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

[DisEnchantment]

Speculate at will

 

[Joe NYC]

Indeed,

A single mobile LPDDR6 CAMM2 memory module will max out at ~500 MB/s. That's double the Strix HALO bandwidth or halve the RTX 4090 bandwidth.

Expect beefy iGPU's in the not so far away future.

View attachment 112566

Trinary signaling can transfer 8x32 bit in 7 cycles over a 24 bit bus. At 14.4 GT/s that amounts to 65.8 GB/s. Eight 24 bit busses add up to ~526 GB/s.

It seems to me that Strix Halo, and its follow up using LPDDR6 are going to be the future of desktops (!!!) as well as laptops.

People thinking about longevity of AM5 or talking about AM6 with DDR6 may be completely missing the point and not seeing the obvious - LPDDRx takeover of client market.

AMD has the latency reduction technology (with V-Cache, MALL), and LPDDRx will deliver the bandwidth.

 

[Hans de Vries]

No. LPDDR6 is NRZ, and transfers 288 bits over 24 cycles per 12-bit subchannel. Of that, 256 is payload, 16 is host-defined metadata and 16 is either link protection RAS or DBI for power reduction.

At 14.4GT/s a single 192bit LPDDR6 module has a total throughput of 14.4*192/8*(256/288) = 307.2GB/s.

The 24 bit channels, with two 12 bit sub channels do allow both PAM3 and NRZ versions in an efficient manner.

For NRZ only you would simply use bus sizes with powers of 2.

I don't know if they have given up on PAM3 versions already?



PAM3 for LPDDR6 presumably
=========================

2 cycles over one line: 3^2 = 9
9 > 8 (3 bits)
2 cycles over 12 lines: 36 bits
16 cycles over 12 lines: 288 bits

So the PAM3 version would have a 16 cycle burst versus the 24 cycle burst of NRZ.

This gives a maximum bandwith of 460.8 GB/s

The data transfer efficiency is: 256/288 x log(8)/log(9) = 84.12%

The trinary to binary conversion efficiency is: 3/2 x log(2)/log(3) = 94.64%



PAM3 theoretically (very high efficiency version)
======================================

7 cycles over a single line: 3^7 = 2187
2187 > 2048 (11 bits)
7 cycles over 3 lines = 33 bits = 32 + 1 bit

This gives a maximum bandwidth of 526 GB/s

The data tranfer efficiency is: 32/( 3 log(3^7)/log(2)) = 96.14%

The trinary to binary conversion efficiency is: 11/7 x log(2)/log(3) = 99.15%

 

[Doug S]

The 24 bit channels, with two 12 bit sub channels do allow both PAM3 and NRZ versions in an efficient manner.

For NRZ only you would simply use bus sizes with powers of 2.

I don't know if they have given up on PAM3 versions already?



PAM3 for LPDDR6 presumably
=========================

2 cycles over one line: 3^2 = 9
9 > 8 (3 bits)
2 cycles over 12 lines: 36 bits
16 cycles over 12 lines: 288 bits

So the PAM3 version would have a 16 cycle burst versus the 24 cycle burst of NRZ.

This gives a maximum bandwith of 460.8 GB/s

The data transfer efficiency is: 256/288 x log(8)/log(9) = 84.12%

The trinary to binary conversion efficiency is: 3/2 x log(2)/log(3) = 94.64%



PAM3 theoretically (very high efficiency version)
======================================

7 cycles over a single line: 3^7 = 2187
2187 > 2048 (11 bits)
7 cycles over 3 lines = 33 bits = 32 + 1 bit

This gives a maximum bandwidth of 526 GB/s

The data tranfer efficiency is: 32/( 3 log(3^7)/log(2)) = 96.14%

The trinary to binary conversion efficiency is: 11/7 x log(2)/log(3) = 99.15%

PAM3 is too power hungry for "LP"DDR.

 

[Tuna-Fish]

The 24 bit channels, with two 12 bit sub channels do allow both PAM3 and NRZ versions in an efficient manner.

For NRZ only you would simply use bus sizes with powers of 2.

I don't know if they have given up on PAM3 versions already?

There maybe will be one later, presumably called LPDDR6X. But LPDDR6 is set in stone, does not use it, and as far as I know, was never planned to use it.

 

[Josh128]

Good evening folks.

My 9800X3D Dawntrail result. 1.4v+, 5.86Ghz, AIO cooling + liquid Metal.

Bruh. World record?? Be very careful with that 1.4V mate, Anyway, whats the secret to reducing CL? My Teamgroup 6000 30-36-36-76 wont even post if I try to go CL28 without touching anything else from EXPO settings. Does increased RAM voltage help with lower CL? I know it helps with upping freq. Im not looking to break any records, just trying to pick low hanging fruit.

 

[igor_kavinski]

My Teamgroup 6000 30-36-36-76 wont even post if I try to go CL28 without touching anything else from EXPO settings.

Don't use EXPO. Switch to manual timings and only enter 6000 for speed and 28 for CL and let everything else remain on AUTO. Then the mobo should try to train the memory at that latency. If it fails, then you need to reduce speed to 5800 or even 5600. For a kit that can do CL28 at 6000, it should be rated for 6600 or 6800.

 

[gaav87]

I did it guys im top1 on userbenchmark in memory score
Beat every 14900k and 14900ks

 

[Saylick]

I did it guys im top1 on userbenchmark in memory score
Beat every 14900k and 14900ks View attachment 112710View attachment 112709

Inb4 Userbenchmark owner changes how the score is calculated to put Intel back on top.

