Can AMD RIP the Core i9? - Ryzen Threadripper

[Alpha297]

2 minutes ago, GabenJr said:

It's being calculated based on the weighted results across all testing (gaming specifically average FPS and 3DMark+Superposition). It's based on the methods @Enderman and @patthehat pointed out in the Core i9 correction thread. The technically comment is just calling out the oddity there at <$100.  

Thank you, I'll have a peek over there.  Threadripper is way outside my budget, but I still like to understand how the numbers are reached on these things.  I appreciate your time. 

[LukaH]

Well, you made a long journey from LABS to FABS, Ryzen! Ryzen! 

You never stopped hoping; now you're THE BEST, Ryzen! Ryzen! 

When the naysayers 'nay' you picked up your pace.

You said Intel ain't gonna stop me so get out of my face.

I'm having adventures all over the place, Ryzen! RYYYYYZZZZZEEEENNN!

 

 

seinfield anyone?

 

also yes there should be a lot of *(footnotes) but whatever tear me a new one.

[Enderman]

52 minutes ago, GabenJr said:

It's being calculated based on the weighted results across all testing (gaming specifically average FPS and 3DMark+Superposition). It's based on the methods @Enderman and @patthehat pointed out in the Core i9 correction thread. The technically comment is just calling out the oddity there at <$100.  

I saw that graph in the video, glad you guys decided to include it.

I'm interested to see that graph in future videos too, especially for stuff like r5 vs r7 or i5 vs i7.

 

The bigger price difference will make that graph a lot more useful, it will show where an i5 is better price/performance for a certain range of system price, and the i7 is better for more expensive builds

[SpaceGhostC2C]

1 hour ago, GabenJr said:

It's being calculated based on the weighted results across all testing (gaming specifically average FPS and 3DMark+Superposition). It's based on the methods @Enderman and @patthehat pointed out in the Core i9 correction thread. The technically comment is just calling out the oddity there at <$100.  

 

It's sad you decided for that, since they were wrong then and remain wrong now  

 

First, it is not the correct way to assess a decision at the margin. When you change a component, you obtain a particular marginal performance change, and you want to compare it to its marginal cost. You gave the correct response in the original thread, when stating that adding the cost of components you keep constant (hence, sunk costs), you are muddying the waters and not coming to the right conclusion (known as "sunk cost fallacy"). This is more transparent in @patthehat's post thanks to his use of math, which makes clear that he's calculating average costs, not marginal ones, hence not the decision-relevant costs.

The correct way to compare a choice of an individual component was and still is to look at the marginal cost of choosing A over B, and the marginal performance obtained by using A instead of B. Because we are looking at a discrete choice, instead of the usual marginals (i.e., derivatives) we use deltas: the relative value can be summarized by the delta in performance divided the delta in cost  (the only complication for CPUs being that you also need a different motherboard, so CPU price vs CPU price isn't the full story. For other components, like GPUs, it would be trivial). The other complication, of course, is defining "performance", but that's another topic: this is all assuming you are happy with whatever average or summary of performance index you ended up with.

In any case, there is no "comparison graph", just a single number: the additional performance per dollar of additional spending. Then it's up to each buyer to decide how much extra per dollar is enough, negative numbers being obvious poor choices.

Now, having said that, the graph could still be useful to see how building systems around these CPUs with different budget points delivers different overall performance per dollar ratios. It would even show how hitting certain bottlenecks flattens out performance and ruins value, while a little investment in the relevant "bottleneck" can restore value at a higher price point.

