The AMD Threadripper 2 Teaser: Pre-Orders Start Today, Up to 32 Cores
anandtech.com
anandtech.com
Until such a time I can get the equivalent tool to stop the hardware spyware built into the CPU I can have no enthusiasm or motivation to buy AMD chips. Not to say I want to buy Intel parts - they have nothing to do with third party efforts to nullify their backdoor - but if I were buying a chip tomorrow it would be a begrudging Intel purchase just for me_cleaner.
If AMD wants to dominate they don’t need to care about your particular use-case at all in fact. They just need to produce the fastest x86 chips at the cheapest price.
The best outcome would be to have AMD provide a first-party, auditable option to disable it, otherwise the community will have to do it themselves, which will probably take longer if there's less people using them. Until then, the main focus would be the buy the best performing product, because that's the only real differentiator.
https://www.raptorcs.com/TALOSII/
Pricy, but they don't have the ME/PSP problem.
I do think that Intel will manage to get silicon out a bit earlier then their current end of CY19, but if they don't, TR and Ryzen may be the default go-to for manufacturers.
The difference between 14nm and 14nm+++ is pretty close to a full node step...
> And just recently they pushed their 10nm products even further back to the second half of 2019. AMD might have the first 7nm CPUs out by then. Of course process technologies are not easily comparable between fabs but it is still crazy to see Intel starting to fall behind in process technology - a game they dominated for decades.
The numbers used for process 'sizes' don't really hold any meaning anymore and can't be directly compared. Intel's 14nm is 'better' in almost every metric than other fabs 10nm process. Intel's 10nm process is expected to be similarly comparable to other fabs 7nm.
Intel isn't really behind on process tech, but the competition has closed the gap considerably.
The still have tons of money and spend more time breaking that wall.
Good for us anyway. Don't mind getting my Intel CPUs cheaper.
I fail to see how that's relevant. What matters is what products are available in the market and how much they cost, and AMD's offering is both cheaper and more performant than anything Intel managed to put together.
No, Intel's 10nm is worse than TSMC's 7nm in every metric. Not significantly so, but it is so.
It also seems to be viable whereas Intel's hasn't been proven yet (the only released product is a mobile chip with the iGPU disabled due to defects and low yields).
Yes -- AMD got lucky, this time. I sincerely doubt they (or anyone else) had any idea how Intel will struggle with their 10nm when they started on Zen in 2012. Back in 2013 Intel said they will have Cannon Lake out in 2015 (!) https://www.theregister.co.uk/2015/07/16/intel_10nm_14nm_pla... and today, mid 2018, the only released Cannon Lake CPU is a sad little dual core with the graphics chip disabled ( i3-8121U) selling in "very limited quantities".
65nm: 2006
45nm: 2007
32nm: 2010
22nm: 2012
14nm: 2014
10nm: 2019 (?)
That being said... Processor pre-orders? I had No idea that was a thing. Hope you get a QA discount. It doesn't talk about sockets but for sake of preorderers I hope it's compatible with original threadripper boards. edit: Actually it does, and they mention it's compatible with existing main boards too. I missed that this was multiple pages.
AMD has already committed to using the same socket until 2020.
I disagree based on my personal experience. The CPU upgrade path for my current desktop computer ended roughly a year after I bought it, even though that CPU generation had been introduced very recently by Intel. I got a relatively high-end CPU, so upgrading to anything but the highest-end CPUs of that generation doesn't really make sense from a performance point of view. Yet those still cost a surprising amount of money today and offer worse performance than the midrange options of newer generations.
Simply put, extending the lifetime and thus upgrade path out by at least 150% means that the options available 5 years down the line will be 150% "better", even though no CPUs for that socket might even be in production anymore at that time. For me that seems like a significant advantage.
Pre-orders don't make sense in the case of downloadable games, where there is no scarcity.
that's 1 good reason to preorder, plus some bonuses that tend to be lame.
Turbos are going to be pretty iffy under sustained load, let alone Precision Boost and XFR, which are (iirc) enabled by default.
https://www.youtube.com/watch?v=aKe7CnZT9ZE
AMD is shipping all the review kits with an unspecified "add-on VRM cooling kit" to try and help that, and there are second-gen X399 boards coming out with 16 phases (up from 6-8 in the first gen).
Ultimately it depends on how cache friendly your workload is.
The 32c/64t is kind of a marketting gimmick selling binned Eypcs. If your doing memory bound HPC consider the 16c/32t model for roughly equal performance at a lower price.
Also, your the Intel setup has a 25 GB/s QPI bus between the processors. Threadripper runs an Infinity Fabric bus, with 42 GB/s between each die internally (aggregate 170 GB/s). While a couple of those internal dies are not directly connected to the memory, you should have little trouble sharing meals among those hungry processors.
I've seen some people hit 3466MT/s on higher-end more expensive RAM on AMD's 12nm stuffs (Ryzen 2). So that's probably the practical limit.
3200 MT/s on Hynix dies is probably a reasonable expectation.
For actual production workloads you want EPYC which has a memory controller for each die, meaning it's 8 channel.
A lot of these pins aren't used on TR, since you only get half the memory controllers you'd have with EPYC
Then Intel was on top of the game again with the core architecture and while bulldozer (iirc) improved in that regard again Intel was still ahead, especially as bulldozer sucked pretty much in every other regard. Ryzen seems to be about on par with current Intel CPUs (performance per watt), depending on benchmark.
Funny thing is that Core, AFAIK, came from a separate laptop chip branch. For 64 bits Intel was betting on Itanium, which crashed and burned like iAPX-432 did before. And so Intel once again was forced to stay on the 8080 treadmill ...
