Discussion AMD Ryzen MegaThread! FAQ and Resources

[Nope 1151]

https://twitter.com/CPCHardware/status/876193860946468865
In french but reads:
"Ah here: B2 stepping approach to Zen. It fixes many hardware bugs."

More importantly: "B2-step seems to focus on Uncore/SoC erratas (PCIe controllers, etc.) . I don't see many fixed bugs within the core"

So no Linux ASLR fix sadly

Some thought in the forums that the PCI-E controller was causing poorer then expected gaming results.

I personally don't know if that is true.

I'm not sure. Has (have??) any test been done on that yet?


Edit: https://community.amd.com/thread/215773?start=105&tstart=0
Quoting mcl100: "Upgrading to kernel 4.11.6.
Enabling or disabling NUMA in the kernel
Enabling or disabling ASLR
Using the performance cpu scaling governor (normally I use ondemand)
Recompiling the system (gcc 6.3) with -O2 -pipe -march=znver1 -mtune=haswell compiler flags
Recompiling the system (gcc 6.3) with just -O2 -pipe
"Overclocking" the RAM to 2933 MHz
Overclocking the CPU to 3.7 GHz, VCore 1.25V (N.B. can run prime95 overnight without failure and without the CPU ever exceeding 50 degrees C)"

Oh dear god this is gonna be a looong wait.

 

[randomizer]

Edit: https://community.amd.com/thread/215773?start=105&tstart=0
Quoting mcl100: "Upgrading to kernel 4.11.6.
Enabling or disabling NUMA in the kernel
Enabling or disabling ASLR
Using the performance cpu scaling governor (normally I use ondemand)
Recompiling the system (gcc 6.3) with -O2 -pipe -march=znver1 -mtune=haswell compiler flags
Recompiling the system (gcc 6.3) with just -O2 -pipe
"Overclocking" the RAM to 2933 MHz
Overclocking the CPU to 3.7 GHz, VCore 1.25V (N.B. can run prime95 overnight without failure and without the CPU ever exceeding 50 degrees C)"

Oh dear god this is gonna be a looong wait.

What about turning it off and on again?

 

[Nope 1151]

Edit: https://community.amd.com/thread/215773?start=105&tstart=0
Quoting mcl100: "Upgrading to kernel 4.11.6.
Enabling or disabling NUMA in the kernel
Enabling or disabling ASLR
Using the performance cpu scaling governor (normally I use ondemand)
Recompiling the system (gcc 6.3) with -O2 -pipe -march=znver1 -mtune=haswell compiler flags
Recompiling the system (gcc 6.3) with just -O2 -pipe
"Overclocking" the RAM to 2933 MHz
Overclocking the CPU to 3.7 GHz, VCore 1.25V (N.B. can run prime95 overnight without failure and without the CPU ever exceeding 50 degrees C)"

Oh dear god this is gonna be a looong wait.

What about turning it off and on again?

If you read that section of posts, the majority of the segfaults occur in the ASLR and SMT, a third option exists for bumping the SOC voltage.
In other words more questions than answers.

 

[randomizer]

If you read that section of posts, the majority of the segfaults occur in the ASLR and SMT, a third option exists for bumping the SOC voltage.
In other words more questions than answers.

I got the impression that nobody actually knows a reliable solution (or even workaround). That's why I gave another suggestion

 

[Nope 1151]

If you read that section of posts, the majority of the segfaults occur in the ASLR and SMT, a third option exists for bumping the SOC voltage.
In other words more questions than answers.

I got the impression that nobody actually knows a reliable solution (or even workaround). That's why I gave another suggestion

My guy feeling is that the bug originates from several issues. Perhaps it is not a single bug but a variety of which that are causing related issues.

 

[juanrga]

Kernel developer for DragonFly BSD found a bug on RyZen. He has worked a partial workaround for the BSD kernel. His solution doesn't solve the bug but reduces the frequency of suffering the bug

https://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/b48dd28447fc8ef62fbc963accd301557fd9ac20

 

[Nope 1151]

Kernel developer for DragonFly BSD found a bug on RyZen. He has worked a partial workaround for the BSD kernel. His solution doesn't solve the bug but reduces the frequency of suffering the bug

https://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/b48dd28447fc8ef62fbc963accd301557fd9ac20

Nice catch, I wonder if this affects the performance of the cpu when it comes out.