The problem is, performance is constant here, based on whatever the test bench was, so the assumption is that you can reach different price points by just burning money, because the additional expenditure has zero impact on performance. Hence, while it could be interesting, in its current form it's just a mix up of what would happen depending on how lucky or unlucky you are in terms of pricing when trying to reproduce the testbench with some average cost calculation that naturally flattens as you increase the fixed cost of reaching the same performance (it's not about "diminishing returns", since we are assuming the same performance as we change the fixed cost, so it's literally useless overhead - the return doesn't diminish, it's flat at zero). It is essentially saying "look, given performance, A is better value than B"; "but what if a I pay $1,000,000 for my mouse?"; "oh no, wait, in that case buy B and enjoy the same performance as before changing the price of the mouse, because that's somehow better value!". It is indeed a stark example of why the sunk cost fallacy is, well, a fallacy.

Perhaps noticeably, a graph that takes into account that additional expenditure goes somewhere, impacting performance, would most of the times have the opposite shape: consider for example a gaming benchmark. It would be flat at low budgets, since essentially at that point all your expenditure goes to the CPU and you end up with the cheapest GPU, turning the CPU choice irrelevant, and it will get steeper as you increase the budget enough to accommodate a GPU that would make your system CPU-bound.

[mariushm]

The value thing is difficult. 

Was the math done for a $550 motherboard or for the cheapest x399 motherboard that would be around $330 ?  That Zephyr board is way more expensive compared to cheaper boards.

Was the math done with 4 memory sticks or 2? Was the math done with cheapest water cooler or some 360mm expensive fan..

 

Obviously the x299 platform may have cheaper boards today since it's already on the market for a longer time.

[SpaceGhostC2C]

2 hours ago, GabenJr said:

It's being calculated based on the weighted results across all testing (gaming specifically average FPS and 3DMark+Superposition). It's based on the methods @Enderman and @patthehat pointed out in the Core i9 correction thread. The technically comment is just calling out the oddity there at <$100.  

It's not an oddity: what you are doing by following patthehat's formula is equivalent to assign weights to price and to performance. Because instead of looking at the price difference (either absolute or relative to one of the prices) you look at the price difference relative to arbitrary constant x, and move x around, you are effectively tilting the weight towards price or towards performance. As x goes to zero, dPrice/x goes to infinity, so price is everything and the cheapest CPU wins. As x goes to infinity, dPrice/x goes to zero, so performance is everything and you effectively ignore prices, it's a pure performance comparison.

Linking the constant x to fixed costs is fine, by fixed costs being fixed are irrelevant. It is however misleading to call this "value comparisons" for the reasons I mention in my previous post.

[Serin]

Whomever is doing the benchmarking - I have a message for yee.
The Blender benchmarks - If I'm not mistaken you're leaving them on the default 256x256 tile size, at least for the BMW file.
This whilst default, it isn't a useful output for us to compare.
For CPU's its best to go 16x16 or I suppose for threadripper 32x32.
GPU can go between 256 to 512 to 1024. 256 is the most used, but 512 will get the fastest result.
It varies pretty obviously for CPU, less so for GPU.

If I might another suggestion - note the tile size on the results so we can check with already known results.
The BlenderArtists thread on benchmarks is a good resource, at least more reliable than blenchmark etc.

(For reference a 5960x at 4.5ghz gets under a minute and a half on the BMW, which is what twigged me to your tile size settings)

[dizmo]

As cringe as that picture is, you could at least have been ripping off an Intel shirt

I've gotta say though, for what it is that sleeving kit for the AIO's is a rip off.

[Emily Young]

41 minutes ago, Serin said:

Whomever is doing the benchmarking - I have a message for yee.
The Blender benchmarks - If I'm not mistaken you're leaving them on the default 256x256 tile size, at least for the BMW file.
This whilst default, it isn't a useful output for us to compare.
For CPU's its best to go 16x16 or I suppose for threadripper 32x32.
GPU can go between 256 to 512 to 1024. 256 is the most used, but 512 will get the fastest result.
It varies pretty obviously for CPU, less so for GPU.

If I might another suggestion - note the tile size on the results so we can check with already known results.
The BlenderArtists thread on benchmarks is a good resource, at least more reliable than blenchmark etc.

