Steam Deck - SteamOS, Zen2 4C/8T, RDNA2 1.0-1.6 TF, 16 GB LPDDR5 88 GB/s, starting at $399 [2021-12]

Yea i perfer the "xbox" way of analog sticks myself but i can see why it wouldn't work with this design.

I'm also thinking that since the controls are at the top you might be able to design a rubber bumper case that kinds rounds on the bottom to allow for better weight distribution if your using it on your lap or on a plane tray or something

 

A small mobile pc in a switch case, could be intresting.

 

Some Linux kernel dumps actually pointed out to Van Gogh using 256bit, though those were eventually discarded as reading errors. The fact that it only has 8 active CUs (at least in the case of Steam Deck) also points to 256bit being too unbalanced.

If wee look at what LPDDR5's offers, to get a total of 16GB at 64bit we'd need availability of either a single 128Gb 64bit-wide chip or 2x 64Gb 32bit-wide chips at 5500MT/s. I can't find either on Samsung, Micron or SK Hynix catalogs.
Nor does 64bit / 44GB/s make much sense performance-wise, unless there's a bunch of Infinity Cache in there (most probably there isn't, considering it's a ~168mm^2 chip).

I know some people argued that Van Gogh must be 64bit because that's what "quad-channel" means for Qualcomm's Snapdragon, but those are 4-5W chips (which AMD probably can't make because of Samsung's contract) and AMD never used 16bit PHYs. There's also some theory that "LPDDR5 only comes in 16bit channels" which is factually false, considering Intel uses dual 64bit LPDDR5 controllers on the 7-15W Tiger Lake refresh.



Regardless, if Steam Deck does come with 44GB/s bandwidth against all odds, I'll cancel my pre-order the next instant since it would mean having an iGPU completely castrated by memory bandwidth and performance would be terrible.

 

whoa! i just watched the ign video. this thing is huuge!

 

I think il keep it to laptops for 'mobile' gaming

 

Nah we'd run two QDPs in subchannel mode.
Kinda hacky but it works!
I.e. 16b subchannel has two 16Gbit devices at 8b each parading as a unified whole.

Means the exact same for AMD!
Note the RNR/CZN slides saying dual channel DDR4/quad channel LPDDR4x.
That's for a 128b part.

Also 8cx is a 7-9W part or so.

Bogus, U42 is yet to ship with LP5 nor will it probably actually ship.
It's also quad 32b which is how LPDDR works.

Dem fightan' words.

 

So... is Van Gogh 64bit or 128bit?

The first sentence hints at 64bit, the other two at 128bit...

 

The former.
Dual-channel LPDDR5 and all (versus quad-channel L4x for RNR/CZN both).

 

Then its performance is going to get obliterated by Tiger Lake U unless there's something like 16MB LLC in the iGPU (which is very unlikely).
Same thing for Cezanne... how is this going to compete with the 15W 5800U and what's the point even?

I guess this isn't the gaming PC handheld I was holding out for, after all.

 

Then why did Vang Gogh "appear" to have a quad-channel setup from the March bootlog claiming a 256bit DDR5 bus?

 

Nope, U42 is an abortion at sub-25W.

Nope.

By having the killiest iGP you've seen in eons.
RMB is even funnier like that (due to wider in every way config).

Someone fucked up or it misread L5 subchannels as individual DDR5 channels each.

 

I just don't see how this will behave like a 1.6 TFLOPs RDNA2 GPU given the 44GB/s split between 4*Zen2 cores and the iGPU.
Moreover the 4*Zen2 cores are probably coming with only 4MB L3 so they'll need to hit the LPDDR5 more often than Cezanne. At least Tiger Lake comes with 12MB L3 of which the Xe iGPU is also a client.

And why didn't they push for LPDDR5 6400MT/s, if memory bandwidth was so constrained?

This was looking like the thing I had been longing for in many years, but now I'm super skeptical about it. Unless I'm missing something, those pre-orders are going to fall once benchmarks for the handheld start pouring in.

