AMD announces open source ray tracing at GDC *Interview update*

[SpaceGhostC2C]

6 minutes ago, leadeater said:

You went from Infiniband to dual Gbe? What? That sounds like 1 step forward 2 steps back lol.

Not that I wanted to  We set up a Debian cluster with Infiniband when I was in grad school, with dual E5440 and E5540 nodes, since a couple of us could benefit from parallelization almost without bound (and the university was buying ). As time went by and we graduated, I would still have access to it from time to time, but later cohorts were not using it much and IT stop paying too much attention to it (I mean, E5540, today... and the electricity costs...). I would have to ask them to turn it on for me, and, after all, It's been years since I left the institution...

 

At some point I built an FX computer at home, swapped some parts around, and eventually decided I could turn the spares and a few purchases from eBay into additional nodes for a home cluster. However, getting 10gig was steep, and Infiniband wasn't even on eBay at all (plus support seems tricky when getting older hardware, where some company acquisitions took place, never intended for home OS, etc). I think it's possible to find some now, but still way up the price ladder compared to a 2- or 4-port Ethernet card.

Overall, it's faster than the old cluster, especially considering I use quite fewer nodes and cores, and that I'm using Windows because laziness (seriously, allowing MPI's smpd through the firewall won't be enough, I have to turn the firewalls completely off or it would still timeout...). Eventually I'll set them up properly, I guess, but right now I have a resubmission to finish  

 

[Doobeedoo]

Yey for open source. Really awesome to see ray tracing focus more. Can't wait to see games with full scene ray tracing one day!

[Razor01]

40 minutes ago, Notional said:

Which is exactly the situation Intel is in. Why do you think their high core count CPU's are so much more expensive than AMD's? It's not just Intel being greedy. And I'm not talking MSRP either. If you are a big customer, you never pay MSRP.

 

And we get a 60% discount on Intel parts, not sure about AMD or even if my IT department is even interested in AMD at this point.  But I can tell you back in the day we always got more discounts from Intel.  They have higher margins.  And leadeater is correct the CPU is actually not that expensive of a part when we are talking about servers.  Memory alone is more expensive even more so if we are talking about dense clusters.  (hardware only, software yeah software can be much more expensive based on needs)

 

About yields, companies will not go into mass production unless yields are over 75% and its probably even higher than this.  What we read about yields on the i-net lol, they are woefully inaccurate.  Semiconductor manufacturing companies have yield model analysis based on their architecture, they can actually predict were errors will show up on what parts of the chip now.  This is why back in the day when they didn't do this, they had redundant parts in the chip, they have dropped that in favor of specific look ups.

 

So they know what the yield of the node is going to be pretty easy right, but the architecture will change the error patterns.  Random errors can still occur, not saying they won't, but they try to limit that as much as possible.

 

Now take a 75% yield for all chips on one wafer, if that is for the top end chip how about the lower end chips, the volume sellers?  What happens to them if yield drops for the full chip?  The hit can be catastrophic for the lower end chips if they need volume. 

 

Lets use a simple example.  i5 vs i7, i5 is just an i7 without hyper threading right?  Are we assuming that the errors when yields are concerned are going to be only in the hyperthreading portions of the silicon?  That is pretty accurate if they can figure out that right?  That isn't the case though.  i5 is a high selling part, actually it is the most volume part of all of the entire Intel line up, its also the part that gives intel's its profits for the most part.

[leadeater]

23 minutes ago, Razor01 said:

Original statement, infinity fabric introduced certain problems from Ryzen and Epyc.  Those problems can't be overlooked when talking about infinity fabric and multiple dies on the same package.  That is what I was getting at.  Infinity fabric is not a cure all for multi dies. 

 

We shouldn't even think of it that way.

 

Just because AMD marketed the hell out of it, as the next big thing. 

 

Infinity fabric what it solves

 

1) Easy of integration of multiple different components as one subsystem

2) Cost savings because of the above

 

That is all it does.  That is all you can ask for an interconnect to do.

 

All the other hub bub about increasing performance, solves deep seeded problems with programmable models.

 

Nada doesn't stop any of that.

 

And then we have the other down side, its harder to program for since its not NUMA aware hardware.

