market, there's more than delivering 90% of perforamnce for 50% of the price - there's reliability, driver support and all other costs which means that a higher Intel CPU price might only translate to <5% higher TCO or something like that, in which case I'd have gone with Intel myself (no one got fired for buying IBM/Intel)
I beg to differ. The last time AMD was competitive, my employer was buying AMD server hardware from HPaq for their VMware environment. Granted, all our desktops/laptops were c2d or early core i-whatever.
If you look around, you can find AM4 platform in prebuilts from some of the biggest PC OEMs. Intel is still the majority of the lineup, of course, but APU systems have given AMD leverage here that they didn't have the last time they had competitive performance. It makes the OEMs offer the platform for an energy efficient/SFF build option, while also giving leeway to do an upgrade down the road.
I wonder, what makes intel chips more expensive? Is it a matter of money sunk into research? Because it seems that in terms of FLOPS per $, AMD always wins, and I have a hard time understanding that.
Other way around: AMD has to be cheaper to compete with Intel, which until now basically could charge what they wanted.
And this specific example here is clearl in the high-end enthusiast space, where whoever has the best can charge the highest price and the target market will pay it.
> I wonder, what makes intel chips more expensive?
The fact that if you want a chip that powerful, there is only Intel. An example in GPUs are the NVIDIA Quaddro cards. If your frame needs more than 16GB memory to render, there is no other card available. So they sell the same chips packaged as a Quaddro at 5x the price of a Titan.
I believe the biggest reason the workstation/enterprise GPU lines (AMD’s too) are priced so much higher than the gaming/consumer lines is because the “certifications” for CAD[1] and 3D graphics[2] software cost a king’s ransom plus require expensive testing and validation of every individual card before it’s sold.
> How does, e.g., POWER8 compare? (And POWER9, soon?)
POWER8 is not well suited for HPC workloads as their vector instructions (VMX/AltiVec) are only 128-bit wide vs Intel's latest Skylake Xeon CPUs have vector instructions (AVX512) that are 512-bit wide.
For non HPC workloads (e.g. integer arithmetic) the POWER8 CPUs were somehow competitive with Intel's CPUs when they were first released in 2015 but today AMD and Intel's CPUs are clearly faster.
From what I can tell CPU enthusiasts were not impressed by the POWER8 CPU architecture.
It is not so much a "chip that powerful" but a "chip that correct". Gaming GPUs are allowed errors because the rendered images are known to be visual output only, whereas an engineering GPU may be used for hard analysis, which cannot afford any rounding or other classes of the acceptable errors gaming GPUs are allowed.
It is just complete nonsense. In fact deep learning/AI chips used for neural networks usually have reduced floating point precision because they don't need it.
We don't need it because errors in any particular value aren't important - the calculations assume an amount of inherent error greater than that produced by the hardware.
Do you have a source for those claims? It seems highly unlikely to me that certain types of cpu's or gpu's are allowed errors to cut costs. Rounding mechanics are complicated, but I'd say a bug is a bug. As mentioned in a comment by jackmott in this thread, "engineering purposes" also often actively reduce precision for increased performance.
Double precision floating point and the addition of ECC memory.
In the case of nVidia the quadro cards are the only ones sold as manufactured by the company and are incredibly high binned.
Lots of people on this thread are conflating a few things such as ML, but I work in an industry where we deal a lot with GPUs and its certainly a fact that flipped bits are significantly more common than many think. If you do any work that requires the output to be nonvolatile then you can run into issues.
Perhaps he is referring to ECC memory - high end pro model NVIDIA cards have it, and consumer cards don’t. Here is a paper showing that there are memory errors on the consumer non-ecc cards.
But for any particular version of "that powerful", the exclusivity doesn't last long. The burning need for hardware that didn't exist a couple years ago and will be mainstream in a couple more years isn't all that strong.
Costs of production rarely determine prices directly for any product - beyond setting the baseline from which the margin measures on. Market dynamics tend to do the rest (competition, market position, etc). Intel plays a role in it but not always.
There may be some that are down to the wire per unit cost (typically the cheaper ones) but I have a feeling Intel operates with plenty of margin on most of their higher end products. And plenty of them are the same chip repackaged.
For me, AMD tends to do better with multithreaded/parallel applications, but that tradeoff is at the expense of single core speeds. So mostly it depends on what you are using it for. I built a 64 core opteron server for doing physics calculations and it was probably my favorite build ever, so I can't wait to try the new AMD server lineup, this time I'll only need two CPUs to get 64 cores!