(For reference a 5960x at 4.5ghz gets under a minute and a half on the BMW, which is what twigged me to your tile size settings)

For CPU rendering, the default is much lower - I believe either 16x16 or 32x32, which seems "sane" enough. I do regret not testing both to see if that would make a difference.

 

I will be noting the tile size(s) that perform best and use those times as the results in future videos; This first became an issue for me with GPUs, and if it's also an issue for CPU performance, then that's how it will look from now on regardless.  

[boss20007]

@LinusTech  Show US PLEASEEE  A THREADRIPPER 1950X WITH MOTHERBOARD x399 WITH 3 RX VEGA 64 WITH CROSSFIREX BRIDGE!!!

[Marinatall_Ironside]

6 hours ago, mariushm said:

The value thing is difficult. 

Was the math done for a $550 motherboard or for the cheapest x399 motherboard that would be around $330 ?  That Zenith board is way more expensive compared to cheaper boards.

Was the math done with 4 memory sticks or 2? Was the math done with cheapest water cooler or some 360mm expensive fan..

The motherboard and the memory kits were provided to all reviewers.

 

Obviously for the value, they're looking at that one component only. Factoring in all of the components for the total cost to then measure the value would be like comparing a Maingear computer to an iBuyPower computer. Since most of us here buy all of the components individually, you generally look at the value of each individual component.

 

The motherboard in question is the Asus ROG Zenith Extreme.

Edited August 11, 2017 by JurunceNK

[Marinatall_Ironside]

3 hours ago, Serin said:

Whomever is doing the benchmarking - I have a message for yee.
The Blender benchmarks - If I'm not mistaken you're leaving them on the default 256x256 tile size, at least for the BMW file.
This whilst default, it isn't a useful output for us to compare.
For CPU's its best to go 16x16 or I suppose for threadripper 32x32.
GPU can go between 256 to 512 to 1024. 256 is the most used, but 512 will get the fastest result.
It varies pretty obviously for CPU, less so for GPU.

If I might another suggestion - note the tile size on the results so we can check with already known results.
The BlenderArtists thread on benchmarks is a good resource, at least more reliable than blenchmark etc.

(For reference a 5960x at 4.5ghz gets under a minute and a half on the BMW, which is what twigged me to your tile size settings)

This ^^

 

Although I no longer use Cycles (I recently transitioned to AMD's Radeon ProRender), this is something worth looking at going into the future.

[Nesagsar]

I'd love to see a video that lays out the requirements for a serious professional content creation PC. My current job is as a designer and graphic artist but our computers and IT infrastructure are nowhere near what we are supposed to have, they have even tried to take away our real PCs and use thin clients. This network cant even transfer a 200 MB file reliably, I'm not going to attempt running After Effects over it. If LMG could make a Fast as Possible or something which lays out what a real professional system looks like compared to a standard business, home, and gaming PC setup it would be really useful in helping us explain why the graphics and media team needs different hardware than people who work in Excel and Word. Threadripper seems like a good starting point for such a video.

[patrick3027]

On 10/08/2017 at 3:24 PM, Goku-sama said:

Apparently it can .

 

But not in CS:GO xp.

The usefulness of testing a low quality game on a 16 core CPU remains a mystery for me. 

 

Anyone who buys a CPU with more than 8 cores for gaming is retarded. 

[Guest]

Just now, patrick3027 said:

The usefulness of testing a low quality game on a 16 core CPU remains a mystery for me. 

 

Anyone who buys a CPU with more than 8 cores for gaming is retarded. 

Well... they might just have the money and desire to go all out instead.

 

The show offs certainly do these things all the time, without even maximizing the product potential.

[Marinatall_Ironside]

16 hours ago, Nesagsar said:

I'd love to see a video that lays out the requirements for a serious professional content creation PC. My current job is as a designer and graphic artist but our computers and IT infrastructure are nowhere near what we are supposed to have, they have even tried to take away our real PCs and use thin clients. This network cant even transfer a 200 MB file reliably, I'm not going to attempt running After Effects over it. If LMG could make a Fast as Possible or something which lays out what a real professional system looks like compared to a standard business, home, and gaming PC setup it would be really useful in helping us explain why the graphics and media team needs different hardware than people who work in Excel and Word. Threadripper seems like a good starting point for such a video.

Everyone has vastly different needs for their work. Some don't need as much computational power while others do need that, and that's okay. However, for any kind of good content creation work, thin clients are out of the question, nor is doing work over the local network (as even a high-speed, low-latency infrastructure is bound to get hammered real quick when dozens of people move data around fast).

 

I don't think LMG is going to make a video on that topic, because as I said, everyone has different needs. But it would be cool nevertheless.

 

To me, a workstation for use in 3D animation and modeling, graphic design, video production, and CAD/CAM is defined as a personal computer (running Windows (the vast majority of workstations I know of runs this), macOS (won't discuss how to perform this, but worth mentioning anyways), or Linux (or any other open-source OS)) that is:

• robust in I/O connectivity (very important for graphics processors, storage, networking, etc.).
• high performance to take on even the most demanding workloads.
• very reliable (no silly software/firmware bugs will even bring down a workstation).
• user-serviceable (need to upgrade or replace a bricked component, no problem).
• possesses an upgrade path for later on down the road, even if it looks maxed out for its time (more storage, GPUs, network adapters, whatever in between).

Hell, I don't see the iMac Pro as having most of these features I've listed (it's high in performance, yes, but the I/O I feel is lacking (no variety in USB connectors, SD card readers, and I feel that there's too many Thunderbolt 3 connectors (I know there's people that feel different about the Thunderbolt thing)), and they're made to be disposable and have no upgrade path once you commit the purchase), despite being an excellent refresh to the Pro desktop lineup from Apple.

Edited August 12, 2017 by JurunceNK

[Nesagsar]

On 8/11/2017 at 10:27 PM, JurunceNK said:

Everyone has vastly different needs for their work. Some don't need as much computational power while others do need that, and that's okay. However, for any kind of good content creation work, thin clients are out of the question, nor is doing work over the local network (as even a high-speed, low-latency infrastructure is bound to get hammered real quick when dozens of people move data around fast).

 

I don't think LMG is going to make a video on that topic, because as I said, everyone has different needs. But it would be cool nevertheless.

 

To me, a workstation for use in 3D animation and modeling, graphic design, video production, and CAD/CAM is defined as a personal computer (running Windows (the vast majority of workstations I know of runs this), macOS (won't discuss how to perform this, but worth mentioning anyways), or Linux (or any other open-source OS)) that is:

• robust in I/O connectivity (very important for graphics processors, storage, networking, etc.).
• high performance to take on even the most demanding workloads.
• very reliable (no silly software/firmware bugs will even bring down a workstation).
• user-serviceable (need to upgrade or replace a bricked component, no problem).
• possesses an upgrade path for later on down the road, even if it looks maxed out for its time (more storage, GPUs, network adapters, whatever in between).

Hell, I don't see the iMac Pro as having most of these features I've listed (it's high in performance, yes, but the I/O I feel is lacking (no variety in USB connectors, SD card readers, and I feel that there's too many Thunderbolt 3 connectors (I know there's people that feel different about the Thunderbolt thing)), and they're made to be disposable and have no upgrade path once you commit the purchase), despite being an excellent refresh to the Pro desktop lineup from Apple.

I certainly agree that graphics shops all have different requirements but I would argue that there are some general items that we can all agree on. Our computers are doing okay on CPU and RAM with Xeons and 32 GB of DDR3. However - they put NVS 510 graphics in there, the HDD isnt in RAID and we have bricked three systems because the HDDs had mechanical failures, and these things are running on 350 watt PSUs. 

