Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
 
  • Post
  • Reply
Anarchist Mae
Nov 5, 2009

by Reene


Lipstick Apathy

eames posted:

Yes but I'd argue that big hook is X299/KBL-X/SL-X

There are people on the AnandTech forums who think it's superior to Threadripper. They cite poor single threaded performance mainly.

Adbot
ADBOT LOVES YOU

Anime Schoolgirl
Nov 28, 2002






Measly Twerp posted:

There are people on the AnandTech forums who think it's superior to Threadripper. They cite poor single threaded performance mainly.
A category rapidly losing its relevance because it's beyond the point of diminishing returns outside of special snowflake cases and lovely adobe software

Cygni
Nov 12, 2005

raring to post



honestly the 8 core TR is the least interesting of the bunch to me. its 8% cheaper than the 7820X that has true quad channel memory and the same number of cores with better IPC and turbo clocks, and 10% more expensive (plus a more expensive motherboard) than the 1800X that will probably match it in performance for everything thats not extremely memory bandwidth hungry and NUMA friendly.

it has a place if you desperately need PCIe lanes, but with crossfire/SLI kinda dying out in games, thats mostly gonna be NVMe raid people, multiple NICs, machine learning, or other very specific niches.

Arzachel
May 12, 2012


SwissArmyDruid posted:

...no, no, I call bullshit. There is *no* goddamn way they are pricing TR _THAT_ loving low! Goddamnit, AMD, what happened to not leaving money on the table anymore!? You don't have to do that in order to make Intel hurt anymore, your product is-

*noise of disgust*

Fine. Is that the bottom end of the stack? Are we only getting three variants of TR?

A wafer unused is a wafer you still have to pay GF for, everything must go!

Besides, it makes sense to get everyone on the platform if they plan to provide an upgrade path.

Scarecow
May 20, 2008

3200mhz RAM is literally the Devil. Literally.


Lipstick Apathy

Cygni posted:

honestly the 8 core TR is the least interesting of the bunch to me. its 8% cheaper than the 7820X that has true quad channel memory and the same number of cores with better IPC and turbo clocks, and 10% more expensive (plus a more expensive motherboard) than the 1800X that will probably match it in performance for everything thats not extremely memory bandwidth hungry and NUMA friendly.

it has a place if you desperately need PCIe lanes, but with crossfire/SLI kinda dying out in games, thats mostly gonna be NVMe raid people, multiple NICs, machine learning, or other very specific niches.

what on earth do you mean by "true quad channel memory"

Cygni
Nov 12, 2005

raring to post



Scarecow posted:

what on earth do you mean by "true quad channel memory"

My understanding of Intel's mesh arch from STH articles is that there are 2 memory controllers, 2 channel each ( i think they are actually 3 wide in hardware, but only 2 enabled in SLX vs Xeon?) on opposite "sides" of the mesh, but all cores can get the full bandwidth. In TR/Epyc, each die has its own local 2 channel memory controller, but dies can reach across IF to access the other banks local to other dies. In STH's testing, it was a significant bandwidth and latency hit to reach across IF. So in reality, any one thread/die cant get all of the quad-channel bandwidth. At best, it can get 100% of its local, and about 60% of the other dies'.

I honestly dont know how much that will matter in the real world, and i imagine in some workloads that distribute well and play well with NUMA, it wouldnt be an issue. STH were pretty high on their Epyc test servers, and Intel's memory design make compromises in other places. Guess we will have to see when real stuff shows up.

e: btw as another caveat, i probably dont know what im talking about. i just like talkin bout puters my dudes.

Cygni fucked around with this message at 07:23 on Jul 31, 2017

MaxxBot
Oct 6, 2003

you could have clapped

you should have clapped!!


FaustianQ posted:

On AMD's own facebook page



I'm dying, this is loving absurd.

Ahahahaha that's amazing

MaxxBot
Oct 6, 2003

you could have clapped

you should have clapped!!