 

[goldstone77]

AMD is reportedly able to use 99.9% of Ryzen CPU dies
Published: 22nd June 2017
Source: Bits and Chips
Author: Mark Campbell
"
AMD's new Ryzen CPU design offers a lot of interesting new features, with Infinity Fabric allowing it to offer more scalability than any other CPU before it. With the same Ryzen dies AMD are able to offer 4, 6 and 8 core offerings in their Ryzen 5 and 7 lineup while also being able to utilise two of these dies with Threadripper and four of these dies in EPYC, allowing AMD to create a massive product stack with only a single core/die design.

With this, it is unsurprising that AMD is getting high yields out of their Ryzen silicon, as even slightly defective silicon will be able to neatly fit into this huge product stack, be it CPUs with defective cores, cache or I/O functionality.

A new report from Bits and Chips states that AMD is able to make use of 99.9% of all Ryzen silicon produced, with 80% having eight fully functional CPU cores. This is an amazing feat for the company, as minimising waste in this way will allow AMD to maximise their profit margins while still offering competitive pricing.

Typically a single silicon design is used in only a handful of products, but AMD's Ryzen design allows their CPUs to be scaled in a way that has never been seen before, offering CPU products to end users to deliver a minimum of 4 cores with Ryzen 5 and up to 32 cores with EPYC. With this AMD has been able to take advantage of the efficiencies of smaller silicon dies in even their largest of server CPUs, which bodes well for AMD's future."

"In the future AMD has the potential to also deliver these same efficiencies on future GPUs, with their Navi GPU architecture promising Scalability. We already know that Vega uses AMD's Infinity Fabric, so it is not out of the question to see multi-die GPU products within just a few years. Does AMD plan to create an EPYC-style GPU?"

https://www.overclock3d.net/news/cpu_mainboard/amd_is_reportedly_able_to_use_99_9_of_ryzen_cpu_dies/1

 

[James Mason]

AMD is reportedly able to use 99.9% of Ryzen CPU dies

Won't this lead to a bunch "how do I enable the locked cores on my r5/r3?" threads.

 

[goldstone77]

AMD is reportedly able to use 99.9% of Ryzen CPU dies

Won't this lead to a bunch "how do I enable the locked cores on my r5/r3?" threads.

The staggered launch makes so much more since now. They are selling the "left overs" that couldn't be salvaged for 8 cores parts. It's a pretty ingenious approach, which leads to almost no waste and a great price to performance ratio. Infinity fabric is pure genius! More computers for everyone! This is how they can offer a 1700 for $299 that will rival the multitasking performance of $1000 Intel parts.

 

[goldstone77]

It makes you wonder how the "X" and non "X" CPU's were binned for the 1600's. Maybe placement of the non usable cores for X variants allowed higher clock speed?

 

[jdwii]

The down side is Ryzen 3 is taking some time to come since Amd is getting such great yields

 

[James Mason]

The down side is Ryzen 3 is taking some time to come since Amd is getting such great yields

Yeah, that makes me concerned that people are gonna be able to turn all their 3s into 5s most likely if they're getting a 99.9% usage rate.
Or AMD is just trying to stockpile enough "r3" grade CPUs to actually have enough to release, because they will be popular.
Though, if AMD gets such high yields, doesn't that meant he price of R7s might fall even further, so they become so affordable it's like "don't bother with the R3s" and this would completely wreck Intel if they had R7s going for like i5/high end i3 prices.

 

[-Fran-]

The down side is Ryzen 3 is taking some time to come since Amd is getting such great yields

Yeah, that makes me concerned that people are gonna be able to turn all their 3s into 5s most likely if they're getting a 99.9% usage rate.
Or AMD is just trying to stockpile enough "r3" grade CPUs to actually have enough to release, because they will be popular.
Though, if AMD gets such high yields, doesn't that meant he price of R7s might fall even further, so they become so affordable it's like "don't bother with the R3s" and this would completely wreck Intel if they had R7s going for like i5/high end i3 prices.

I would love for the latter to happen. Let's hope that is the case, but I have to say I don't think it will.

Cheers!

 

[aldaia]

It may be the case that AMD is having lower production cost than intel. AMD has lower die area per core with the aditional benefit of using only small dies.

RYzen (8-core) 1x195 mm2
Threadriper (16-core) 2x195 = 390 mm2
EPYC 8 to (32-core) 4x195 = 780 mm2
vs
LCC 3x4 (10-core) >300 mm2
HCC 4x5 (18-core) ~500 mm2
XCC 5x6 (28-core) ~700 mm2

Despite it's evident shorcomings, the idea of using the same zepelin die from ryzen to EPYC ended up being a brilliant idea, specially considering the limited resources AMD has compared to intel.