I guess I should start redirecting my anticipation to handhelds using Rembrandt U and Alder Lake U..

RMB = Right Mouse Button? Raster.. erm.. Memory Buffer?



I also don't get how Van Gogh at 162mm^2 ended up being larger than the 156mm^2 Renoir with 128bit memory and 8 Zen2 cores, considering it has half the memory PHYs and half the die area dedicated to CPU cores.
Even if Van Gogh's RDNA2 8CUs are ~60% larger than Renoir's GCN5 8CUs, it still shouldn't translate into a bigger die.
Here's Renoir's xray edited by me:






Where did all the red die area go?

 

Well then you have to wait some!

Gen12 sucks anyhow.
Why even mention U42, it's not a part that's usable at 7-10W.

PoP-only until waaaaay into late this year.

Hawt!

Rembrandt.

Because it has a sizeable CVML block that won't ever be touched here.

VGH was NOT made for this handheld, it's not a semi-custom part at all and was never supposed to be sold for those prices.

 

It's usable enough at 15W which is how all the gaming benchmarks we know of were done, and at least on the One XPlayer it's configured as such.

I thought the Machine Learning capabilities simply referred to RDNA2's higher rate INT8/INT4 operations. Why would they have that if not even the dGPUs use it?
Moreover, the same mention to CVML also appears in Rembrandt on the roadmaps, and that's just the socket-able APU that will replace Cezanne.

 

Oh not really.
Not in hybrid loads.

Most relevant laptops run it at 25-28W PL1.

Nope.

It also has the same block and more.
It's a pretty chungus part clocking at ~209mm^2 (exactly Raven, ha) worth of N6 goodies.

 

What do you want to do on this exactly ?


as I have shown in this thread and others the aya neo which is a 6/6 zen 2 + a vega with ddr 4 is able to play games very well


The steam deck gpu should be much much better

For me its about having my library in a more portable on the go form factor. There are most likely tens of thousands of games this will play well on the go. a laptop doesn't really work well on bus or train during my commute or even an airplane. even a surface go or pro would be tough to do.

 

The GPD Win3 runs at 20W PL1 and the results are excellent, especially considering it's pretty tiny compared to the Steam Deck:
https://www.notebookcheck.net/GPD-Win-3-Review-The-Unofficial-Microsoft-Switch.514859.0.html

I would expect the One XPlayer to be even faster.

That's my point. I'm not sure the Steam Deck's gaming performance will be much better than AYA NEO's in all occasions, considering it only has less than 2/3rds the memory bandwidth of the Vega 6 in the AYA Neo.

Couple that with Van Gogh's 4-core CCX probably only having 4MB L3 (the 4500U has 8MB), it's going to starve the main memory bus even more.
For the PS4, Sony had estimated around 20GB/s for the 8-core 1.6GHz Jaguars. Van Gogh's 3GHz 4-core Zen2 are likely to consume a similar amount in gaming loads. This leaves as little as 24GB/s for the iGPU, which is... around 10 times less the VRAM bandwidth that a RX 5500 has available.

I can't stress enough how castrating a 64bit bus will be for Van Gogh's iGPU. If this spec is confirmed, I'm expecting the performance to be pretty bad.

 

Can you elaborate on this.

 

Uh-oh they have the same amount of L3 actually.
(AMD L3 is NOT unified at all).

Yea it will be.
You severely overestimate the amount of b/w any RDNA2 part needs.

6WGP in two SAs backed by a 2MiB L2.

 

So 2WGP disabled, leaving 4WGP active. I wonder if they reduced the number of ROPS per RBE. 32 seems to be overkill for 800p.

Also in regards to BW, 44GB/s is definitely sufficient for ~1 teraflops.

That being said I don't think Valve had input into Van Gogh's design the same way MS and Sony dedicated their needs for the consoles. I would have asked for 8 memory controllers.

In your opinion, what can we expect performance wise from Van Gogh? I.e. similar to Vega 11, Xbox One, RTX 1050 etc.