 

[gaav87]

Inb4 Userbenchmark owner changes how the score is calculated to put Intel back on top.

Ye he will somehow change my 100% memory score so im below that 8800mt/s 14900ks who i beat xD

 

[Josh128]

I did it guys im top1 on userbenchmark in memory score
Beat every 14900k and 14900ks View attachment 112710View attachment 112709

With what CPU/setup?

 

[gaav87]

With what CPU/setup?

9800x3d duuh.

 

[Josh128]

9800x3d duuh.

You got 34ns memory latency?? Lowest Ive seen for any Zen 5 is in the 50s in AIDA64. Must be reading something differently.

 

[gaav87]

You got 34ns memory latency?? Lowest Ive seen for any Zen 5 is in the 50s in AIDA64. Must be reading something differently.

Lol its userbenchmark dude. His numbers mean nothing xD
But i beat his meme benchmark.

 

[igor_kavinski]

You got 34ns memory latency??

UB code loop got stuck in V-cache and kept bouncing around

 

[maddie]

But i beat his meme benchmark.

Set up a pool for how long.

 

[SolidQ]

9800X3D Starcraft 2/Tarkov/RPCS3

Spoiler

 

[igor_kavinski]

No wonder it's OOS everywhere. It's a legendary CPU, the 9800X3D.

 

[Win2012R2]

No wonder it's OOS everywhere. It's a legendary CPU, the 9800X3D.

And this is before it gets paired with 5090...

 

[Josh128]

New AGESA 1.2.0.2.b out that supposedly reduces AIDA64 mem latency figure. WCCFTECH cites a website that tested and supposedly it helps gaming up to 5% or 6% as well. Anybody get a chance to test it out yet? I havent upgraded yet. I hate messing with BIOS when I have a smooth running system.

 

[igor_kavinski]

New AGESA 1.2.0.2.b out that supposedly reduces AIDA64 mem latency figure.

I've read the opposite. It increases the latency because the way AIDA64 calculates the latency doesn't take into account the change they made. Performance in games may be better. If you don't like the change, the BIOS will have a legacy option which resets the behavior back to launch one, according to AMD but they can only provide guidance to their partners. It's up to the mobo maker to decide what to name that option.

https://www.reddit.com/r/Amd/comments/1h8siwi/asus_intros_core_tuning_config_for_gaming_feature

 

[Josh128]

I've read the opposite. It increases the latency because the way AIDA64 calculates the latency doesn't take into account the change they made. Performance in games may be better. If you don't like the change, the BIOS will have a legacy option which resets the behavior back to launch one, according to AMD but they can only provide guidance to their partners. It's up to the mobo maker to decide what to name that option.

https://www.reddit.com/r/Amd/comments/1h8siwi/asus_intros_core_tuning_config_for_gaming_feature

According to WCCFTECH, its the previous 1.2.0.2a patch that increased the way AIDA64 reads it. In any case, their cited source claims BOTH faster gaming AND reduced reported latency in AIDA64. 🤷‍♂️

Gaming Benchmarks Reveal AMD AGESA 1.2.0.2b Patch Brings Slightly Better Gaming Performance On Ryzen CPUs

The Beta BIOS with the AMD AGESA 1.2.0.2b patch seems to be working great for Ryzen CPUs in gaming applications.

wccftech.com

As mentioned in the results, this was achieved via BIOS 3065 Beta, which isn't yet official. Currently, you can only download the BIOS 3057 for this particular model, which comes with the AGESA 1.2.0.2a patch. However, the latest one should come out in a few days. One more thing to point out here is that the 1.2.0.2b patch shows improved memory latency as tested by the user in AIDA64 Extreme. As mentioned in the intro, the memory latency decreased to 64 ns from over 68 ns. This is similar to pre-AGESA 1.2.0.2a patch latency.

 

[Gideon]

Strix Halo GPU performance leak:

https://videocardz.com/newz/amd-ryzen-ai-max-pro-395-strix-halo-apu-with-raden-8060s-graphics-leaks-out

Though it sure seems AMD still hasn't fired the right marketing people: "AMD Ryzen AI MAX+ PRO 395 (w/ Radeon 8060S)" Seriously??????

Why couldn't it just be "AMD R395 (PRO)" or just "Ryzen 395 (PRO)" (the R would stand for Ryzen anyway). Why do you need multiple useless acronyms and buzzwords?

• The 395 number ALREADY tells you it's a "MAXXX+++++," SKU
• "AI" in product names is BS and people HATE that.
• The people that care already know Ryzen is AMD anyway (or they don't and don't care)

Apple just puts their logo on their badge, next too the SKU name (that's very concise) and that's it. AMD should also just keep the Arrow logo, "R" (for Ryzen) and a single 3-4 digit product number. Ryzen + the product name is the only thing people will remember or verbally express anyway.

Imagine a PC shop talking about these over a phone. What do you think would say? 99% it would be either a "I have a shipment of Ryzen 395 incoming" or even "Strix Halo". Nobody is gonna pronounce the "AI MAX+++++ PRO" BS!

 

[FlameTail]

Only 5.1 GHz boost clock speed?

 

[SteinFG]

Only 5.1 GHz boost clock speed?

Kind of sane. Though I wonder why they didn't go higher. Probably to reduce power consumption.

 

[Thibsie]

Strix Halo GPU performance leak:

https://videocardz.com/newz/amd-ryzen-ai-max-pro-395-strix-halo-apu-with-raden-8060s-graphics-leaks-out

Thanks. How good/bad is this ?