 

[juanrga]

The staggered launch makes so much more since now. They are selling the "left overs" that couldn't be salvaged for 8 cores parts. It's a pretty ingenious approach, which leads to almost no waste and a great price to performance ratio. Infinity fabric is pure genius! More computers for everyone! This is how they can offer a 1700 for $299 that will rival the multitasking performance of $1000 Intel parts.

Infinity fabric is not anything special. In fact some call it a HTT+ indicating it is hypertransport plus enhancements.

Also the $329 R7 1700 is currently offering similar multithreading performance than $389 Intel parts like the i7 7800x.

 

[juanrga]

It may be the case that AMD is having lower production cost than intel. AMD has lower die area per core with the aditional benefit of using only small dies.

RYzen (8-core) 1x195 mm2
Threadriper (16-core) 2x195 = 390 mm2
EPYC 8 to (32-core) 4x195 = 780 mm2
vs
LCC 3x4 (10-core) >300 mm2
HCC 4x5 (18-core) ~500 mm2
XCC 5x6 (28-core) ~700 mm2

Despite it's evident shorcomings, the idea of using the same zepelin die from ryzen to EPYC ended up being a brilliant idea, specially considering the limited resources AMD has compared to intel.

195mm² is not the size of Zeppelin die. The die is ~10% bigger than that.

For ThreadRipper and EPYC you have to count the cost of the MCM package as well.

 

[aldaia]

It may be the case that AMD is having lower production cost than intel. AMD has lower die area per core with the aditional benefit of using only small dies.

RYzen (8-core) 1x195 mm2
Threadriper (16-core) 2x195 = 390 mm2
EPYC 8 to (32-core) 4x195 = 780 mm2
vs
LCC 3x4 (10-core) >300 mm2
HCC 4x5 (18-core) ~500 mm2
XCC 5x6 (28-core) ~700 mm2

Despite it's evident shorcomings, the idea of using the same zepelin die from ryzen to EPYC ended up being a brilliant idea, specially considering the limited resources AMD has compared to intel.

195mm² is not the size of Zeppelin die. The die is ~10% bigger than that.

For ThreadRipper and EPYC you have to count the cost of the MCM package as well.

Any source for the area?
All sources I've seen are around 195
e.g. https://en.wikichip.org/wiki/amd/ryzen_7/1800x

 

[8350rocks]

The staggered launch makes so much more since now. They are selling the "left overs" that couldn't be salvaged for 8 cores parts. It's a pretty ingenious approach, which leads to almost no waste and a great price to performance ratio. Infinity fabric is pure genius! More computers for everyone! This is how they can offer a 1700 for $299 that will rival the multitasking performance of $1000 Intel parts.

Infinity fabric is not anything special. In fact some call it a HTT+ indicating it is hypertransport plus enhancements.

Also the $329 R7 1700 is currently offering similar multithreading performance than $389 Intel parts like the i7 7800x.

Infinity fabric decouples the control fabric and the data fabric, which allows much greater granularity, and more options to optimize compute.

It may seem quite simple, but truthfully, this is a game changer across the board. The feature set of HTX is still present, but expanded, and the capabilities of the HTX existing feature set were also improved.

There was nothing wrong with the HTX design beyond the limitations of the original design itself. The technology is sound, now it is far more complete in terms of doing what the vision of the original HTX was designed to do.

 

[aldaia]

One of the inconvenients of using multiple dies is that packaging has an extra cost. However, judging by pictures, it seems that packaging for both EPYC and Threadripper wont be signifficantly more expensive than packaging for a single big die.

EPYC package

Ryzen (delided)

 

[con635]

AMD is reportedly able to use 99.9% of Ryzen CPU dies

Won't this lead to a bunch "how do I enable the locked cores on my r5/r3?" threads.

The staggered launch makes so much more since now. They are selling the "left overs" that couldn't be salvaged for 8 cores parts. It's a pretty ingenious approach, which leads to almost no waste and a great price to performance ratio. Infinity fabric is pure genius! More computers for everyone! This is how they can offer a 1700 for $299 that will rival the multitasking performance of $1000 Intel parts.

4c8t r5 is just 133£ atm, unbelievable value there for a budget gamer and an upgrade path into the bargain.

 

[juanrga]

It may be the case that AMD is having lower production cost than intel. AMD has lower die area per core with the aditional benefit of using only small dies.

