Zen 3 speculation - Page 2

henryzz · 2019-10-10, 10:28

Am I correct in thinking that one AVX512 instruction should use less than double the power of two FMA3 instructions(or AVX512 instructions running on 256 bits) doing the same thing?
I would have thought that this should be true even if one AVX512 instruction is internally done in two halves.
Part of me wonders why they don't make the size variable and just have an internal loop until it is done. This means the same code will run faster on newer CPUs with wider units.

M344587487 · 2020-05-16, 18:38

AMD's current stance on supporting Zen3 on older chipsets is that they aren't. They're not supporting Zen3 AGESA updates on B450/X470 or earlier, which means that any motherboard using older chipsets cannot support Zen3 as all mobo manufacturer's rely on this AGESA binary blob for the core of their bios. There may be some unsupported hackery to get Zen3 working on some older boards, unlikely though.

This is a problem as B450 is the go-to budget AM4 chipset to this day (B550 should be released this month as the new budget choice about a year late, some B450 boards were "refreshed" as a stop-gap last year). The cited reason for dropping older boards (that a 16MB bios chip is not big enough) is simply false, there exist older boards with a 32MB bios chip and X570 boards with a 16MB bios chip. Given that B450 is the current budget option (that I'd bet good money the vast majority of people buy instead of X570), dropping support so soon likely affects the majority of people that were eyeing up Zen3 as a drop-in upgrade. At the very least AMD should have been clear when B550 was delayed that B450 boards would not support Zen3. 6 months of heads up is better than 0 but a lot worse than 18. For all intents and purposes post-Zen2 AM4 chips might as well be on a new socket for the majority of people, not an exciting prospect when AM5/PCIe5/DDR5 is a much more meaningful upgrade and only a few years away (expected late 2022).

mackerel · 2020-05-17, 08:21

If something is supported or not supported is one of the annoying things about AMD messaging, but also in another way understandable. The chart that sparked all this off also shows X370 as not supporting Zen 2. I have a 3700X on an X370 mobo, and it works apart from a minor fan speed quirk which others report going away with newer bios, but there is no newer bios for this model.

So it is kinda like unsupported ECC on some AMD platforms. Functionality may be there, but it might work, it might not. If it doesn't, you're on your own. Later bios may even explicitly NOT have support at all for unsupported combinations. It is hard to build a business without support through the chain.

Personally I'm not bothered at all by this. I was ok with the Intel approach already. A chipset may support two generations of CPU, and you move on. Don't expect any support beyond the current offering unless explicitly stated otherwise.

*I write "bios" out of habit, I guess it more specifically refers to the AGESA.

M344587487 · 2020-05-19, 15:44

AMD have reversed the decision to not support Zen3 on 400 series boards, supposedly because of the backlash: https://old.reddit.com/r/Amd/comment..._amd_x470_and/

tl;dr

"Optional bios", whether this means optional for the mobo manufacturer to implement or optional for the user to upgrade to is unknown. Worth noting that Asus opposed Zen3 support on 400 series boards so they might back out if they can, as may smaller manufacturers
One-way update, so it's only for those upgrading from pre-Zen3 to Zen3
Zen3 is the last that the 400 series will support
"Timing and availability of the BIOS updates will vary and may not immediately coincide with the availability of the first Zen 3 based processors" - Another out for manufacturers that are reluctant to support Zen3

Quote:

Personally I'm not bothered at all by this. I was ok with the Intel approach already.

I'm fine with the intel approach if it's clear from the start. Deciding to support 400 but not 300 is a good compromise, the 300 series is very hit and miss (from a time before Ryzen had proven itself and AMD only had Bulldozer and co. to show for the past decade, many manufacturer's didn't exactly put their best foot forward). Hopefully the plan for AM5 will be crystal clear.

mackerel · 2020-05-19, 16:32

With hindsight, I think their mistake was not to clarify compatibility earlier. This scenario we are entering is exactly why they didn't want to support it at all. Will be interesting to see who does what once the dust settles. Where was the Asus info from?

I saw elsewhere the Zen 3 for older platform bios will be labelled "beta" so prevent cases where people might mistakenly think it a routine update, and prevent their system from working. If this will be anything like my experiences with Zen 2 on X370, I'll give it a miss. I want a bit more support than "it seems to work". Still debating a B550 board for Zen 2 as neither board I current have I want to use seriously.

Just wondering, has there been any speculation on chipsets for Zen 3? Will we get 600 series, or will 500 series continue on? Especially given how late 550 has been. Don't want to see a X670 then wait best part of a year for B650.

M344587487 · 2020-05-19, 19:45

The Asus info was from here ( https://youtu.be/T5X-8vZtml8?t=1434 ), less harshly worded than I remember so he may have talked about it differently elsewhere or I'm confusing it with info from reddit. Many reddit threads indicate that Asus poorly supports older chipsets (that was also my experience waiting months for an outdated bios), they mainly focus only on the latest chipset probably down to lack of manpower.

My understanding of the roadmap may be outdated, from my understanding there'll be Zen 4 next year on AM4, no Zen3+, and that Zen 5 will be on AM5 set for around Q4 2022. That doesn't leave much time for a B650 board to take hold, but I wouldn't be surprised if an X670 was released as a minor refresh of X570, maybe introducing Thunderbolt support but that's pure conjecture (support for the next USB standard which has Thunderbolt support baked in makes more sense). It seems likely that the roadmap may be stretched from when they made it for a number of reasons, perhaps adding many months to the longevity of each generation. AM5 was expected Q4 2022, but if that's pushed to H1 2023 there may be room for a full 600 series AM4 next year. Even if AM5 is released in 2022 it'll probably be a slow mover thanks to DDR5.

M344587487 · 2020-10-08, 17:07

https://www.youtube.com/watch?v=iuiO6rqYV4o

https://www.youtube.com/watch?v=n-AanO3Axzk

tl;dr from the AMD presentation:

8 core CCX
Unified L3 cache, each core on a chiplet has full and equal access to the L3 cache
FPU and integer "wider and more flexible"
Increased loads/stores
More branch prediction bandwidth
~20% more performance per watt
5900X single core cinebench R20 score of 631, compared to 544 score of i9-10900K
Beating intel in (albeit cherry-picked) gaming benchmarks for the first time in a long time
$50 higher RRP than the previous gen started at, now that AMD has overcome its biggest performance deficit in the consumer space (single-threaded performance) they are free to start milking their cash cow like everybody else
Available November 5th
Tempering Big Navi expectations

mackerel · 2020-10-08, 19:40

Quote:

Originally Posted by M344587487

FPU and integer "wider and more flexible"

To check if people heard it the same way I did. I took that to mean they could dispatch more operations, not necessarily that there was more execution resource. I didn't hear anything to make me think the peak execution rate would be significantly better (for FP64 workloads).

Regardless, I'll look forward to someone else getting it and testing it as I'm not paying for any of the announced models. If they don't take too long to release a 5700X then I might get one. It needs to be purchasable before I get a 3080.

M344587487 · 2020-10-28, 10:35

It's always best to err on the side of pessimism, you're probably right that they mean the existing units are more easily utilised instead of bigger. I'm also not getting the first wave of Zen 3 processors, maybe a non-X down the line is on the cards but more likely my pennies are being saved for big navi iff OpenCL support is there and it looks like big navi might be viable in some shape or form. We'll know more on that front in 6 hours or so.

They've added a few instructions ( https://www.phoronix.com/scan.php?pa...nutils-Support ) including the 256 bit form of this:

Quote:

VPCLMULQDQ - This instruction is to perform a carry-less multiplication of two quadwords. VPCLMULQDQ allows VEX-encoded 256-bit version but part of the path to the AVX-512 extensions supported by Intel.

It's a precursor to AVX512 support supposedly coming in Zen 4, it seems odd that they're adding the 256 bit instruction standalone but there we are. Maybe it won't be exposed for use with Zen 3, but if that's the case why show their hand at all?

M344587487 · 2020-10-31, 09:57

Some early benchmarks, these are a bit of a mess but still interesting to decipher: https://www.sisoftware.co.uk/2020/10...d-performance/

Xyzzy · 2020-11-03, 15:13

What would be the best way of securing a 5600X on release day?

Is it a sure thing that the 5600X has six cores on one chiplet instead of three cores each on two chiplets?

2020-10-08, 17:07	#18
M344587487 "Composite as Heck" Oct 2017 2×5²×19 Posts	Zen 3 AMD presentation https://www.youtube.com/watch?v=iuiO6rqYV4o https://www.youtube.com/watch?v=n-AanO3Axzk tl;dr from the AMD presentation: 8 core CCX Unified L3 cache, each core on a chiplet has full and equal access to the L3 cache FPU and integer "wider and more flexible" Increased loads/stores More branch prediction bandwidth ~20% more performance per watt 5900X single core cinebench R20 score of 631, compared to 544 score of i9-10900K Beating intel in (albeit cherry-picked) gaming benchmarks for the first time in a long time $50 higher RRP than the previous gen started at, now that AMD has overcome its biggest performance deficit in the consumer space (single-threaded performance) they are free to start milking their cash cow like everybody else Available November 5th Tempering Big Navi expectations

2019-10-10, 10:28	#12
henryzz Just call me Henry "David" Sep 2007 Liverpool (GMT/BST) 17FD₁₆ Posts	Am I correct in thinking that one AVX512 instruction should use less than double the power of two FMA3 instructions(or AVX512 instructions running on 256 bits) doing the same thing? I would have thought that this should be true even if one AVX512 instruction is internally done in two halves. Part of me wonders why they don't make the size variable and just have an internal loop until it is done. This means the same code will run faster on newer CPUs with wider units.

2020-05-16, 18:38	#13
M344587487 "Composite as Heck" Oct 2017 2·5²·19 Posts	AMD's current stance on supporting Zen3 on older chipsets is that they aren't. They're not supporting Zen3 AGESA updates on B450/X470 or earlier, which means that any motherboard using older chipsets cannot support Zen3 as all mobo manufacturer's rely on this AGESA binary blob for the core of their bios. There may be some unsupported hackery to get Zen3 working on some older boards, unlikely though. This is a problem as B450 is the go-to budget AM4 chipset to this day (B550 should be released this month as the new budget choice about a year late, some B450 boards were "refreshed" as a stop-gap last year). The cited reason for dropping older boards (that a 16MB bios chip is not big enough) is simply false, there exist older boards with a 32MB bios chip and X570 boards with a 16MB bios chip. Given that B450 is the current budget option (that I'd bet good money the vast majority of people buy instead of X570), dropping support so soon likely affects the majority of people that were eyeing up Zen3 as a drop-in upgrade. At the very least AMD should have been clear when B550 was delayed that B450 boards would not support Zen3. 6 months of heads up is better than 0 but a lot worse than 18. For all intents and purposes post-Zen2 AM4 chips might as well be on a new socket for the majority of people, not an exciting prospect when AM5/PCIe5/DDR5 is a much more meaningful upgrade and only a few years away (expected late 2022).

2020-05-17, 08:21	#14
mackerel Feb 2016 UK 2⁶·7 Posts	If something is supported or not supported is one of the annoying things about AMD messaging, but also in another way understandable. The chart that sparked all this off also shows X370 as not supporting Zen 2. I have a 3700X on an X370 mobo, and it works apart from a minor fan speed quirk which others report going away with newer bios, but there is no newer bios for this model. So it is kinda like unsupported ECC on some AMD platforms. Functionality may be there, but it might work, it might not. If it doesn't, you're on your own. Later bios may even explicitly NOT have support at all for unsupported combinations. It is hard to build a business without support through the chain. Personally I'm not bothered at all by this. I was ok with the Intel approach already. A chipset may support two generations of CPU, and you move on. Don't expect any support beyond the current offering unless explicitly stated otherwise. *I write "bios" out of habit, I guess it more specifically refers to the AGESA.

2020-05-19, 16:32	#16
mackerel Feb 2016 UK 448₁₀ Posts	With hindsight, I think their mistake was not to clarify compatibility earlier. This scenario we are entering is exactly why they didn't want to support it at all. Will be interesting to see who does what once the dust settles. Where was the Asus info from? I saw elsewhere the Zen 3 for older platform bios will be labelled "beta" so prevent cases where people might mistakenly think it a routine update, and prevent their system from working. If this will be anything like my experiences with Zen 2 on X370, I'll give it a miss. I want a bit more support than "it seems to work". Still debating a B550 board for Zen 2 as neither board I current have I want to use seriously. Just wondering, has there been any speculation on chipsets for Zen 3? Will we get 600 series, or will 500 series continue on? Especially given how late 550 has been. Don't want to see a X670 then wait best part of a year for B650.

2020-05-19, 19:45	#17
M344587487 "Composite as Heck" Oct 2017 1110110110₂ Posts	The Asus info was from here ( https://youtu.be/T5X-8vZtml8?t=1434 ), less harshly worded than I remember so he may have talked about it differently elsewhere or I'm confusing it with info from reddit. Many reddit threads indicate that Asus poorly supports older chipsets (that was also my experience waiting months for an outdated bios), they mainly focus only on the latest chipset probably down to lack of manpower. My understanding of the roadmap may be outdated, from my understanding there'll be Zen 4 next year on AM4, no Zen3+, and that Zen 5 will be on AM5 set for around Q4 2022. That doesn't leave much time for a B650 board to take hold, but I wouldn't be surprised if an X670 was released as a minor refresh of X570, maybe introducing Thunderbolt support but that's pure conjecture (support for the next USB standard which has Thunderbolt support baked in makes more sense). It seems likely that the roadmap may be stretched from when they made it for a number of reasons, perhaps adding many months to the longevity of each generation. AM5 was expected Q4 2022, but if that's pushed to H1 2023 there may be room for a full 600 series AM4 next year. Even if AM5 is released in 2022 it'll probably be a slow mover thanks to DDR5.

2020-10-31, 09:57	#21
M344587487 "Composite as Heck" Oct 2017 2·5²·19 Posts	Some early benchmarks, these are a bit of a mess but still interesting to decipher: https://www.sisoftware.co.uk/2020/10...d-performance/

2020-11-03, 15:13	#22
Xyzzy Aug 2002 20723₈ Posts	What would be the best way of securing a 5600X on release day? Is it a sure thing that the 5600X has six cores on one chiplet instead of three cores each on two chiplets?