Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[Geddagod]

5:30 AM for me. I might get late for office. Dang it. Couldn't they pick a time neutral for the majority of timezones in the world???

Though if it were during office hours...

It's 9 30 PM on a Sunday for me, which tbh is very reasonable... except for that fact that I also have a physics 2 test like 2 days after.....

 

[uzzi38]

https://twitter.com/x/status/1796471089532723676

Standout things to note:

1. Support for DDR5-5600 JEDEC (up from DDR5-5200), up to DDR5-8000 with OC (up from DDR5-7200)
2. New chipsets are indeed 800 series
3. B840 reuses the A620A chipset and does not support CPU OC
4. Both 600 and 800 series mobos will support Raphael, Phoenix and Granite Ridge
5. CCD support Zen 5 (improved IPC) and is produced on 4nm (improved frequency)
6. New chips are called 9950X, 9900X, 9700X, 9600 (notably, only the 9600 is a non-X part)
7. FCLK now goes up to 2400MHz (up form 2000MHz on Raphael).
8. ECC is supported on Raphael, Phoenix Pro (not regular Phoenix) and Granite Ridge
9. Seems to say something about support for DP2.1 UHBR20? Also support for DP2.1 daisy chaining.
10. Also support for USB4 now as well
11. Also seeing something about Wifi7, unclear on the details of it.

 

[FlameTail]

Do FP and SIMD use the same physical units in a CPU core?

 

[igor_kavinski]

@ Videocardz now: https://videocardz.com/newz/aorus-x870-leak-confirms-ryzen-9-9950x-zen5-flagship-cpu

Yay for leaky Gigabyte, btw!

 

[igor_kavinski]

1. Support for DDR5-5600 JEDEC (up from DDR5-5200)

Disappointing. Couldn't they have worked just a bit harder for JEDEC DDR5-6400 1.1V?

 

[lightmanek]

AMD won the German DIY market over by building a fab in Dresden 25 years ago when that was still considered cool. There has been pro-AMD purchasing behavior ever since. But they don't make anything in Germany anymore and that fab is GloFo now.

I would rather see AMD doing well in regions where Amiga computers did well back in 80s and 90s.
Tinkerers, people wanting something a bit different to canon and generally not very fond of Intel as a company.

But 98% of sales is driven by other metrics, like performance, efficiency and TCO (platform longevity).

 

[tsamolotoff]

7. FCLK now goes up to 2400MHz (up form 2000MHz on Raphael).

I've asked a person who knows Chinese what is said there and it seems that this phrase is generic, not really related to some new arch / CPU. I guess it might refer to the fact that FCLK is set to 2400 by default for Zen4 APUs and 2000 for chiplet-based Zen4 CPUs

 

[eek2121]

EDIT: Possibly not. It is early, I will think about it once I wake up.

 

[tsamolotoff]

But 98% of sales is driven by other metrics, like performance, efficiency and TCO (platform longevity).

There is also different views on what a PC is in different countries. I've noticed that people in poorer parts of the world (former CIS, India, South America, Eastern Europe etc) strongly (almost pathologically) prefer the 'winner / leader' brands like Intel or nVidia even if price / performance ratio is heavily disfavouring such SKUs. I personally know several people who bought 3050 instead of 6600 "because it's nVidia" and "its future resale value is higher" (lots of people sell their old HW to buy newer one, some even do that well in advance of new generation releases - one guy just sold his 4090 to buy 5090 at release)

 

[itsmydamnation]

Do FP and SIMD use the same physical units in a CPU core?

zomg you are a noob

yes , and some people are even adventurous enough to do scalar int ops on the FPU ( hand rolled code )

 

[Mahboi]

Uh, DP 2.1? uhbr20? Either AMD was holding out on us or we are getting a new IO die.

IIRC RDNA 3 already had both of those, couldn't they just have put 2 RDNA 3.5 CUs into the chips?

AMD Ryzen 5 7600X Review - Affordable Zen 4 for Gaming

Ryzen 5 7600X is the company's most affordable offering for the Zen 4 family. Our review of the 7600X confirms that this new $300 CPU offers huge performance gains over Zen 3, and can even beat the Ryzen 5800X3D in gaming. In applications, the 7600X is faster than the Intel Core i9-11900K.

www.techpowerup.com

Seems the 7600X used RDNA 2, so my guess is that they replaced it with the cheaper/smaller RDNA 3.5 (or maybe they'll dump extra RDNA 3 stock lel, wouldn't be above them for sure).