Intel operates at 60% gross profit margin, AMD operates primarily in the red.
In the past 15 quarters AMD had a single quarter in 2014 where it didn't suffered an operational loss (when the bulk of this generation of console royalties were paid out most likely)
Tensorflow isn’t supported yet.
The ROCm libraries are optimized for AMD hardware explicitly this isn’t democratizing anything it just locks you into a different vendor other than NVIDIA.
If anything AMD should follow NVIDIA in supporting OpenACC which is the way to go as far as true probability goes.
No, ROCm is an amalgamation of a lot of things and it's been very inconsistently represented, you are thinking of HIP which in theory should that but in practice it's a tool to more easily convert CUDA applications to ROCm, since while you can in theory use NVCC as a backend for HIP the design directive for HIP make it nonviable.
In all effectiveness ROCm is AMD's version of CUDA it includes it's own run time and driver stack which effectively replaced the OpenCL frontend for AMD GPUs, basically it gives AMD the ability to do with their GPUs what CUDA allowed NVIDIA to do and it's not to be dependent on a camel (OpenCL).
This is what ROCm is: https://rocm.github.io/images/ROCm_Stack.png or at least this is what AMD is pushing it to be, currently a lot of things are still unfinished or non-existent.
The representation of it being open is misleading, while more of the source code is available than for CUDA, ROCm isn't vendor agnostic on any tangible level.
In 2-3 years once it's somewhat in a production ready state it will lock you into a single vendor just like CUDA does.
> Yes, Intel just surpassed them, but at twice the price.
Which to me says that Intel is not competitive. If you can only release a slightly better product at twice the price, then you failed. Of course Intel's products have some branding inertia behind them, too, but that will be short-lived if the products don't become more competitive in the near future.
On the other hand, you could also say that the i7-7740X[1], at a third TR's price, will (significantly) outperform TR in most real-world usage, and even for video-editing/compiling-Chromium type tasks, the $700 difference would probably be better spent on a better graphics card, more RAM or, since we're getting into overkill territory, a nicer monitor. And, in theory, I suspect that a dual-7740X build might match or beat TR in multithreaded, too.
My point is it's the category most enthusiasts would be happiest with, even if they think otherwise.
>>>Is that possible?
Looks like I was mistaken, I thought the Sky/Kaby Lake-X chips used the same socket as the Xeons. So no.
>>>And I really doubt you'll get 8 cores to beat 16.
If it were possible, it would be 8 cores operating at a 50% higher frequency (with substantially more room for overclocking, too) and that support AVX-512 instructions. And you get quickly-diminishing returns from throwing more threads at a program (just mathematically, even if it's embarrassingly parallel and CPU-bound, each time you double the threadcount you get half the improvement of the previous doubling).
With 3 glaring factual errors across two posts, I'm going to go out on a limb and suggest to you that the average enthusiast purchasing Threadripper is in fact more aware of the trade offs between the products available on the market than you are.
For price of 7980XE you can buy Epic and have even more performance, if you care about multithreading.
The only thing that Intel surpassed them is single-thread performance. It's important even on workstation, IMO, that's the main reason I'm thinking to stick with Intel for now, everything else AMD made better.
Intel must be pretty desperate, according to the review article -
* Intel doubled the price of a threadripper 1950x for giving you a 8% increase on multithread performance (measure by Cinebench R15 scores).
* when Intel could't afford to match AMD high end CPU's 60 PCIE lanes, they invented the term "platform PCIE lanes" to include those from chipsets.
* when Intel couldn't match AMD's TDP, they came up with some different TDP definition, as a result the 165w TDP Intel 7980xe draws 15w more than a 180w threadripper 1950x.
Intel must have a pretty good propaganda department known as "marketing"?
for those who believe Intel is responding to AMD's challenge - no, it is not, according to leaks, Intel is about to release i3-7360x, which is going to be 1.25% faster (yeah, 1.25% faster, no typo, source linked) than the previous generation. when you can fully utilize your well established marketing department to fool consumers to buy your overpriced slow processor with far less cores and PCIE lanes, why bother giving those customers anything better?
I haven't followed the market for a while but a few years ago it was more or less established that the top Intel CPU had about 10% better performance and 2x the price of the nest one. Has it changed much?
The new chip is smaller (and I'd presume more energy efficient) while being slightly faster. Those characteristics are what a good chunk of the server market craves (more blade server density per rack at the same power draw).