I was responding to your statement that it introduces wild performance loss and is most effected on EPYC. Actual reviews and benchmarks have shown no such thing, other than tests that are designed to specifically show it. There is no 50%-75% performance slow downs in existing software, even on non NUMA aware badly optimized things, we don't even see that on Threadripper.

 

You're talking about a bunch of theoretical issues that were all perpetuated before the release of the products and like everything to do with AMD stick around even when there is direct evidence that it's not true. They were legitimate concerns and I was very interested to see how it played out, luckily for AMD the worst case didn't eventuate.

 

Also you don't really have to program for NUMA unless you're doing OS level stuff; Windows, Linux, ESXi etc are all NUMA aware and will take care of the majority of it for you. EPYC properly presents the NUMA nodes through, it has to. It's not all sunshine and rainbows, for example you don't want to use EPYC for hosting VMs with many vCPUs due to those NUMA boundaries. It's good at hosting many many 1-4 core VMs, decent up to 8 cores but don't do more than that, if you need to go back/stick to Intel.

 

Anyway off topic.

 

Edit:

Spoiler

Quote

root@testsystem:~# lscpu
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 128
On-line CPU(s) list: 0-127
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 2
NUMA node(s): 8
Vendor ID: AuthenticAMD
CPU family: 23
Model: 1
Model name: AMD EPYC 7601 32-Core Processor
Stepping: 2
CPU MHz: 1200.000
CPU max MHz: 2200.0000
CPU min MHz: 1200.0000
BogoMIPS: 4391.67
Virtualization: AMD-V
L1d cache: 32K
L1i cache: 64K
Linux® Network Tuning Guide for AMD EPYC™ Processor
Based Servers
56224 Rev. 1.00 November 2017
8 Optimize Interrupt Handling for Your NIC
L2 cache: 512K
L3 cache: 8192K
NUMA node0 CPU(s): 0-7,64-71
NUMA node1 CPU(s): 8-15,72-79
NUMA node2 CPU(s): 16-23,80-87
NUMA node3 CPU(s): 24-31,88-95
NUMA node4 CPU(s): 32-39,96-103
NUMA node5 CPU(s): 40-47,104-111
NUMA node6 CPU(s): 48-55,112-119
NUMA node7 CPU(s): 56-63,120-127

 

 

[Razor01]

7 minutes ago, leadeater said:

Also you don't really have to program for NUMA unless you're doing OS level stuff; Windows, Linux, ESXi etc are all NUMA aware and will take care of the majority of it for you. EPYC properly presents the NUMA nodes through, it has to. It's not all sunshine and rainbows, for example you don't want to use EPYC for hosting VMs with many vCPUs due to those NUMA boundaries. It's good at hosting many many 1-4 core VMs, decent up to 8 cores but don't do more than that, if you need to go back/stick to Intel.

 

Anyway off topic.

 

 

Last thing, hmm no databases and other programs with NUMA, can't do all that automatically, you must specify what core is doing what and where the data must go to what core to preserve performance.   What NUMA does for current multi CPU's is management of that data transfer can't do that with CCX modules, yet, maybe in the future we will have to see, that is the OS level stuff.

 

This is not an OS level only.  Use it all the time with NUMA, its called CPU pinning and multi core memory caching.  Some times  if you need data for multi cores coming from one core its better to cache that data on all the nodes.  It wastes resources but if performance is the key, it needs to be done.

[leadeater]

1 minute ago, Razor01 said:

Last thing, hmm no databases and other programs with NUMA, can't do all that automatically, you must specify what core is doing what and where the data must go to what core to preserve performance. 

 

This is not an OS level only.  Use it all the time with NUMA, its called CPU pinning and multi core memory caching.

MSSQL is NUMA aware, so is Oracle. MySQL is rather bad at NUMA awareness, hence those sysmark tests being rather bad.

 

You certainly do not under normal situations do CPU pinning or try and override what the OS does with your threads, the OS is already NUMA aware it will put your program thread in the correct place for you. It's very rare you need to override this and if you do you have to be very careful else you'll do a worse job at it compared to the OS.

[leadeater]

41 minutes ago, SpaceGhostC2C said:

and that I'm using Windows because laziness (seriously, allowing MPI's smpd through the firewall won't be enough, I have to turn the firewalls completely off or it would still timeout...).

Windows, MPI? Where's the flying pigs?

 

[Razor01]

26 minutes ago, leadeater said:

MSSQL is NUMA aware, so is Oracle. MySQL is rather bad at NUMA awareness, hence those sysmark tests being rather bad.

 

You certainly do not under normal situations do CPU pinning or try and override what the OS does with your threads, the OS is already NUMA aware it will put your program thread in the correct place for you. It's very rare you need to override this and if you do you have to be very careful else you'll do a worse job at it compared to the OS.

I don't think its that rare. But again, not my cup of tea, only can talk about what I've been exposed to.  Gaming servers this is a top priority, and its damn complex to work with.  Many cloud servers also recommend this.

[dalekphalm]

56 minutes ago, Razor01 said:

About yields, companies will not go into mass production unless yields are over 75% and its probably even higher than this.

Source?

[Razor01]

9 minutes ago, dalekphalm said:

Source?

 

Not going to find it on the inet, at least not in articles.  But you can ask engineers they will tell ya this.

 

I found this out from a guy that works at TSMC.

 

Here is something

 

https://electronics.stackexchange.com/questions/79874/amd-intel-cpu-yield-failure-rate

 

 

Quote

The yield numbers are trade secrets. However, it is widely known fact that it becomes more and more difficult to get high yields which more advanced technologies. The number of 60% yield seems too low to me - this might be an initial yield when the technology is not mature enough, but this is too low for large scale manufacturing.

 

[Notional]

1 hour ago, Razor01 said:

About yields, companies will not go into mass production unless yields are over 75% and its probably even higher than this.

There is a 0% chance that Volta 100 GPU had a yield over 75%. I doubt Titan V has anywhere near 75% yield rate either. The halo (non gaming/consumer) products from NVidia and Intel do not have 75% yield or close to it.

[Razor01]

2 minutes ago, Notional said:

There is a 0% chance that Volta 100 GPU had a yield over 75%. I doubt Titan V has anywhere near 75% yield rate either. The halo (non gaming/consumer) products from NVidia and Intel do not have 75% yield or close to it.

 

It has to be that high since there is no real cut down parts yet for Volta, We are talking about one memory bank difference between V100 parts so far.

 

Every nV chip, only has one cut down part, so @ 50% yield, they will not make any money on the entire wafer.

[Notional]

Just now, Razor01 said:

 

It has to be that high since there is no real cut down parts, We are talking about one memory bank here.

 

Every nV chip, only has one cut down part, so @ 50% yield, they will not make any money on the entire wafer.

ALL V100 chips are cut down. You can't get a full V100 based GPU.

[Razor01]

6 minutes ago, Notional said:

ALL V100 chips are cut down. You can't get a full V100 based GPU.

 

 

That is what we think, but that doesn't mean there isn't full dies being made. 

 

12nm node is just a 16nm modified node, its a half node, half nodes tend to have better yields without the need for maturity.

 

Those cut dies can just be because of thermal limitations on where those products go to.  They also have to look at thermal limits vs performance what is the best mix.

 

 

[dalekphalm]

27 minutes ago, Razor01 said:

 

Not going to find it on the inet, at least not in articles.  But you can ask engineers they will tell ya this.

 

I found this out from a guy that works at TSMC.

 

Here is something

 

https://electronics.stackexchange.com/questions/79874/amd-intel-cpu-yield-failure-rate

 

 

 

That link does not support your argument that there are at minimum, 75% yields. The article goes on to say some people (who claim to be in the industry) see yields as low as 30%. Obviously a 30% yield isn't normal, and we shouldn't consider it as, but to blindly say that 75% is the magic number seems misleading at best.

 

Yes, they are trade secrets - Intel isn't going to voluntarily tell the world their yield. But the "real" answer seems to be "it varies, a lot".

[Razor01]

11 minutes ago, dalekphalm said:

That link does not support your argument that there are at minimum, 75% yields. The article goes on to say some people (who claim to be in the industry) see yields as low as 30%. Obviously a 30% yield isn't normal, and we shouldn't consider it as, but to blindly say that 75% is the magic number seems misleading at best.

 

Yes, they are trade secrets - Intel isn't going to voluntarily tell the world their yield. But the "real" answer seems to be "it varies, a lot".

I'm not blindly saying it, its from an engineer that works TSMC lol.  I will take his word for it over what others on the inet say.

 

 

And what article are you talking about, cause I didn't link an article it was a forum for EE's and guys talking about yields. 

 

What these guys didn't get into as node side decreases steps increase, each stepping has its own yield analysis, to hit 75% with 50 steps, which is about where they are at right now I think, not sure as no one really talks about this.  they need 99% yields per stepping.  This is how they know well before had when nodes will have enough yields to go into mass production for a particular product.

[Mihle]

The yield of the biggest GPUs (like v100) or like the 28 core Xeon is definitely bellow 75%.

[Dylanc1500]

5 minutes ago, Mihle said:

The yield of the biggest GPUs (like v100) or like the 28 core Xeon is definitely bellow 75%.

I can only speak on behalf of the 28 core Xeon, however do remember that while the yield on the top SKUs produced from the wafer may not be high, the yield of working SKUs is much higher. Intel makes numerous SKUs from the same die, by cutting them down, for a reason. I'd explain further but I'm at work and on mobile.

[Mihle]

35 minutes ago, Dylanc1500 said:

I can only speak on behalf of the 28 core Xeon, however do remember that while the yield on the top SKUs produced from the wafer may not be high, the yield of working SKUs is much higher. Intel makes numerous SKUs from the same die, by cutting them down, for a reason. I'd explain further but I'm at work and on mobile.

When I wrote that I was thinking of full usable as the top end product on the market. I know the damaged xeons is just used as xeons with the broken stuff and some more turned off. But I didn't think of that when I wrote it.

 

I think I red somewhere that Ryzen have all 8 cores fully working on 85% of chips. I don't know if all threads or cache is included in that or not.

Edited March 23, 2018 by Mihle
Typo

[Dylanc1500]

5 minutes ago, Mihle said:

When I wrote that I was thinking of full usable as the top end product on the market. I know the damaged xeons is just used as xeons with the broken stuff and some more turned off. But I didn't think big that when I wrote it.

 

I think I red somewhere that Ryzen have all 8 cores fully working on 85% of chips. I don't know if all threads or cache is included in that or not.

I figured that! With multiple people saying it though, I wanted to clarify it for others that might misunderstand and think otherwise.

[JR88]

I don't Listen to anything that Company Says anymore Lol...Probably will go Bankrupt by 2020 Lol Ryzen appears to be a lot better. but a lot of those chips were kinda shitty and a lot that didn't hit the clocks many wanted....

[dalekphalm]

Just now, JR88 said:

I don't Listen to anything that Company Says anymore Lol...Probably will go Bankrupt by 2020 Lol

Which company? AMD? The one that's on track for a financial recovery after years of planning?

 

Yeah... doubt it.

 

If you mean some other company, please be specific.

[JR88]

1 minute ago, dalekphalm said:

Which company? AMD? The one that's on track for a financial recovery after years of planning?

 

Yeah... doubt it.

 

If you mean some other company, please be specific.

It's an Evil Company with evil Practices and turns around and Blames Intel and Nvidia any chance they get lol

[dalekphalm]

Just now, JR88 said:

It's an Evil Company with evil Practices and turns around and Blames Intel and Nvidia any chance they get lol

Evil?

 

Hookay.

 

You're seriously going to say that, and keep a straight face, and use Intel and NVIDIA in the same sentence?

 

Frankly, neither Intel, nor NVIDIA, nor AMD are evil. They're companies, who's responsibility is to the shareholders, and is to make profit. They aren't altruistic, nor are they straight up evil.

 

All three have done anti-competitive practices in the past. So if AMD is evil, so is Intel and NVIDIA (which could well be correct), but it's definitely not an "either-or" type situation.

[WMGroomAK]

4 minutes ago, dalekphalm said:

Which company? AMD? The one that's on track for a financial recovery after years of planning?

 

Yeah... doubt it.

 

If you mean some other company, please be specific.

I think they might mean Apple ...  /s

 

(Sounds like a troll looking to start a flame war honestly.)