What to expect from AMD on October 8th?

[Jimmy Weirdarms]

I doubt they'll mention 16 core 5950X at this point because it won't be released until a fair bit later (like before).

Shame, that's the one I'm, going to be after!
As I'm assuming it's going to be the biggest/best AM4 chip ever made, it seems like the obvious choice to max out this system before the world moves on to AM5...

 

[kitfit1]

Shame, that's the one I'm, going to be after!
As I'm assuming it's going to be the biggest/best AM4 chip ever made, it seems like the obvious choice to max out this system before the world moves on to AM5...

Same here as well. Will be going from a 3950x which was released just before last Christmas. I fully expect a 5950x to be be around 4 months after the main release. At the moment none of us knows even when the main release will actually be, but i expect the 8th October to just be an announcement of the release date not the actual release.
I don't see the 5950x seeing the light of day until March or April next year. Which to be honest, suits me fine.

 

[pc-guy]

The 8th is the date when AMD are supposedly releasing (or at least announcing) their Zen 3 CPUs. I've seen so many rumours my head is spinning.

Besides with Nvidia's shenanigans with the 3080 launch I'm wondering what AMD will come out with Re the Zen 3 CPUs.

So if you think your able to read the runes I'd love to hear what you think will happen.

i hope none of the over hyped crap and then reality is single digital performance gain or relying on some other stuff like PCIe4 to differentiate the part.

AMD has been good with their CPU promises...so my hopes are high.#

Nvidia is clearly a bunch of lying scumbags...shame their driver and software support is just so much better than Radeon that I would still buy nvidia even if AMD beats 3080 for like less than 10% for the same price or even £50 less. basically i d rather have a 2060 super / 2070 super than a 5700xt even tho 5700XT is clearly better value all round because of the driver is better and CUDA core acceleration.

 

[pc-guy]

5600x 6 cores single CCX
5700x 8 core cross 2 CCXs
5800x 8 core single CCX
5900x 12 core dual CCX

at launch, 6 cores and 8 cores will be siungle CCX and 12 core will be a doubling up of the 6 core with dual CCX.

after a year there maybe super budget 6core dual CCX and 8 core dual CCX parts to target the low end i3 and i5 from intel

an uber 16core dual CCX to put down intels i9 whenever that gets put out.

then refresh like the XT CPUs to put down intel even more...before the next gen in 6months to 1 year.

 

[LePhuronn]

3100 is cheap to make, because it doesn't have two chiplets.

So are you saying the 3900X is expensive to make? The Matisse substrate is designed for 2 chiplets and IO die. It is the same substrate regardless of how many chiplets are put on it. So pray tell how is it significantly more expensive for a machine to drop a 2nd chiplet onto a substrate?

Unless you have an actual source to cite, I'd suggest you're talking tosh.

 

[FoxEye]

Probably more teasing of future products.

I don't expect an actual release of much/anything in October. Late November, December, even 2021 for some consumer stuff.

We know directly from the mouth of AMD that they will be completing their line-up in 2021. They only have to release a couple products in 2020 to meet the promises they've made.

 

[eeek]

So are you saying the 3900X is expensive to make? The Matisse substrate is designed for 2 chiplets and IO die. It is the same substrate regardless of how many chiplets are put on it. So pray tell how is it significantly more expensive for a machine to drop a 2nd chiplet onto a substrate?

Unless you have an actual source to cite, I'd suggest you're talking tosh.

I actually thought it was a single line for all chips with them being rated at the end as to whether they are a 5950x, 5900x downward depending on the testing result. If I'm wrong I'm sure someone will correct me

It's why I suspect we will see chips running over both CCXs because the binning will result in chips in various combinations just work out which combination makes the most money.

 

[LePhuronn]

I actually thought it was a single line for all chips with them being rated at the end as to whether they are a 5950x, 5900x downward depending on the testing result. If I'm wrong I'm sure someone will correct me

You're correct in that chiplet stocks are built up for each Ryzen SKU based on testing, but the actual assembly of a full Matisse package will be done on different lines. So you have X number of machines making 3600X and fed with a load of 6 core chiplets, and then you'd have another bunch of machines making 3900X. I've never seen the actual production line, but the assembly machines would need to be programmed differently if they're attaching 1 chiplet to the substrate or 2. That's literally a couple of button presses though.

So I really don't know what EsaT is on about in saying dual chiplet SKUs are significantly more expensive to make.

 

[EsaT]

So are you saying the 3900X is expensive to make? The Matisse substrate is designed for 2 chiplets and IO die. It is the same substrate regardless of how many chiplets are put on it. So pray tell how is it significantly more expensive for a machine to drop a 2nd chiplet onto a substrate?

Unless you have an actual source to cite, I'd suggest you're talking tosh.

It's you who is saying gravity doesn't exist and hence needs to provide proof that you can walk over the edge of your house's roof without falling to ground.

Two chiplets models are certainly clear amount more expensive than single chiplet models:

• That second chiplet, made on expensive front line node, itself has clear BOM cost and AMD certainly isn't going to dump more of them than necessary into one CPU.
  