> when Intel couldn't match AMD's TDP, they came up with some different TDP definition, as a result the 165w TDP Intel 7980xe draws 15w more than a 180w threadripper 1950x.
So far it was AMD which had a weird way to calculate the TDP, and it was them that went over while with Intel TDP pretty much equaled max power draw. That Intel now goes over is clearly a sign that they are stressing their boundaries and the of their architecture, but it's not really a big distinction against AMD.
Intel's TDP measures power consumption/heat production in the intended application. The same silicon has a TDP of 5W, 15W or 35W depending on whether it's going in an ultralight laptop, pro laptop or desktop.
All modern Intel CPUs will draw more than their rated TDP if you run them hard.
And these are high end 'prosumer' parts. Running the CPU at peak 80-100% load for a significant amount of time should be considered the intended application.
What does 80-100% even mean, though? The CPU cores are merely one component of the CPU 'package'. The GPU on most Intel parts consumes way more power than the CPU cores when running flat-out.
Even just within the CPU cores, different ops consume vastly different amount of energy while still 'maxing out' utilisation. And we haven't even considered accelerator hardware within the CPU which is totally independent of the CPU cores but can still consume a large amount of energy (e.g. H.264, AES).
Intel no doubt calculate a maximum power consumption figure where every bit of hardware on the CPU is consuming as much as it can. It'd be utterly meaningless -- probably on the order of 100W for a 2-4 core laptop part or 400W for a server part.
You can design a computer to handle those figures, but you don't want to. So TDP reflects intended usage rather than peak demand.
If you're a 'prosumer', you're presumably smart enough to size your power supply and cooling appropriately. TDP is not the number you're looking for.
Ah yes fair point, my personal bias is showing a bit. I would consider numerics (AVX512 or AVX2) to be the main interest, and it is often what the type of workstations I'm familiar with need. It also tends to consume the most power when run at full load by a significant margin.
the 80-100% is bit of a vague number, I was just kinda referring to the cpu utilization as reported by the OS. For numerics I would consider a realistic full load 75%+ pipeline fill time on the AVX codepaths. With a lot of compute you can get quite close to 100% quite quickly if your software is highly optimized (and uses enough compute to hide the memory loads). TDP should just be 100% pipelined load of the highest consuming codepath.
But my general point is still that these benchmarks should not exceed the TDP for these types of parts. For everyday use these loads might not be realistic. But for a lot of people looking to spend $2000(!) on a single CPU, these benchmarks are indicative of real life usage.
> All modern Intel CPUs will draw more than their rated TDP if you run them hard.
If hard does not include AVX I think that is not correct. All locked i5s and i7s stayed under or very near to their TDP. And also the K units, before overclocking. https://www.anandtech.com/show/10968/the-intel-core-i7-7700k... as example. However, I noticed that the reviewer mentioned that wasn't true for the 4790K and 6700K.
I should mention that I'm talking about the desktop versions here. The laptop parts are a different story.
> according to leaks, Intel is about to release i3-7360x, which is going to be 1.25% faster (yeah, 1.25% faster, no typo, source linked) than the previous generation
That CPU is built for overclocking competitions, don't know why so many people are upset over it.
Intel's staunch refusal to support ECC RAM on non-Xeon chips seems incredibly chintzy at this price level.
Who is the customer that's simultaneously willing to spend $2,000 on a CPU and so desperate to save a few bucks that they'd settle for non-ECC RAM? I want to meet that person.
I'm running VMs in a LAB and I need multples of 6 virtual machines so these high core count CPUs are exactly what I'm after and I don't mind the occasional bit flip as it's for testing only, not production. But I agree supporting ECC RAM wouldn't add much to the cost of the CPU itself.
I think these things are primarily aimed at gamers and overclocking enthusiasts. They don't care about ECC at all, or might even worry that gets in the way of overclocking (since at least for now, nobody optimizes ECC DIMMs for that)
Those who care are supposed to spend even more on Xeons.
I mean, I guess they WILL - not many people invest in HEDT boards and CPUs to not milk every last bit of performance they can for their dollar, but even with a 360mm rad and a high flow pump you can’t push these very far, the thermal and electrical window just isn’t there.
I've got a nagging question about this, since this topic comes up again and again and I'm planning to buy a new machine soon after having run an overclocked i7 920 for many years:
How important is ECC RAM if I want 32 GB or more RAM? What are the consequences of not having ECC RAM?