 

This is not how professionals are supposed to work. I just need some sort of resource that tells what a content creation PC is supposed to be so I can show my leadership how much they are screwing us. We are doing extremely important work and everybody should be concerned that our systems are holding us back.

[Marinatall_Ironside]

22 minutes ago, Nesagsar said:

I certainly agree that graphics shops all have different requirements but I would argue that there are some general items that we can all agree on. Our computers are doing okay on CPU and RAM with Xeons and 32 GB of DDR3. However - they put NVS 510 graphics in there, the HDD isnt in RAID and we have bricked three systems because the HDDs had mechanical failures, and these things are running on 350 watt PSUs. 

 

This is not how professionals are supposed to work. I just need some sort of resource that tells what a content creation PC is supposed to be so I can show my leadership how much they are screwing us. We are doing extremely important work and everybody should be concerned that our systems are holding us back.

Well if you want an example build of this, I'll be happy to oblige. Generally speaking, you want to have RAID 1 for the HDDs, you want to use higher-than-average wattage power supplies of much higher quality (Seasonic is the go-to on this one), and have a strong GPU with enough VRAM. The rest of the hardware should be strong enough to handle the workloads too, and therefore the CPU and RAM shouldn't be overlooked either.

 

Using this philosophy I've outlined above, here's what a high-performance workstation that is extremely reliable, robust in expansion and I/O connectivity, high-performance for the applications being used, and has an upgrade path for the future using the most current architectures for their respective components, look like, with AMD in mind.

 

PCPartPicker part list / Price breakdown by merchant

CPU: AMD - Threadripper 1950X 3.4GHz 16-Core Processor  ($1079.98 @ Amazon) 
CPU Cooler: Enermax - LiqTech TR4 360 102.2 CFM Liquid CPU Cooler 
Motherboard: Asus - ROG ZENITH EXTREME EATX TR4 Motherboard  ($593.98 @ Amazon) 
Memory: G.Skill - Ripjaws V Series 64GB (4 x 16GB) DDR4-3200 Memory  ($734.38 @ Newegg) 
Storage: Samsung - 960 PRO 512GB M.2-2280 Solid State Drive  ($302.38 @ Amazon) 
Storage: Samsung - 960 PRO 512GB M.2-2280 Solid State Drive  ($302.38 @ Amazon) 
Storage: Samsung - 960 Pro 1TB M.2-2280 Solid State Drive  ($626.38 @ Amazon) 
Storage: Seagate - IronWolf 10TB 3.5" 7200RPM Internal Hard Drive  ($365.58 @ Amazon) 
Storage: Seagate - IronWolf 10TB 3.5" 7200RPM Internal Hard Drive  ($365.58 @ Amazon) 
Video Card: NVIDIA - Titan Xp 12GB Video Card  ($1296.00) 
Case: Corsair - Carbide Clear 600C ATX Full Tower Case  ($151.08 @ Amazon) 
Power Supply: SeaSonic - PRIME Platinum 1200W 80+ Platinum Certified Fully-Modular ATX Power Supply  ($268.92 @ Amazon) 
Operating System: Microsoft - Windows 10 Pro Full - USB 32/64-bit  ($205.18 @ Newegg) 

Base Total: $5921.82

Tax (8%): $370.00

Total: $6291.82
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2017-08-13 00:11 EDT-0400

 

The point of this build is to be extremely reliable, robust in expansion and I/O connectivity, high-performance for the applications being used, and has an upgrade path for the future. Need more storage? No problem. More RAM? Sure, go for it. Want to slam 3 more GPUs? The board and power supply has got you covered.

 

HDDs should be configured in RAID 1 for redundancy (while sacrificing total storage). If you want to do RAID 10, then you'll want to toss in two more 10TB HDDs for 20TB maximum unformatted space for mass storage (due to the mirroring of the two drives before being striped to the other two).