Measly Twerp posted:

There are people on the AnandTech forums who think it's superior to Threadripper. They cite poor single threaded performance mainly.

It would have to be non-gaming single-threaded applications since the L3 cache structure of the Skylake-X is currently having issues with games and as a consequence it's no better than Ryzen even with a big clock advantage. The cache restructuring and possibly the mesh bus has totally ate away at the potential advantages that Skylake-X would have over Ryzen and Threadripper. Coffee Lake seems more compelling if the rumors are true.

MaxxBot fucked around with this message at 08:08 on Jul 31, 2017

Scarecow
May 20, 2008

3200mhz RAM is literally the Devil. Literally.


Lipstick Apathy

Newegg has prices up for Ripper boards, looks like 3200mhz is gonna be the max OC the boards will claim which I'm a little disappointed in

Anarchist Mae
Nov 5, 2009

by Reene


Lipstick Apathy

Cygni posted:

honestly the 8 core TR is the least interesting of the bunch to me. its 8% cheaper than the 7820X that has true quad channel memory and the same number of cores with better IPC and turbo clocks, and 10% more expensive (plus a more expensive motherboard) than the 1800X that will probably match it in performance for everything thats not extremely memory bandwidth hungry and NUMA friendly.

Don't forget that the 1900X has an all core base clock of 3.8GHz which is shockingly high for Ryzen, so out of the box it's basically running at what many people are overclocking their 1700's to. There will be a cheaper 8c/16t part, the 1900.

This I think makes Threadripper at quite competitive with the 7820X, especially if you'd rather spend that 8% somewhere else in the system (someone is going to spend it on RGB lighting and now I feel bad).

MaxxBot posted:

It would have to be non-gaming single-threaded applications since the L3 cache structure of the Skylake-X is currently having issues with games and as a consequence it's no better than Ryzen even with a big clock advantage. The cache restructuring and possibly the mesh bus has totally ate away at the potential advantages that Skylake-X would have over Ryzen and Threadripper. Coffee Lake seems more compelling if the rumors are true.



There may have been some decent arguments over there against these benchmarks, unfortunately it got entirely lost in a sea of the stupidest bullshit I've ever had to read. I think the most compelling argument was that these benchmarks are just GPU bottlenecked and therefore irrelevant. Which personally I think is bullshit.

Over on the AnandTech forums certain members wanted to see these same tests repeated at a lower resolution in order to "gauge future performance". This I think is based on the faulty assumption that games will continue to utilise available CPU resources in the exact same way in the future. Often the failure of Bulldozer to influence game developers to do so is brought up as part of this assertion, however with Ryzen prompting Intel to finally introduce 6 cores to the mainstream with Coffee Lake, and with AMD themselves developing partnerships that have already brought performance improvements to existing games, I don't think it holds up at all.

Reviewers have even done that exact kind of testing in the past, however to the best of my knowledge none have ever followed it up and demonstrated that it had any predictive value what so ever which is a huge problem since people constantly bring it up as if it has.

We should be testing games for the performance you can reasonably expect to get today, with the graphics fidelity that people exepct to enjoy today. To the best of my understanding that means people want to see 1080p Ultra and sometimes 4k too.

If that means we're testing CPUs on a 1080ti and they're still GPU bottlenecked in some games then so be it. That's more important to know than how the same game performs at 720p, especially if you're buying on a budget and looking to get the most cost effective computer for playing the games you want to play.

eames
May 9, 2009



Measly Twerp posted:

There are people on the AnandTech forums who think it's superior to Threadripper. They cite poor single threaded performance mainly.

They're not wrong but the Intel consumer CPUs (7700K/8700K) are much better value for single threaded performance up to 6 cores and Ryzen/TR is cooler/better value at 8 cores and beyond.