I'll probably go for AMD this time anyway, but was wondering in general how important that is. Are non-ECC machines known to crash? Wouldn't that be a factory design flaw that would need to be fixed, at least according to EU consumer protection laws?
Typically what happens is silent corruption. Rarely, very rarely you'll get a crash, but even if a bit flips inside of program code there's no guarantee that that code will ever run again -- and if it does, the malfunction might not be obvious.
Most memory is non-program data, however, and most non-program data won't cause obvious errors if corrupted. A picture that renders incorrectly, a broken model in a game, a single burst of static in a video...
It's insidious precisely because it's hard to notice, and corruption never gets better over time.
That certainly makes sense for files, which are read into memory and written back to disk, potentially corrupt - but I wouldn't expect most compiled code to be written back to disk? (I am still on an old cheese-grater Mac Pro, so my 32GB are ECC which means I'm not worried personally, just curious).
Another important advantage of ECC RAM is early detecting of malfunctioning memory. I've yet to hear from anyone that he's regularly running memtest. While cosmic rays might be rare event, bad memory stick is certainly possible and it'll produce much more errors than cosmic rays. Some people live with it for a long time, cursing their blue-screened OS or "access violation error" programs. It's hard to distinguish software bugs from hardware bugs. ECC will let you know about it immediately.
>>>I believe there was data that with 64GB of RAM you can expect at least a single bit error or two a month.
But this would be if you were using 64GB nonstop for a month, which is quite a bit rarer than just having 64GB. And certain filesystems (ZFS, Btrfs) and most databases have software checksumming functionality to detect bitrot.
Not to say ECC isn't useful or multiple layers of protection aren't prudent, just that the problem it solves is something you're unlikely to encounter below datacenter scale.
Some research yes. To answer your final question directly - more like hourly.
Google did a big study in 2009 [0] and found error rates between 2.4E-11 and 6.7E-11 errors/bit-hour, or, in more familiar units: about 1 error per 2-GByte of memory per hour.
Obviously most errors go unnoticed, but for anything where random bit flips are unacceptable, ECC is necessary.
Note too that error rates go up with memory density, not down, so things might be worse with more recent chips.
Wow! So that means a 16GB laptop will on average experience 8 memory errors per hour under moderate load? I had no idea it was that high.
This likely explains why my Mac crashes randomly but incredibly rarely. It's like once every few months. I guess a bit flip happens at the wrong place about that often?
I would say that bit flipped in much less wrong place than the one that ended up corrupting some file on disk, unless you have a lot of insignificant data stored.
That and refusing to up the number of directly connected PCI lanes. The latest chips support 40 lanes but some of them are outboard of the CPU and connected via a controller. High speed cards like a 10 gigabit Mellanox can use 8 PCI lanes all by itself. Throw in some M.2 devices and you're out of directly connected lanes.
Per lane (unidirectional) bandwidth by PCIe revision [0]:
PCIe 1.0: 250MB/s
PCIe 2.0: 500MB/s
PCIe 3.0: 985MB/s
PCIe 4.0: 1.969GB/s
PCIe 5.0: 3.9GB/s
> A 10gb card does not need 8 PCIe lanes.
It depends on the PCIe revision(s) the card supports. If the card needs to support 10GBit on a PCIe 1.0 link, then it will need an x8. The bandwidth of PCIe 1.0 is ~250MB/s per x1 link, but there is protocol overhead and 4 lanes doesn't provide enough bandwidth, so you'd have to use an x8 (since x6 keyed cards are not a thing).
The most popular cards with enthusiasts are refurb Mellanox cards, which are PCIe 2.0 but support PCIe 1.1, so they need to utilize PCIe x8 to provide full speed 10GBit:
> High speed cards like a 10 gigabit Mellanox can use 8 PCI lanes all by itself. Throw in some M.2 devices and you're out of directly connected lanes.
The original post was saying there is a shortage of PCI lanes because of situations like a 10gb network card using 8 lanes. The truth of the matter is that if someone is running out of lanes in that situation, it is because of exceptionally inefficient use of the lanes, not because 40 lanes is as limiting as what is being implied.
> it is because of exceptionally inefficient use of the lanes
No, it's because the customer is installing the cards they want to use, and those cards are consuming PCIe lanes.
A system with more PCIe lanes lets the customer install more of whatever they want. The PCIe revision doesn't matter, and neither does the "efficiency" of the cards they're installing. More lanes means the customer can install more of whatever they want/need to install.
Sure, you could spend more on a 10GBit card that supports a newer PCIe revision and doesn't use 8 lanes, or just go and buy an AMD system which costs less than Intel and provides 64 PCIe lanes anyway.
If someone has a USB wifi dongle that takes up 4 USB ports instead of 1, would you say that the they need to get a computer with more USB ports or that they should avoid such an inefficient design if they think a lack of ports will be a problem?
> USB wifi dongle that takes up 4 USB ports instead of 1, would you say that the they need to get a computer with more USB ports or that they should avoid such an inefficient design if they think a lack of ports will be a problem?
You can recommend whatever you want, but at the end of the day the client is going to use what they want to use. I find it's best just to accept that and not argue with the client about how they're doing things wrong. You can of course make a recommendation on how to improve things, but if the client doesn't accept it, move on. You're usually hired for your professional opinion, not to stand around whining about how no one implements your great ideas.
If the client decides not to take your professional opinion, then either:
1) accept that they're paying you for your opinion which they're not following (which is their choice, it's their money)
2) quit and find another client who treats you like the demigod you so clearly are
My point about AMD providing more PCIe lanes than Intel at a lower price point still stands. The original point of OP was that Intel is not providing more than 40 PCIe lanes on anything but Xeon SKUs:
> That and refusing to up the number of directly connected PCI lanes. The latest chips support 40 lanes but some of them are outboard of the CPU and connected via a controller.
I did that, more or less. I only dropped $700 on a 10 core Xeon CPU for my workstation, which does support ECC, but I went with 128GB of non-ECC RAM because the price for the ECC stuff was ridiculous.
I wish I'd known about AMDs chips at the time though. I suspect my next machine will be an AMD, which hasn't been true in a decade.
I got 128GB DDR4 for $600. You can't even get 128GB ECC DDR for double that a year later. I was a bit concerned about memory errors as well, but my workstation has been rock solid.
Using 8x16gb sticks. But even at $250 for a pair, that's $1000. The choice was between 64GB ECC and 128GB non-ECC and I decided to live dangerously. So far, no regrets.
Its not even ECC its also multiple processors. If I want two i7's, too bad, I also need to buy Xeon. I think this is a way AMD can differentiate itself. I don't think anyone is selling dual socket ryzen boards either. Last I heard you'll need to pay for Naples for that.
These are odd moves to make during the long tail of Moore's law. Not sure if protecting the server market from a small amount of overly cheap buyers putting workstations into server use is worth hobbling everyone else and forcing them to pay the Xeon and Naples premium.
No idea who is spending $2k on this either, especially when the 1800X is $449 right now. Whale gamers I suppose.
I recently put together a Threadripper PC. Overclocked to 4.0ghz (stable) using a Noctua air cooler, my machine beats both these processors in the multi-threaded CineBench 15 benchmark (3288cb vs 3281cb & 3083cb.) Single core is a bit slower (161cb vs 189cb.)
This is my first AMD system in ages and I can't say I'm disappointed, performance is great, value is better than Intel's offerings.
101 comments
[ 2.7 ms ] story [ 175 ms ] threadI'm not really part of the "prosumer" market (what a word!), since I don't have a burning need for that much power, but it helps move prices lower in the middle and bottom of the performance range when the top end improves.
In generally it actually doesn't. The AMD Ryzen 7 1800x is selling for less than half the price of an Intel i7-6900k ($420 Vs $1010), and in benchmarks the Ryzen system beats the i7 or practically matches the performance.
Getting the same performance for 40% of the cost is something that does not depend on your use case.
Sure, but in many cases the practical difference between an Intel i7-6900k and for example an Intel i3-7XXX (or a 4 year old i5/7) is basically negligible, and the i3 is less than half the price of the Ryzen
Is there any chance that they are not having all OEMs by the balls this time too?
They were bribing and blackmailing OEMs.
What makes AMD more interesting?
And this specific example here is clearl in the high-end enthusiast space, where whoever has the best can charge the highest price and the target market will pay it.
The fact that if you want a chip that powerful, there is only Intel. An example in GPUs are the NVIDIA Quaddro cards. If your frame needs more than 16GB memory to render, there is no other card available. So they sell the same chips packaged as a Quaddro at 5x the price of a Titan.
[1] http://www.nvidia.com/object/quadro-design-and-manufacturing...
[2] http://www.nvidia.com/object/media-and-entertainment.html
How does, e.g., POWER8 compare? (And POWER9, soon?)
That said, POWER tends to be even more expensive than this.
POWER8 is not well suited for HPC workloads as their vector instructions (VMX/AltiVec) are only 128-bit wide vs Intel's latest Skylake Xeon CPUs have vector instructions (AVX512) that are 512-bit wide.
For non HPC workloads (e.g. integer arithmetic) the POWER8 CPUs were somehow competitive with Intel's CPUs when they were first released in 2015 but today AMD and Intel's CPUs are clearly faster.
From what I can tell CPU enthusiasts were not impressed by the POWER8 CPU architecture.
In the case of nVidia the quadro cards are the only ones sold as manufactured by the company and are incredibly high binned.
Lots of people on this thread are conflating a few things such as ML, but I work in an industry where we deal a lot with GPUs and its certainly a fact that flipped bits are significantly more common than many think. If you do any work that requires the output to be nonvolatile then you can run into issues.
http://ieeexplore.ieee.org/document/5493404/?arnumber=549340...
There may be some that are down to the wire per unit cost (typically the cheaper ones) but I have a feeling Intel operates with plenty of margin on most of their higher end products. And plenty of them are the same chip repackaged.
If anything AMD should follow NVIDIA in supporting OpenACC which is the way to go as far as true probability goes.
It doesn’t matter that the rocm libs are optimised for AMD.
In all effectiveness ROCm is AMD's version of CUDA it includes it's own run time and driver stack which effectively replaced the OpenCL frontend for AMD GPUs, basically it gives AMD the ability to do with their GPUs what CUDA allowed NVIDIA to do and it's not to be dependent on a camel (OpenCL).
This is what ROCm is: https://rocm.github.io/images/ROCm_Stack.png or at least this is what AMD is pushing it to be, currently a lot of things are still unfinished or non-existent.
The representation of it being open is misleading, while more of the source code is available than for CUDA, ROCm isn't vendor agnostic on any tangible level. In 2-3 years once it's somewhat in a production ready state it will lock you into a single vendor just like CUDA does.
Which to me says that Intel is not competitive. If you can only release a slightly better product at twice the price, then you failed. Of course Intel's products have some branding inertia behind them, too, but that will be short-lived if the products don't become more competitive in the near future.
[1] Kaby Lake-X, 4/8 cores/threads, 4.3/4.5 freq/turbo, $329
https://www.anandtech.com/show/11549/the-intel-kaby-lake-x-i...
https://www.anandtech.com/show/11697/the-amd-ryzen-threadrip...
> And, in theory, I suspect that a dual-7740X build might match or beat TR in multithreaded, too.
Is that possible? And I really doubt you'll get 8 cores to beat 16.
My point is it's the category most enthusiasts would be happiest with, even if they think otherwise.
>>>Is that possible?
Looks like I was mistaken, I thought the Sky/Kaby Lake-X chips used the same socket as the Xeons. So no.
>>>And I really doubt you'll get 8 cores to beat 16.
If it were possible, it would be 8 cores operating at a 50% higher frequency (with substantially more room for overclocking, too) and that support AVX-512 instructions. And you get quickly-diminishing returns from throwing more threads at a program (just mathematically, even if it's embarrassingly parallel and CPU-bound, each time you double the threadcount you get half the improvement of the previous doubling).
That's fine, but use that category for both manufacturers.
> 8 cores operating at a 50% higher frequency
26% higher base, 10% higher boost.
With 3 glaring factual errors across two posts, I'm going to go out on a limb and suggest to you that the average enthusiast purchasing Threadripper is in fact more aware of the trade offs between the products available on the market than you are.
The only thing that Intel surpassed them is single-thread performance. It's important even on workstation, IMO, that's the main reason I'm thinking to stick with Intel for now, everything else AMD made better.
Intel's site says Q3 (which ends this month): https://ark.intel.com/products/126699/Intel-Core-i9-7980XE-X...
And in a blog post they appeared to indicate the 7980XE would be available Sept 25 (at least the 14-core version, which also hasn't yet appeared, depending on parsing): https://newsroom.intel.com/news/intel-unveils-full-intel-cor...
So the availability lag might be worse than normal.
* Intel doubled the price of a threadripper 1950x for giving you a 8% increase on multithread performance (measure by Cinebench R15 scores).
* when Intel could't afford to match AMD high end CPU's 60 PCIE lanes, they invented the term "platform PCIE lanes" to include those from chipsets.
* when Intel couldn't match AMD's TDP, they came up with some different TDP definition, as a result the 165w TDP Intel 7980xe draws 15w more than a 180w threadripper 1950x.