 

For the NVMe SSDs, you can't make a bootable NVMe RAID, but you can make a RAID array for use on active projects and act as a high-speed scratch disk solution. The 1TB NVMe drive is for the OS and all of the applications to be used. The HDDs act as the storage for completed projects, that's at least how I would use this kind of computer.

 

For the Nividia Titan Xp's, Nvidia released the prosumer drivers, which reportedly gives a 3x boost in content creation applications like Autodesk 3ds Max and Maya for example.

 

I have a similar parts list like this one saved that is much more ridiculous (4 SSDs in RAID on a special add-in card and two Nvidia Titan Xp's), but I dialed it back so it is more practical in a sense.

[mariushm]

I actually have to disagree with you on that build.

 

Depending on the software packages, user may want to go with Threadripper 1920x (12 core / 24 thread) and just overclock all cores to 3.9-4 ghz all the time.

No need for a 550$ board, can go with a 440$ board like Asrock x399 professional and get all those features. Why would you pay extra for 802.11 ad for example, which dies a few feet from your desk or when you place a sheet of paper between the pc and the router?

You're adding 3 x samsung 960 pro for storage ... no boot drive

I'd rather have a <=512 GB boot drive for just the OS, and buy a SSD which uses MLC or even SLC memory for more endurance... you only need to store the OS and some applications and maybe

Yeah the 960 Pro uses MLC but it's also expensive... normally it's 330$, now i see it at 280$ for the 512 GB version. 

I was looking at a much cheaper Kingston KC1000 480 GB SSD for 250$.  Yeah the Samsung 960 Pro can do  3.5 GB/s read / 2.1 GB write while the Kingston can only do 2.7 GB/s read / 1.6 GB write but nobody cares, it's not like you're constantly loading applications from the SSD which are hundreds of megabytes in size.... the sequential speed just doesn't matter for a boot drive.  The Kingston is rated for 550 TB writes for the 480 GB ssd, while the Samsung is only rated for 400 TB for the 512 GB drive, and the Kingston also has a 2 mil hours mtbf compared to Samsung's 1.5 mil hours. 

 

For actual storage, do you really care if you read at 600 MB/s or at more? Do you actually have workloads that have to sustain more than 600 MB/s sustained? Unless you work with raw 4K / Prores footage / something like that (and even that is usually something like 500-800 mbps) maybe you'd be better off to just go with regular SSD instead of squeezing 1 TB or more into a m.2 footprint. Better thermals, higher nm MLC chips which means more endurance etc etc

 

But yeah, if you want you can definitely put two additional SSDs in RAID 1 for redundancy or whatever and use them as scratch disks for editing

But maybe it would be a better deal to just fill the system with ram and go with 128 GB of memory and just create a ram disk of let's say 96 GB , and when you work on a project, just copy the files from SSD to ram and work on stuff... just make sure you have an UPS and do backups to disk from time to time.

 

And yeah, you'd want to go with ECC unbuffered memory, not regular desktop memory, if you're going to build such a system ... ex go with kingston valueram 16 GB 2400 Mhz ECC ubuffered for 170$ per stick since you're not really going to get much higher than 2400 Mhz if you're going to fill all 8 memory slots.

 

 

Seagate ironwolf 10tb ... bleah.. rather go with hgst nas or wd helium drives

 

and 1200w psu? really not needed. Even a 1950x overclocked doesn't use more than 300 watts. Unless you plan to install more than 2 graphics cards, you don't really need 1200 watts.

[Nesagsar]

This is why I figured a "Fast as Possible" would work better than a build guide. I dont accomplish the task by recommending specific hardware, what I need is an explanation for why better hardware is necessary. Essentially, run through each of the components and explain why professionals need better components than business users and gamers. Explain why high core counts are better than clock speed when multitasking. Explain why pros need lots of RAM due to RAM enabled preview in linear editors. Explain why RAID is important to data security. Essentially the same thing as the "What is the difference between a PC and a server?" video but for workstations instead.