RYzen (8-core) 1x195 mm2
Threadriper (16-core) 2x195 = 390 mm2
EPYC 8 to (32-core) 4x195 = 780 mm2
vs
LCC 3x4 (10-core) >300 mm2
HCC 4x5 (18-core) ~500 mm2
XCC 5x6 (28-core) ~700 mm2

Despite it's evident shorcomings, the idea of using the same zepelin die from ryzen to EPYC ended up being a brilliant idea, specially considering the limited resources AMD has compared to intel.

195mm² is not the size of Zeppelin die. The die is ~10% bigger than that.

For ThreadRipper and EPYC you have to count the cost of the MCM package as well.

Any source for the area?
All sources I've seen are around 195
e.g. https://en.wikichip.org/wiki/amd/ryzen_7/1800x

Unfortunately, I am not allowed to say the exact size. However, if you google a bit you can people mentioning values as 214mm² that are very very close to the real size.

 

[juanrga]

The staggered launch makes so much more since now. They are selling the "left overs" that couldn't be salvaged for 8 cores parts. It's a pretty ingenious approach, which leads to almost no waste and a great price to performance ratio. Infinity fabric is pure genius! More computers for everyone! This is how they can offer a 1700 for $299 that will rival the multitasking performance of $1000 Intel parts.

Infinity fabric is not anything special. In fact some call it a HTT+ indicating it is hypertransport plus enhancements.

Also the $329 R7 1700 is currently offering similar multithreading performance than $389 Intel parts like the i7 7800x.

Infinity fabric decouples the control fabric and the data fabric, which allows much greater granularity, and more options to optimize compute.

It may seem quite simple, but truthfully, this is a game changer across the board. The feature set of HTX is still present, but expanded, and the capabilities of the HTX existing feature set were also improved.

There was nothing wrong with the HTX design beyond the limitations of the original design itself. The technology is sound, now it is far more complete in terms of doing what the vision of the original HTX was designed to do.

That is why I said "hypertransport plus enhancements".

 

[goldstone77]

The staggered launch makes so much more since now. They are selling the "left overs" that couldn't be salvaged for 8 cores parts. It's a pretty ingenious approach, which leads to almost no waste and a great price to performance ratio. Infinity fabric is pure genius! More computers for everyone! This is how they can offer a 1700 for $299 that will rival the multitasking performance of $1000 Intel parts.

Infinity fabric is not anything special. In fact some call it a HTT+ indicating it is hypertransport plus enhancements.

Also the $329 R7 1700 is currently offering similar multithreading performance than $389 Intel parts like the i7 7800x.

Sorry it's actually cheaper than quoted previously. And comes with that pretty Wraith Spire LED Cooler!
AMD Ryzen 7 1700 Processor with Wraith Spire LED Cooler (YD1700BBAEBOX) Price: $296.71
https://www.amazon.com/dp/B06WP5YCX6

ri·val: compete for superiority with; be or seem to be equal or comparable to.
"the efficiency of the Bavarians rivals that of the Viennese"
synonyms: match, compare with, compete with, vie with, equal, measure up to, be in the same league as, be on a par with, touch, challenge; informalhold a candle to
"few countries can rival ours for natural resources"

AMD Ryzen 7 1700X Review
by Paul Alcorn March 27, 2017 at 6:00 AM
http://www.tomshardware.com/reviews/amd-ryzen-7-1700x-review,4987-4.html

Those bench marks prove that the sub $300 1700 does compete with, and in some tests beat the 6900k some $1,000 Intel CPU. So, the word rival is correct in my statement.

"In fact some call it a HTT+ indicating it is hypertransport plus enhancements." Some people say all kinds of slang words and terms. Why would they not just use Infinity Fabric like AMD calls it? Is there a specific reason? Is it an attempt to trivializing or defaming it? And for what purpose would that serve?

Defame: damage the good reputation of (someone); slander or libel.

Why would you say,"Infinity Fabric not anything special." Near 100% CPU yields sounds like a very ingenious approach. What kind of yields do Intel receive from their unique fab process? Do they have a "special process" like infinity fabric the creates efficiency to get some near 100% yields, or do they just throw away the 20% failures?

 

[gamerk316]

Those bench marks prove that the sub $300 1700 does compete with, and in some tests beat the 6900k some $1,000 Intel CPU. So, the word rival is correct in my statement.

Not really. Gaming benchmarks are bottlenecked by the GPU, compressing the performance of the CPU. Using your logic, I can prove Ryzen competes against my 2600k.