 

[leoneazzurro]

If the I/O die is the same, then there is no chance we'll see RDNA3/3.5 on Granite Ridge.

 

[Mahboi]

Well we have faster memory, new DP/uhbr features, and some other things.
It's probably still on TSMC 6 cause why change, but otherwise yeah, I'd say it's a new or rather slightly upgraded I/O die.

 

[FlameTail]

yes , and some people are even adventurous enough to do scalar int ops on the FPU ( hand rolled code )

So FP performance is tied to SIMD performance.

 

[DrMrLordX]

Disappointing. Couldn't they have worked just a bit harder for JEDEC DDR5-6400 1.1V?

JEDEC timings can be pretty bad.

 

[itsmydamnation]

So FP performance is tied to SIMD performance.

depends,

you have scalar FP , SIMD FP and SIMD INT.

now you have different pipelines that support different operations and that can vary massively per uarch.

so scalar FP is generally going to be OOOE/latency/etc limited because its normally mixed in with normal INT/branchy code etc.

SIMD FP and SIMD INT performance is going to be dictated by lots of different things, but the biggest issues tends to be number of Load/Store ops compared to arithmetic. Load and store of vectors is slow/ expensive compared to execution.

So the real question is what do you mean by FP performance.

 

[leoneazzurro]

Well we have faster memory, new DP/uhbr features, and some other things.
It's probably still on TSMC 6 cause why change, but otherwise yeah, I'd say it's a new or rather slightly upgraded I/O die.

That can be due to improvement in packaging and validation of new memory support, in any case. Ofc a new I/O die would open other improvements as well, but so far there are no evidences in support of that.

 

[poke01]

What will be funny as hell is when Zen 5 is priced like a premium Intel KS SKU and it still sells like crazy.

Zen 5 will sell like crazy. I say this as an Apple M sheep. Any hardware nerd will go nuts over it.

 

[blackangus]

You are being sarcastic, but that is 100% true. If it weren’t, The term AI wouldn’t be screamed from the rooftops. Hint: The actual market for AI is smaller than an AAA game. No joke. Some companies are toying with it, but it solves nothing, might (read: does, according to current ongoing legal cases) violate copyright, and gets a ton of stuff wrong (ask Google how many rocks you should eat per day if you disagree, hopefully they patched that one LOL)

Its not 100% true, like everything in the world, its much more grey.
But that conversation is not for here.

 

[poke01]

I might build my first all AMD rig. I already have a 6750 XT. Should I replace my 12400 with a 9600?

Come on AMD, impress me with that keynote.

 

[reggie_fils_aime]

Those B840 boards are going to be very interesting to anyone who might want to whitelabel build on AM5 - an A-series board in a 10 dollar suit, sure, but I'm keen to see how skint they are.

 

[Det0x]

I've asked a person who knows Chinese what is said there and it seems that this phrase is generic, not really related to some new arch / CPU. I guess it might refer to the fact that FCLK is set to 2400 by default for Zen4 APUs and 2000 for chiplet-based Zen4 CPUs

And with OC ~2500-2600mhz FCLK just like the 8000 series APU ?

 

[Det0x]

I also found this very interesting in the leaked gigabyte slide:
Pretty clearly Indicating FCLK will be the bottleneck again i guess

 

[blackangus]

So if FCLK is 2000 then we are looking at DDR5 6000 being the optimal 1:1 speed? As I said I have not paid alot of attention here over the past few years so forgive my noob questions.

 

[tsamolotoff]

So if FCLK is 2000 then we are looking at DDR5 6000 being the optimal 1:1 speed? As I said I have not paid alot of attention here over the past few years so forgive my noob questions.

Optimal FCLK is the highest you can get, unless you have a board and CPU capable of running RAM at 8000 MT/s and if you don't care about bandwidth and only care about latency ()then you go for 2000 to lower latency by ~3-5 ns). And I guess it's not always true in this case. Currently I've been running 7800 / 2166 for several months, it's extra +17% calc speed in my sparse matrix multiplication app as compared to 6400/2166 (which is 2x faster than my previous setup of 5900x + 3733 CL14 ballistix 4x8 set).