"They worked on it in their spare time and it was really a passion project for about a year before they sought the green light from management, which is quite unusual – it was something they really cared about."
The Ryzen 2000 chips have a smarter SpeedStep-type functionality than Intel does, it pretty much squeezes out all the OC performance from the chip while maintaining normal power management, while Intel (and Ryzen 1000) needs to manage voltages on a processor-wide basis, which eats more power.
Ryzen is default choice now for me and my brother. This pushes even more people toward DEFAULT RED.
It's called "marketing", and it's definitely not limited by the laws of physics as we know them.
You too can learn the ways of marketing, but it comes at a very steep price: your eternal soul. (JK, that's sales.)
With a modular design the lower bound is, indeed, linear, allowing AMD to severely undercut Intel.
Cycles is easily 5+ minutes a frame, maybe 20+ minutes on a threadripper if you have an interior scene or something hard to light. Even with Cycle's denoiser, you need a lot of rays to get a decent animation.
Even a short 10-second animation through Cycles will take multiple days to render on a Threadripper.
Note that most price comparisons show that the iMac Pro is actually quite the deal.
Do remember that this 32-core Threadripper offers quad-channel memory, 64x PCIe lanes, and ECC support (!!). So the Threadripper offers pretty much everything a professional wants.
The only downside to Threadripper is that its multi-threaded performance is similar to a quad-socket / quad-CPU design (~200ns to ~300ns latencies to memory, depending on how many "hops" and NUMA nodes and stuff...). So audio professionals and gamers who are IIRC latency bound may want to stick to Intel.
But almost everyone else is bandwidth bound (video editing, 3d rendering, compilers, web servers, LT Spice...). So most professionals would probably prefer a Threadripper.
It's the case that building a comparable computer (in terms of the specific iMac Pro hardware) is very similarly priced BUT requires labor, the purchase of an OS and additional peripherals: https://youtu.be/SONKIJd8xRM?t=187
That said, your point that Apple could've chosen better hardware at its price, is reasonable but off topic.
[update]Sorry he said $20,000. SO you are right
Same with GPU systems like redshift, alternatively I can dedicate half the machine to rendering and the other half to continued work.
I know
Unless you’re using Houdini, I think you’ll get the same result using Redshift with 16 threads. Houdini is the only DCC that uses more than one thread to prep the scen
It can take a very long time just to process a 5 minute video.
I've got my costs down for a TR machine and it comes in at around $4K CAD.
Statistically I’m pretty sure most of them aren’t too scientific about their assumptions.
Hence the huge market for lots of HW which no normal person would buy, explicitly targeted at gamers.
TLDR: reality doesn’t matter as long as the x CPUs are faster :)
They check benchmarks and see what kind of gains they can expect over what they currently have.
Even for rendering it's still the CPU preparing the commands and doing a first-pass curation to minimize rendering load.
That said, in this case it's not so much running the games better, as yes the GPU is usually the bottleneck. It's more for gamers that also stream or similar, which is an increasingly popular thing. You can GPU offload it, but that hurts your game performance and GPU encoding quality tends to be relatively fixed and relatively poor. CPU encoding is more flexible and higher quality, and if you've got an extra 8 cores sitting around idle anyway might as well use those instead of eating into the GPU power budget.
Also you've got the PCI-E lanes to run things like SLI or multi-GPU along with RAID NVME drives. Threadripper has 60 lanes. Something like an i7-8700k only has 16 PCI-E lanes - you can't even run a single GPU and an NVME drive without multiplexing your PCI-E lanes.
The sole advantage is that it's a lot of PCIe lanes for the money, so it can make sense for storage builds that want to address a lot of NVMe, or GPU compute builds that need very little CPU horsepower.
Also, like all TR boards, the motherboards are extortionate. They start at about $350 and go up from there. And that doesn't even buy you a futureproof system - the higher core counts on the TR 2000 series mean that first-gen boards likely won't be able to turbo the new processors, you're running base-clocks.
That's an excellent deal compared to the Talos POWER9 boards :)
That's what I was thinking. The 1900x looks useful for quad-GPU Redshift rendering setups.
Its certainly not "mainstream". The 1920x / 2920x is far better bang for your buck. But the price from AMD scales very well. If you're willing to spend bigger bucks on cooling, the 2990WX is great. But 250W TDP is going to be a tough-one to design around.
Custom-cooling is probably the answer. Enermax's TR4 AIO turned out to be all sorts of awful, unfortunately. And Noctua's Air coolers cap out at 180W TDP designs.
The solution, of course, is to name them. Can you share which motherboards you've seen people complaining about?
Remember: DDR4 goes straight to the CPU these days. I'm fairly certain that motherboard makers can make a simple wire connection between the DDR4 pins and the CPU itself without much issue.
https://community.amd.com/thread/217871
The AMD Community knows to avoid Hynix and to buy Samsung. Hynix did improve with some BIOS updates (in particular: increasing V_soc voltage to 1.1 and other tricks). Update your BIOS if you have any issues.
Zen+ dies have better DDR4 compatibility.
AMD's memory controller is actually from Rambus.
The good thing is as more EPYC moves inside DC, there will be lots more memory testing done for AMD.
It doesn't have 256 bit wide units, so it's 2 clock cycles instead of 1… but that's actually good. Intel's 256 bit units cause horrible downclocking issues: https://blog.cloudflare.com/on-the-dangers-of-intels-frequen...
What even is "SSE speeds"?!