Intel must have a pretty good propaganda department known as "marketing"?
for those who believe Intel is responding to AMD's challenge - no, it is not, according to leaks, Intel is about to release i3-7360x, which is going to be 1.25% faster (yeah, 1.25% faster, no typo, source linked) than the previous generation. when you can fully utilize your well established marketing department to fool consumers to buy your overpriced slow processor with far less cores and PCIE lanes, why bother giving those customers anything better?
https://videocardz.com/72775/intel-preps-dual-core-i3-7360x-...
* why Intel released Q6600 10 years ago and made quad core mainstream? because AMD was competitive.
* why Intel repeatedly refused to make 8 cores mainstream in the last 10 years? because AMD was not competitive.
* why Intel finally made 6 cores mainstream 10 years after the release of Q6600? because AMD is competitive again.
> when Intel couldn't match AMD's TDP, they came up with some different TDP definition, as a result the 165w TDP Intel 7980xe draws 15w more than a 180w threadripper 1950x.
So far it was AMD which had a weird way to calculate the TDP, and it was them that went over while with Intel TDP pretty much equaled max power draw. That Intel now goes over is clearly a sign that they are stressing their boundaries and the of their architecture, but it's not really a big distinction against AMD.
everything else is just cheap talk.
All modern Intel CPUs will draw more than their rated TDP if you run them hard.
Even just within the CPU cores, different ops consume vastly different amount of energy while still 'maxing out' utilisation. And we haven't even considered accelerator hardware within the CPU which is totally independent of the CPU cores but can still consume a large amount of energy (e.g. H.264, AES).
Intel no doubt calculate a maximum power consumption figure where every bit of hardware on the CPU is consuming as much as it can. It'd be utterly meaningless -- probably on the order of 100W for a 2-4 core laptop part or 400W for a server part.
You can design a computer to handle those figures, but you don't want to. So TDP reflects intended usage rather than peak demand.
If you're a 'prosumer', you're presumably smart enough to size your power supply and cooling appropriately. TDP is not the number you're looking for.
the 80-100% is bit of a vague number, I was just kinda referring to the cpu utilization as reported by the OS. For numerics I would consider a realistic full load 75%+ pipeline fill time on the AVX codepaths. With a lot of compute you can get quite close to 100% quite quickly if your software is highly optimized (and uses enough compute to hide the memory loads). TDP should just be 100% pipelined load of the highest consuming codepath.
But my general point is still that these benchmarks should not exceed the TDP for these types of parts. For everyday use these loads might not be realistic. But for a lot of people looking to spend $2000(!) on a single CPU, these benchmarks are indicative of real life usage.
If hard does not include AVX I think that is not correct. All locked i5s and i7s stayed under or very near to their TDP. And also the K units, before overclocking. https://www.anandtech.com/show/10968/the-intel-core-i7-7700k... as example. However, I noticed that the reviewer mentioned that wasn't true for the 4790K and 6700K.
I should mention that I'm talking about the desktop versions here. The laptop parts are a different story.
That CPU is built for overclocking competitions, don't know why so many people are upset over it.
https://www.reddit.com/r/Amd/comments/6svy1a/tdp_vs_tdp/dlg8...
Who is the customer that's simultaneously willing to spend $2,000 on a CPU and so desperate to save a few bucks that they'd settle for non-ECC RAM? I want to meet that person.
Those who care are supposed to spend even more on Xeons.
How important is ECC RAM if I want 32 GB or more RAM? What are the consequences of not having ECC RAM?
I'll probably go for AMD this time anyway, but was wondering in general how important that is. Are non-ECC machines known to crash? Wouldn't that be a factory design flaw that would need to be fixed, at least according to EU consumer protection laws?
Most memory is non-program data, however, and most non-program data won't cause obvious errors if corrupted. A picture that renders incorrectly, a broken model in a game, a single burst of static in a video...
It's insidious precisely because it's hard to notice, and corruption never gets better over time.
Suppose I'm doing scientific computing, is ECC solving a super rare problem, or something that might happen weekly or monthly?
Thus ECC would be super helpful to those that expect that most of the 64GB will be filled with bits that shouldn't be changed.
This article by a system integrator is a pretty decent data point: https://www.pugetsystems.com/labs/articles/Advantages-of-ECC...
But this would be if you were using 64GB nonstop for a month, which is quite a bit rarer than just having 64GB. And certain filesystems (ZFS, Btrfs) and most databases have software checksumming functionality to detect bitrot.