  Also having multiple CCDs causes delay into communication between cores on different CCDs from going through IF.
  Just like in case of 3100 vs 3300X
  That makes using more than necessary CCDs bad also for performance, besides AMD's under the line result.​
• And "installing" CCD certainly won't happen by just dropping piece of glue on silicon interposer (not usual substrate) and chiplet on top of that glue.
  
  Time spent in machine positioning silicon dies increases certainly by notable percentage compared to single chiplet model and in mass production time is money.
  When AMD first published chiplet design there were some doubts about capacity to manufacture interposers and especially "mate" dies on them in sufficient numbers.​

 

[EsaT]

I actually thought it was a single line for all chips with them being rated at the end as to whether they are a 5950x, 5900x downward depending on the testing result. If I'm wrong I'm sure someone will correct me

It's why I suspect we will see chips running over both CCXs because the binning will result in chips in various combinations just work out which combination makes the most money.

What makes money best is doing most QC and binning of CCDs before they're built into CPU package.

I mean would be pretty costly waste having tried to make 3950X only to find out that you can have only 4 operating cores in one CCD and 6 in other.
(which could have been used to make two CPUs)
In Zen2 single CCD having two CCXes made binning big deal, because of need for symmetry between (active) CCXes:
If one CCX had two bad cores, they had to disable two of them also from other CCX (that 3100) and even single bad core made it max 6 core CCD.
Hence only even number of cores per CCD CPUs...
And no 10 core refresh to answer 10900K.

Also clock speed capability varies.
And while Ryzen and Threadripper need high clock speed capability, EPYCs couldn't care less about 4+GHz scaling or even sub 4GHz "clock speed wall", if energy efficiency is good around 3GHz.


Now Zen3 changes that first aspect with CCX covering all 8 cores on CCD.
And hence there's no CCX symmetry requirement preventing CPUs with odd number of cores.
But still binning silicon before you try to make CPU out of them is the best for making money.

 

[4K8KW10]

They could try to be a little more generous and give a 10-core part as Ryzen 7 5800X.

30-40% performance improvement is and will not be enough from gen-to-gen perspective.
In absolute term it's only the jump from 30 FPS to 39 FPS.
Which means lag and stutter would still be there.

 

[clv101]

30-40% performance improvement is and will not be enough from gen-to-gen perspective.

What performance improvement did Intel deliver from gen 8 to 9 to 10?

 

[4K8KW10]

What performance improvement did Intel deliver from gen 8 to 9 to 10?

PassMark shows 28% from 8-core Core i9-9900K (18871) to 10-core Core i9-10900K (24175).
https://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i9-9900K+@+3.60GHz&id=3334
https://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i9-10900K+@+3.70GHz&id=3730

 

[copiedright]

Throwing this speculation out there... so its written down before the launch
5600x 8 cores, 8 threads (1 ccx)
5700x 8 cores, 16 threads (1 ccx)
5800x 16 cores, 16 threads (2 ccx)
5900x 16 cores, 32 threads (2 ccx)

As for the 4xxx vs 5xxx. I am still sceptical, but I can see the 4xxx series being used by the zen2 with Vega series.

 

[4K8KW10]

Throwing this speculation out there... so its written down before the launch
5600x 8 cores, 8 threads (1 ccx)
5700x 8 cores, 16 threads (1 ccx)
5800x 16 cores, 16 threads (2 ccx)
5900x 16 cores, 32 threads (2 ccx)

As for the 4xxx vs 5xxx. I am still sceptical, but I can see the 4xxx series being used by the zen2 with Vega series.

Oh, AMD doing Intel
What performance improvement will be seen going from 12-thread Ryzen 5 3600 to 8-thread Ryzen 5 5600?

If this is true, AMD will wait very long time for my money.

 

[copiedright]

Oh, AMD doing Intel
What performance improvement will be seen going from 12-thread Ryzen 5 3600 to 8-thread Ryzen 5 5600?

If this is true, AMD will wait very long time for my money.

Don't worry too much, I am only speculating! Have no more info than anyone else.

But to speculate further to your question. The 8 core, 8 thread, will use single CCX where as the 3600 has two CCX. So a 3600 to a 5600 upgrade would see an IPC increase as well as no infinity fabric losses. Alternatively, they could make it a 5500x and lower the price. Only having two chiplets would make it cheaper to produce.

Anyways, pure speculation. Let's find out on the 8th Oct

 

[Bottle]

Although they have promised compatibility with AM4 socket, what's the likelihood a new chipset and socket will be released as well?

 

[pc-guy]

Although they have promised compatibility with AM4 socket, what's the likelihood a new chipset and socket will be released as well?

There is no news on any 600 series chipset or boards. It’s only a couple of weeks away so if there is anything it would have been leaked or mentioned by now.

however it would be nice to have X670 boards with loads more pcie lanes. One can only hope.

 

[Stu999]

Hopefully there will be some sort of slide show.

 

[eeek]

Although they have promised compatibility with AM4 socket, what's the likelihood a new chipset and socket will be released as well?

When does DDR5 appear on the market - the answer to that tells you probably everything you need to know.

This is the last AM4 release, next year will introduce DDR5 and an AM5 socket to go with it.