I guess that leaves the i9-7820X for those who need good ST performance, don't care about high FPS, can't live with 6 cores but are satisfied with 8 cores. Pretty niche if you ask me. Not to mention potential upgrade paths and goodies like ECC.

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


Wait are people making the argument that you should sink nearly triple the cost into a system for futureproofing in gaming on an explicitly workstation platform? Hahahaha, loving seriously?

eames
May 9, 2009



Ryzen/Threadripper is shipping with revision "ZP-B1", Epyc is shipping with a newer "ZP-B2" revision.
Looks like there's some truth to the B2 rumors after all. I expect that the newer rev will tricke down to consumer/prosumer eventually but it's interesting that there's a difference at all.

http://www.planet3dnow.de/cms/33149...ter-epyc-zp-b2/

Anarchist Mae
Nov 5, 2009

by Reene


Lipstick Apathy

FaustianQ posted:

Wait are people making the argument that you should sink nearly triple the cost into a system for futureproofing in gaming on an explicitly workstation platform? Hahahaha, loving seriously?

I'm not sure that the attitude was that you should buy the 7800X over anything else for gaming, however let me quote the argument against the Hardware Unboxed benchmarks that I was addressing:

MajinCry posted:

If you want to test the prowess behind a CPU, the games must have all GPU-related settings turned to minimum. Texture detail, polygon count, shadow resolution, shaders, and cloth physics (if done on GPU).

Then the CPU-related settings maxed. Shadow distance, draw distance, and the number of lights.

Otherwise, it's a meaningless CPU test, as the entire test itself is buggered.

I think it's fairly obvious that they didn't like that the 1600X was performing so well in these benchmarks and wanted to find some way of invalidating the benchmarks. Why? I don't know.

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


Measly Twerp posted:

I'm not sure that the attitude was that you should buy the 7800X over anything else for gaming, however let me quote the argument against the Hardware Unboxed benchmarks that I was addressing:


I think it's fairly obvious that they didn't like that the 1600X was performing so well in these benchmarks and wanted to find some way of invalidating the benchmarks. Why? I don't know.

See this poo poo always irks me because I really don't care about theoretical performance but practical performance. Who gives a poo poo if the 7800X is better in an entirely contrived scenario, it's not practically different. It's basically the "Vega is a 1080ti competitor at 8K" of the CPU world.

SoftNum
Mar 31, 2011



Measly Twerp posted:

I think it's fairly obvious that they didn't like that the 1600X was performing so well in these benchmarks and wanted to find some way of invalidating the benchmarks. Why? I don't know.

I'm mostly curious what this person thinks these have to do with CPU performance.

Combat Pretzel
Jun 23, 2004

No, seriously... what kurds?!

eames posted:

Ryzen/Threadripper is shipping with revision "ZP-B1", Epyc is shipping with a newer "ZP-B2" revision.
Looks like there's some truth to the B2 rumors after all. I expect that the newer rev will tricke down to consumer/prosumer eventually but it's interesting that there's a difference at all.

http://www.planet3dnow.de/cms/33149...ter-epyc-zp-b2/
How am I supposed to interpret that? TR has the B1 die? It will keep it through out its lifecycle?

--edit: I sure hope it gets updated. I have to sit it out for GeIL's 3200MHz ECC modules, anyway.

Combat Pretzel fucked around with this message at 16:07 on Jul 31, 2017

eames
May 9, 2009



Combat Pretzel posted:

TR has the B1 die?

yes

Combat Pretzel posted:

It will keep it through out its lifecycle?

nobody knows; it could be that B1 has some (minor?) errata that disqualifies it for server applications and they're offloading old dies via Desktop/HEDT

Combat Pretzel posted:

--edit: I sure hope it gets updated. I have to sit it out for GeIL's 3200MHz ECC modules, anyway.

same although I don't know what advantages B2 would bring

Combat Pretzel
Jun 23, 2004

No, seriously... what kurds?!