Not to say ECC isn't useful or multiple layers of protection aren't prudent, just that the problem it solves is something you're unlikely to encounter below datacenter scale.
Google did a big study in 2009 [0] and found error rates between 2.4E-11 and 6.7E-11 errors/bit-hour, or, in more familiar units: about 1 error per 2-GByte of memory per hour.
Obviously most errors go unnoticed, but for anything where random bit flips are unacceptable, ECC is necessary.
Note too that error rates go up with memory density, not down, so things might be worse with more recent chips.
[0] http://www.cs.toronto.edu/~bianca/papers/sigmetrics09.pdf
Wow! So that means a 16GB laptop will on average experience 8 memory errors per hour under moderate load? I had no idea it was that high.
This likely explains why my Mac crashes randomly but incredibly rarely. It's like once every few months. I guess a bit flip happens at the wrong place about that often?
Per lane (unidirectional) bandwidth by PCIe revision [0]:
PCIe 1.0: 250MB/s
PCIe 2.0: 500MB/s
PCIe 3.0: 985MB/s
PCIe 4.0: 1.969GB/s
PCIe 5.0: 3.9GB/s
> A 10gb card does not need 8 PCIe lanes.
It depends on the PCIe revision(s) the card supports. If the card needs to support 10GBit on a PCIe 1.0 link, then it will need an x8. The bandwidth of PCIe 1.0 is ~250MB/s per x1 link, but there is protocol overhead and 4 lanes doesn't provide enough bandwidth, so you'd have to use an x8 (since x6 keyed cards are not a thing).
The most popular cards with enthusiasts are refurb Mellanox cards, which are PCIe 2.0 but support PCIe 1.1, so they need to utilize PCIe x8 to provide full speed 10GBit:
PCI Express 2.0 x8 (1.1 compatible) [1]
[0] https://en.wikipedia.org/wiki/PCI-Express
[1] http://www.mellanox.com/page/products_dyn?product_family=20&...
No, it's not. Not when the cards people are using are targeted at PCIe 2.0/1.1.
Since the context was 10GBit cards, most people are using the Mellanox card I linked to, which is PCIe 2.0.
Whine about the age of the standard all you want, what's on the market today and in use by a majority of home users is PCIe x8 for 10GBit.
The original post was saying there is a shortage of PCI lanes because of situations like a 10gb network card using 8 lanes. The truth of the matter is that if someone is running out of lanes in that situation, it is because of exceptionally inefficient use of the lanes, not because 40 lanes is as limiting as what is being implied.
No, it's because the customer is installing the cards they want to use, and those cards are consuming PCIe lanes.
A system with more PCIe lanes lets the customer install more of whatever they want. The PCIe revision doesn't matter, and neither does the "efficiency" of the cards they're installing. More lanes means the customer can install more of whatever they want/need to install.
Sure, you could spend more on a 10GBit card that supports a newer PCIe revision and doesn't use 8 lanes, or just go and buy an AMD system which costs less than Intel and provides 64 PCIe lanes anyway.
You can recommend whatever you want, but at the end of the day the client is going to use what they want to use. I find it's best just to accept that and not argue with the client about how they're doing things wrong. You can of course make a recommendation on how to improve things, but if the client doesn't accept it, move on. You're usually hired for your professional opinion, not to stand around whining about how no one implements your great ideas.
If the client decides not to take your professional opinion, then either:
1) accept that they're paying you for your opinion which they're not following (which is their choice, it's their money)
2) quit and find another client who treats you like the demigod you so clearly are
My point about AMD providing more PCIe lanes than Intel at a lower price point still stands. The original point of OP was that Intel is not providing more than 40 PCIe lanes on anything but Xeon SKUs:
> That and refusing to up the number of directly connected PCI lanes. The latest chips support 40 lanes but some of them are outboard of the CPU and connected via a controller.
I wish I'd known about AMDs chips at the time though. I suspect my next machine will be an AMD, which hasn't been true in a decade.
These are odd moves to make during the long tail of Moore's law. Not sure if protecting the server market from a small amount of overly cheap buyers putting workstations into server use is worth hobbling everyone else and forcing them to pay the Xeon and Naples premium.
No idea who is spending $2k on this either, especially when the 1800X is $449 right now. Whale gamers I suppose.
This is my first AMD system in ages and I can't say I'm disappointed, performance is great, value is better than Intel's offerings.
CineBench / UserBenchmark screenshots: https://ibb.co/mapRUQ
No OC vs OC?