AMD Ryzen Threadripper and Ryzen 3 Product Updates (Threadripper price reveal)

[tom_w141]

4 minutes ago, VagabondWraith said:

The 7800X still has more L3 than the 7700K

0.25 more. Less per core. There has to be a reason for the 7800X under performing. Guess we have to wait for more people to test that 1

 

[VagabondWraith]

Just now, tom_w141 said:

0.25 more. Less per core. There has to be a reason for the 7800X under performing. Guess we have to wait for more people to test that 1

 

I just think it's such a rushed launch that performance is not up to par rn. PCGamer stated that Turbo Boost was broken at first with the 7900X only hitting 4.0 on all cores. I think it will take time for things to improve much like Ryzen is continually improving. ?

[VagabondWraith]

4 minutes ago, juri-han said:

not much good if it doesn't show the lows 

True. It will take some time as more reviews touch upon platform improvements down the line. I can say from experience that my 7820X at 4.5 with cache at 2800MHz that gaming is truly fluid. But I would wait to see what Coffee Lake will bring as I hope Intel will price the product much better and the platform isn't rushed.

[themctipers]

7 hours ago, TVwazhere said:

60 PCI-E lanes got me excited

A L L  T H E  E X P A N S I O N  S L O T S 

my p55 motherboard has less PCI slots than my Z77 motherboard

[AnonymousGuy]

11 minutes ago, juri-han said:

not much good if it doesn't show the lows 

Low doesn't really matter much if it's an outlier measurement that only happens for 1/60th of a second one time in 5 minutes.  You won't notice it.

 

Std. deviation or 99.9th percentile would be far more useful of a metric.

[willsamuels]

Sure is cool

[Drak3]

@leadeater Keep in mind, AVX/AVX2 gets a boost from memory bandwidth. And what does EPYC have over Xeon? Octochannel memory, or 8 channels of memory goodness for you heathens that don't like using cool prefixes. Whereas SLSP has Hexachannel, 6, BWEP has Quad channel.

When using AVX, Ryzen has 1/2 the resources available to it compared to a Haswell, Broadwell, Skylake, or Kabylake core. Or as I see it, Haswell, Mini Haswell, Mini Haswell 2.0, Mini Haswell 2.0.1. So it cannot possibly match an equivalent clock of Haswell+ without taking 2 cores to match one of Intel's.

When using AVX2, Ryzen has the same resources available, but the IMC gets in the way, and strangles performance to the point where, again, 2 AMD cores approximate to being 1 Intel core of similar clockspeed with Ryzen.

 

I don't expect this to change moving from dual channel to quad channel. I fully expect the same old song and dance with X299 vs X399.

 

Hell, Reading through the Anantech article you posted, it's clear why EPYC was dynamite in the first two benches - they're running off of cache predominately or exclusively. They're not being held back.

In the updated NAMD run, the dual 8176 rig hit 82% what the dual EPYC 7601. It also has 88% of the core count, and after multiplying core count and clock speed, you've got a "total clock speed" of 140.8GHz for EPYC dual and 117.6GHz for Xeon Plat dual. Intel's system is approximately 83% the "total clock speed," and we can assume the IPC advantage is lost to scaling never being quite 100%.

It takes quite a bit to get a Zen core to perform as it should across the board, relative to theory and Intel's performance.

 

And my stance on Threadripper remains unchanged. It'll have limited value, and where it shines isn't good enough for the target HEDT user.

 

[Brooksie359]

3 hours ago, SpaceGhostC2C said:

No. I mean, yes, that's the approach a company deciding over buying a computer vs. not buying that computer will take. That is not, however, the approach to make a decision between buying individual component A vs. individual component B.

 

So, full price comparisons are relevant for companies buying black boxes, and never upgrading them. For them, it is the price of Black Box 1, for performance A, vs the price of Black Box 2, for performance B. You can argue that's the situation for many companies. However, if we focus on those, then the CPUs' MSRP is irrelevant: they only care about the price Dell, HP, etc charge for their black boxes, which is based on their own markups, the actual demand they face, whatever marketing factor they can add to each model, the deals they sign with their own suppliers, and the wholesale prices they get on hardware, which need not be in line with retail MSRP.

The retail price of a CPU only matters for buyers of individual components, and for them the "right" analysis is in terms of marginal costs of said components, not total prices.

The is a reason why they do analysis that way and it is because that way you get the actual price to performance of the pc as a whole and not an individual part. I mean I love amd and would still recommend their cpus most of the time but to say that the price of the other components in the pc doesn't matter is just plain ridiculous. I mean if you are just replacing the cpu motherboard and ram then the comparison you talked about makes more sense but that isn't always the case. If we just went by pure price to performance of a cpu and motherboard combo then everyone would have a g4560. It's simple economics that you include the whole price of the build when comparing one potential build with another. Now this is when comparing potential builds so there are some cases where the intel option might make more sense in a specific build. I guess what I'm trying to say is that although for the most part the amd cpus are going to make more sense price to performance wise, there are also situations where that won't be true.

[Brooksie359]

10 minutes ago, Drak3 said:

@leadeater Keep in mind, AVX/AVX2 gets a boost from memory bandwidth. And what does EPYC have over Xeon? Octochannel memory, or 8 channels of memory goodness for you heathens that don't like using cool prefixes. Whereas SLSP has Hexachannel, 6, BWEP has Quad channel.

When using AVX, Ryzen has 1/2 the resources available to it compared to a Haswell, Broadwell, Skylake, or Kabylake core. Or as I see it, Haswell, Mini Haswell, Mini Haswell 2.0, Mini Haswell 2.0.1. So it cannot possibly match an equivalent clock of Haswell+ without taking 2 cores to match one of Intel's.

When using AVX2, Ryzen has the same resources available, but the IMC gets in the way, and strangles performance to the point where, again, 2 AMD cores approximate to being 1 Intel core of similar clockspeed with Ryzen.

 

I don't expect this to change moving from dual channel to quad channel. I fully expect the same old song and dance with X299 vs X399.

 

Hell, Reading through the Anantech article you posted, it's clear why EPYC was dynamite in the first two benches - they're running off of cache predominately or exclusively. They're not being held back.

In the updated NAMD run, the dual 8176 rig hit 82% what the dual EPYC 7601. It also has 88% of the core count, and after multiplying core count and clock speed, you've got a "total clock speed" of 140.8GHz for EPYC dual and 117.6GHz for Xeon Plat dual. Intel's system is approximately 83% the "total clock speed," and we can assume the IPC advantage is lost to scaling never being quite 100%.

It takes quite a bit to get a Zen core to perform as it should across the board, relative to theory and Intel's performance.

 

And my stance on Threadripper remains unchanged. It'll have limited value, and where it shines isn't good enough for the target HEDT user.

 

If they don't have a need for avx or don't use it very often then it will be an easy sell. You fail to realize that people don't just go and buy a cpu of that price without thinking if it is a suitable for their usecase.

[Drak3]

Just now, Brooksie359 said:

If they don't have a need for avx or don't use it very often then it will be an easy sell. You fail to realize that people don't just go and buy a cpu of that price without thinking if it is a suitable for their usecase.

That's the thing though, AVX is significantly more common in the tasks that HEDT is meant and used for. What few tasks that benefit from 6 to 8 cores and not AVX, or have limited benefit (beginner content creators namely), are already capably handled by Ryzen 5/7, the 5820K, 5930K, 6800K, 6850K, 7800X, and 7820X, all of which are less expensive than Threadripper's base offering.

 

So far, nothing has been able to change my view of TR. For the HEDT platform, it just isn't up to snuff.

[XenosTech]

10 minutes ago, Brooksie359 said:

If they don't have a need for avx or don't use it very often then it will be an easy sell. You fail to realize that people don't just go and buy a cpu of that price without thinking if it is a suitable for their usecase.

You'd be surprised at the amount of idiots that would buy x99 over ryzen just to do something as simple as video rendering with adobe. I recently got kicked out of a group becasue I pointed out x99 is bad value for a guy who wanted to build a BUDGET video editing rig.

[Drak3]

Just now, XenosTech said:

You'd be surprised at the amount of idiots that would buy x99 over ryzen just to do something as simple as video rendering with adobe. I recently got kicked out of a group becasue I pointed out x99 is bad value for a guy who wanted to build a BUDGET video editing rig.

Adobe's software is AVX enabled. Something like the 5820K/6800K/7800X, which isn't much more expensive than the 1700 or 7700K, will outperform either, and the end cost won't be the astronomic thing that people make Intel HEDT out to be.

 

Hell, the 7700K with a good OC makes for a better choice in a video editing rig if your soft is AVX enabled.

[XenosTech]

2 minutes ago, Drak3 said:

Adobe's software is AVX enabled. Something like the 5820K/6800K/7800X, which isn't much more expensive than the 1700 or 7700K, will outperform either, and the end cost won't be the astronomic thing that people make Intel HEDT out to be.

 

Hell, the 7700K with a good OC makes for a better choice in a video editing rig if your soft is AVX enabled.

Where I live none of those things are cheap hence budget builds... He ended up getting the 1600 and a gtx 1060 and used cuda since he edits with premier pro. Was pointless telling him to buy anything from x99 since it doesn't sell here either way

[PCGuy_5960]

6 minutes ago, XenosTech said:

I recently got kicked out of a group becasue I pointed out x99 is bad value for a guy who wanted to build a BUDGET video editing rig.

That's ridiculous, for a budget editing rig the Ryzen 5 1600 is the go-to CPU

[XenosTech]

Just now, PCGuy_5960 said:

That's ridiculous, for a budget editing rig the Ryzen 5 1600 is the go-to CPU

I said that and I got kicked out.. they were telling to buy things like the 6800k or the 7800x which are well over $1000 here and saying it was good value .-. Like it's one thing to like a company's products but it's another to be dick riding them so hard you can't clearly see that they're bad value when compared to what's on the market.

[TVwazhere]

8 minutes ago, XenosTech said:

I said that and I got kicked out.. they were telling to buy things like the 6800k or the 7800x which are well over $1000 here and saying it was good value .-. Like it's one thing to like a company's products but it's another to be dick riding them so hard you can't clearly see that they're bad value when compared to what's on the market.

Was the group name "Intel or die"? Cause it sounds like it

[XenosTech]

Just now, TVwazhere said:

Was the group name "Intel or die"? Cause it sounds like it

Nah a gaming group for the local gaming clans to organise lan parties

[TVwazhere]

4 minutes ago, XenosTech said:

lan parties

Intelan

There's your problem! :^)

[XenosTech]

3 minutes ago, TVwazhere said:

Intelan

There's your problem! :^)

 

[SpaceGhostC2C]

45 minutes ago, Brooksie359 said:

The is a reason why they do analysis that way and it is because that way you get the actual price to performance of the pc as a whole and not an individual part.

Yes, and the reason is the one I stated.

 

45 minutes ago, Brooksie359 said:

I mean I love amd and would still recommend their cpus most of the time but to say that the price of the other components in the pc doesn't matter is just plain ridiculous.

No. You just don't understand the difference between thinking at the margin and being confused by sunk costs. I explained when system price vs component price is used.

 

45 minutes ago, Brooksie359 said:

I mean if you are just replacing the cpu motherboard and ram then the comparison you talked about makes more sense but that isn't always the case.

Yes, that's why I covered both cases.

45 minutes ago, Brooksie359 said:

If we just went by pure price to performance of a cpu and motherboard combo then everyone would have a g4560.

No, that is not true. The cheapest CPU won't magically turn out to be the optimal purchase for everyone just because they think at the margin.

45 minutes ago, Brooksie359 said:

It's simple economics that you include the whole price of the build when comparing one potential build with another.

No, it is actually comparing marginal benefits and marginal costs that forms the most basic building block of economics.

45 minutes ago, Brooksie359 said:

Now this is when comparing potential builds so there are some cases where the intel option might make more sense in a specific build. I guess what I'm trying to say is that although for the most part the amd cpus are going to make more sense price to performance wise, there are also situations where that won't be true.

I agree 100% that sometimes one option will be the best choice, sometimes the other option will be the best choice (that's also why I disagree about the Pentium argument). What started this exchange was the difference between two ways of reasoning, only one of which is "rational" in a technical sense (i.e., one that will lead to maximizing own's benefit given constraints); it was never implied that the "correct" reasoning can't lead to different choices in different situations.

(also, notice that a company who buys fully built machines is also "thinking at the margin", since for them there is no divisibility between components, just black box A or black box B. Anyone building their own computer would be highly mislead by looking at full build prices when choosing each individual component, though).

[Brooksie359]

12 minutes ago, SpaceGhostC2C said:

Yes, and the reason is the one I stated.

 

No. You just don't understand the difference between thinking at the margin and being confused by sunk costs. I explained when system price vs component price is used.

 

Yes, that's why I covered both cases.

No, that is not true. The cheapest CPU won't magically turn out to be the optimal purchase for everyone just because they think at the margin.

No, it is actually comparing marginal benefits and marginal costs that forms the most basic building block of economics.

I agree 100% that sometimes one option will be the best choice, sometimes the other option will be the best choice (that's also why I disagree about the Pentium argument). What started this exchange was the difference between two ways of reasoning, only one of which is "rational" in a technical sense (i.e., one that will lead to maximizing own's benefit given constraints); it was never implied that the "correct" reasoning can't lead to different choices in different situations.

(also, notice that a company who buys fully built machines is also "thinking at the margin", since for them there is no divisibility between components, just black box A or black box B. Anyone building their own computer would be highly mislead by looking at full build prices when choosing each individual component, though).

Yes you should always look at each individual component but at the end of the day you are spending x amount of money for a pc giving you x amount of performance and if you choose good parts you should be getting a good price to performance. Almost all my economic classes were taught that way because that is what engineers do. They think about how much you are putting in and what type of return you are getting for it. The same applies here too because who really cares about the individual component prices if that leads to you to not buying the more expensive component even though it would result in a way better return on investment. 

[leadeater]

11 hours ago, Drak3 said:

@leadeater Keep in mind, AVX/AVX2 gets a boost from memory bandwidth. And what does EPYC have over Xeon? Octochannel memory, or 8 channels of memory goodness for you heathens that don't like using cool prefixes. Whereas SLSP has Hexachannel, 6, BWEP has Quad channel.

When using AVX, Ryzen has 1/2 the resources available to it compared to a Haswell, Broadwell, Skylake, or Kabylake core. Or as I see it, Haswell, Mini Haswell, Mini Haswell 2.0, Mini Haswell 2.0.1. So it cannot possibly match an equivalent clock of Haswell+ without taking 2 cores to match one of Intel's.

When using AVX2, Ryzen has the same resources available, but the IMC gets in the way, and strangles performance to the point where, again, 2 AMD cores approximate to being 1 Intel core of similar clockspeed with Ryzen.

 

I don't expect this to change moving from dual channel to quad channel. I fully expect the same old song and dance with X299 vs X399.

 

Hell, Reading through the Anantech article you posted, it's clear why EPYC was dynamite in the first two benches - they're running off of cache predominately or exclusively. They're not being held back.

In the updated NAMD run, the dual 8176 rig hit 82% what the dual EPYC 7601. It also has 88% of the core count, and after multiplying core count and clock speed, you've got a "total clock speed" of 140.8GHz for EPYC dual and 117.6GHz for Xeon Plat dual. Intel's system is approximately 83% the "total clock speed," and we can assume the IPC advantage is lost to scaling never being quite 100%.

It takes quite a bit to get a Zen core to perform as it should across the board, relative to theory and Intel's performance.

 

And my stance on Threadripper remains unchanged. It'll have limited value, and where it shines isn't good enough for the target HEDT user.

 

The problem is you are ignoring evidence provided and going with an assumption. The technical details about the AVX execution units in the Zen architecture is true however you can't simply say it's exactly half the performance of Intel, the performance is different.

 

As you can clearly see in the link I provided the AVX workload on a real workload was faster than Intel's. It's not always faster as they stated highly vectored code will get accelerated much more by Intel's AVX execution units but then you have to start using the same argument people knock back AMD for, you have to optimize your code.

 

That test was compiled on Intel's compiler and optimized for AVX, Intel's AVX.

 

We already know on a per core basis Zen is slower, restating that again doesn't change anything. The actual products that were testing show what performance they get, one is double the CPU cost and performs slower. We can predict that yes the 18 core X299 CPU will perform faster than 16 core X399 CPU but who here was saying that wouldn't be the case?

 

Also FYI your own math shows it's not 2 AMD cores to 1 Intel for that workload.

 

And to the actual point I why I actually posted that, that was showing an AVX workload that isn't "fucking worthless for AVX loads compared to an Intel CPU"

 

Edit:

And for the other tests, why bring those up I didn't mention them at all as they are not AVX workloads.

[SpaceGhostC2C]

Just now, Brooksie359 said:

Yes you should always look at each individual component but at the end of the day you are spending x amount of money for a pc giving you x amount of performance and if you choose good parts you should be getting a good price to performance.

At the end of the day you are spending X, but let's say you are deciding over a GPU: you are spending Y<X anyway, and your true choice is whether to spend Y+A or Y+B, depending on your GPU choice. Hence, the marginal cost in one case is A, and in the other is B. And let's say B is weaker and cheaper than A: then your baseline is Y+B, and the marginal cost is A-B. How much do you get at the margin for paying A-B? That is the economic mindset.

Just now, Brooksie359 said:

Almost all my economic classes were taught that way because that is what engineers do. They think about how much you are putting in and what type of return you are getting for it.

How much you are putting at the margin and what type of marginal return you are getting for it. For "do this project" vs "don't do this project" you look at full cost because all of it is marginal in binary choices. But, believe me, once the project is approved, a proper management will polish every cost on its own merit, from the salaries of the people involved to the coffee machine at the offices: there is no other way to optimize than to do it at the margin. I won't comment on what engineers do or don't, because I don't expect them  to be great economists  But I'm sure every engineer will know a thing or two about constrained optimization and first order conditions, and the English language translation of those is not "look at the total cost bro" but marginal benefit vs marginal costs.

Just now, Brooksie359 said:

The same applies here too because who really cares about the individual component prices

Anyone building their own PC

Just now, Brooksie359 said:

if that leads to you to not buying the more expensive component even though it would result in a way better return on investment. 

If it leads to a way better return on investment, you will buy the more expensive component, as that would be the result of marginal analysis.

[Brooksie359]

4 minutes ago, SpaceGhostC2C said:

At the end of the day you are spending X, but let's say you are deciding over a GPU: you are spending Y<X anyway, and your true choice is whether to spend Y+A or Y+B, depending on your GPU choice. Hence, the marginal cost in one case is A, and in the other is B. And let's say B is weaker and cheaper than A: then your baseline is Y+B, and the marginal cost is A-B. How much do you get at the margin for paying A-B? That is the economic mindset.

How much you are putting at the margin and what type of marginal return you are getting for it. For "do this project" vs "don't do this project" you look at full cost because all of it is marginal in binary choices. But, believe me, once the project is approved, a proper management will polish every cost on its own merit, from the salaries of the people involved to the coffee machine at the offices: there is no other way to optimize than to do it at the margin. I won't comment on what engineers do or don't, because I don't expect them  to be great economists  But I'm sure every engineer will know a thing or two about constrained optimization and first order conditions, and the English language translation of those is not "look at the total cost bro" but marginal benefit vs marginal costs.

Anyone building their own PC

If it leads to a way better return on investment, you will buy the more expensive component, as that would be the result of marginal analysis.

Economics is a huge part of engineering especially in manufacturing because at the end of the day the whole point of engineering is to make money and do it efficiently. Often times you think at margin when replacing a part but not when purchasing a new machine even if you are buying the individual parts. I mean there are also various different methods that one can do economic analysis of these types of things and can end up with different results but doing it as a whole makes the most sense form a individual and business standpoint. 

[Drak3]

2 minutes ago, leadeater said:

Also FYI your own math shows it's not 2 AMD cores to 1 Intel for that workload.

My math is specifically for EPYC, which is the outlier with 8 channel memory, double that of Broadwell-EP and 33% more than that of Skylake-SP. It took that, to get AMD to the point where AVX2 performance is where it should be relative to the competiiton. AMD lifting a bottleneck found on other platforms.

Threadripper and Ryzen are different, they're operating with the same memory configurations as the more powerful Intel cores.

5 minutes ago, leadeater said:

The problem is you are ignoring evidence provided and going with an assumption.

I'm not ignoring the evidence. The evidence is that EPYC performs at the same capacity as Intel in AVX, when put under a more realistic load. However, Ryzen doesn't hold up to that.

http://www.anandtech.com/show/11170/the-amd-zen-and-ryzen-7-review-a-deep-dive-on-1800x-1700x-and-1700/20

In the 'Handbrake and HVEC,' where software is AVX enabled, we see the 1800X hanging out with the 7700K. We're getting damn near 1/2 the performance per core. Hence 2 AMD cores = 1 Intel core.

 

If the 7700K can keep up with the 1800X, on the same memory channel config, why should I expect a 16c Threadripper to outpace the 7820K under the same scenario? Both are on the same underlying architectures, Threadripper has the same IMC as Ryzen. Hell, SLX has an improved IMC that allows for more impressive clocks and timings than previously seen on quad channel configs, that might be a bit of bottleneck reducer as well.

 

And when a workload benefits from better memory bandwidth and latency, such as AVX does, it doesn't take much to put the pieces together. EPYC achieves a superior memory config on the bandwidth front than BWEP or SLSP can. Threadripper and Ryzen do not.