Last the B2 stepping came up, things about PCIe/uncore. Apparently there's an issue with FMA3, maybe that got fixed in B2?

SwissArmyDruid
Feb 14, 2014



Combat Pretzel posted:

Last the B2 stepping came up, things about PCIe/uncore. Apparently there's an issue with FMA3, maybe that got fixed in B2?

No, the bug for that was allegedly tied to the firmware itself and was supposed to have been fixed with AGESA 1.0.0.6.

GRINDCORE MEGGIDO
Feb 28, 1985




In the 8core TR, how high is the latency between the dies, vs the CCX latency on Ryzen?

Combat Pretzel
Jun 23, 2004

No, seriously... what kurds?!

Based on ServeTheHome's test with EPYC and I think DDR4-2400 RAM, the inter-CCX latency was 80ish ns while inter-die was 130ish ns.

Scarecow
May 20, 2008

3200mhz RAM is literally the Devil. Literally.


Lipstick Apathy

God i loving hate being in australia. The retailer's here are trying to charge a thousand didgerydoos for a board that newegg has listed for 550usd

And people wonder why brick and mortar stores here are loving dieing off

KillaZilla
Jan 30, 2003


Scarecow posted:

God i loving hate being in australia. The retailer's here are trying to charge a thousand didgerydoos for a board that newegg has listed for 550usd

And people wonder why brick and mortar stores here are loving dieing off

What board are you looking at? I just saw the absurd price of the Asus Zenith at PCCG and surely they or the distributor have to know people are just going to import them and avoid the $300+ mark up. No doubt the retailers will all be crying to the government again in a few months over how internet shopping is the devil and not their stupid prices.

GRINDCORE MEGGIDO
Feb 28, 1985




Combat Pretzel posted:

Based on ServeTheHome's test with EPYC and I think DDR4-2400 RAM, the inter-CCX latency was 80ish ns while inter-die was 130ish ns.

Thanks man. It's going to be bouncing off that a lot, I guess

Scarecow
May 20, 2008

3200mhz RAM is literally the Devil. Literally.


Lipstick Apathy

KillaZilla posted:

What board are you looking at? I just saw the absurd price of the Asus Zenith at PCCG and surely they or the distributor have to know people are just going to import them and avoid the $300+ mark up. No doubt the retailers will all be crying to the government again in a few months over how internet shopping is the devil and not their stupid prices.

Yeah im looking at the zenith and yup 750aud delivered from amazon has me flipping the bird at stores here

Anarchist Mae
Nov 5, 2009

by Reene


Lipstick Apathy

Scarecow posted:

Yeah im looking at the zenith and yup 750aud delivered from amazon has me flipping the bird at stores here

Holy poo poo.

They're not even available yet in this European backwater with 27% VAT, but I'm pretty sure when it is it will be cheaper than that.

eames
May 9, 2009



579 (incl. 20% vat) over in Germany.

https://geizhals.eu/?cat=mbtr4

Anarchist Mae
Nov 5, 2009

by Reene


Lipstick Apathy

I've just ordered one of these:



Overkill a little?

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


Combat Pretzel posted:

Based on ServeTheHome's test with EPYC and I think DDR4-2400 RAM, the inter-CCX latency was 80ish ns while inter-die was 130ish ns.

No this was for 2667, 2400 it was ~150 https://www.servethehome.com/amd-ep...666-a-snapshot/ (thanks Cygni)

It's about an 8% improvement moving from DDR4 2400 to 2667, so I'm guessing you'll get a 10-15% further improvement moving to DDR4 3000-3200, or about 110-120ns for inter-CCX latency. IIRC, this is about what Intel's general latency is.

wargames
Mar 16, 2008

official yospos cat censor


FaustianQ posted:

No this was for 2667, 2400 it was ~150 https://www.servethehome.com/amd-ep...666-a-snapshot/ (thanks Cygni)

It's about an 8% improvement moving from DDR4 2400 to 2667, so I'm guessing you'll get a 10-15% further improvement moving to DDR4 3000-3200, or about 110-120ns for inter-CCX latency. IIRC, this is about what Intel's general latency is.

https://www.youtube.com/watch?v=6laL-_hiAK0

intel seems to be around 80 atleast with the 5960x I don't know what the new intel mesh is at.

Tindahbawx
Oct 14, 2011




What's the point in that?

Anarchist Mae
Nov 5, 2009

by Reene


Lipstick Apathy

wargames posted:

https://www.youtube.com/watch?v=6laL-_hiAK0

intel seems to be around 80 atleast with the 5960x I don't know what the new intel mesh is at.

According to Toms Hardware the new i9 7900 is worse only compared to other Intel products. It averages at around 75ns, where as the previous HEDT processor averaged at 65, faster ram makes things worse and both generations of HEDT are worse than their regular desktop processors.

ufarn
May 30, 2009


Measly Twerp posted:

Holy poo poo.

They're not even available yet in this European backwater with 27% VAT, but I'm pretty sure when it is it will be cheaper than that.
EU is expensive vs US, but we've got nothing on Oz.

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


Measly Twerp posted:

According to Toms Hardware the new i9 7900 is worse only compared to other Intel products. It averages at around 75ns, where as the previous HEDT processor averaged at 65, faster ram makes things worse and both generations of HEDT are worse than their regular desktop processors.

So basically AMD needs about DDR4 4000 to be close but not on par for inter-ccx latency, or to run the IF at 2:1 divider or asynchronously. Wonder why higher ram clocks hurt Intel though?

Watermelon Daiquiri
Jul 10, 2010




latency matters more?

BangersInMyKnickers
Nov 3, 2004

I have a thing for courageous dongles



Cygni posted:

My understanding of Intel's mesh arch from STH articles is that there are 2 memory controllers, 2 channel each ( i think they are actually 3 wide in hardware, but only 2 enabled in SLX vs Xeon?) on opposite "sides" of the mesh, but all cores can get the full bandwidth. In TR/Epyc, each die has its own local 2 channel memory controller, but dies can reach across IF to access the other banks local to other dies. In STH's testing, it was a significant bandwidth and latency hit to reach across IF. So in reality, any one thread/die cant get all of the quad-channel bandwidth. At best, it can get 100% of its local, and about 60% of the other dies'.

I honestly dont know how much that will matter in the real world, and i imagine in some workloads that distribute well and play well with NUMA, it wouldnt be an issue. STH were pretty high on their Epyc test servers, and Intel's memory design make compromises in other places. Guess we will have to see when real stuff shows up.

e: btw as another caveat, i probably dont know what im talking about. i just like talkin bout puters my dudes.

This kind of penalty has always existed on the Intel side, they've just been sorta covering it up by running the inter-node communications async and really fast and basically masking the memory controller layout from software to give consistent latencies and an easier software optimization target at the penalty of some performance. The Cluster-on-Die mode takes all that away and exposes the two numa nodes with their memory controllers so applications that are numa-aware can optimize against them and get improved latency/bandwidth if the workload can take advantage of it.

http://en.community.dell.com/techce...-on-vmware-esxi

Combat Pretzel
Jun 23, 2004

No, seriously... what kurds?!

wargames posted:

https://www.youtube.com/watch?v=6laL-_hiAK0

intel seems to be around 80 atleast with the 5960x I don't know what the new intel mesh is at.
Yea, I recently tested my 5820K at 2400MHz with tight timings, and it gets like 83ns. To me, the inter-die memory latency is OK, given the circumstances. I'd rather have none, but alas.

Ihmemies
Oct 6, 2012



So 1700X is 600-700 cheaper than 1920X with a mobo... 150 per core is a bit meh price.

Adbot
ADBOT LOVES YOU

Anarchist Mae
Nov 5, 2009

by Reene


Lipstick Apathy

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply