Question AMD Rembrandt/Zen 3+ APU Speculation and Discussion

[cherullo]

I guess it's the Samsung chip.

 

[uzzi38]

I'm gonna be a bit lazy and just post my Twitter post instead of downloading both images onto my phone and then uploading them here.

https://twitter.com/i/web/status/1423696744165126147

+50-60% gains over Vega at 15W (AKA would even surpass a fully maxed out stock 5700G iGPU by 10-20%). And as a reminder, Rembrandt boasts 2-3x the compute throughput of Van Gogh here (2x with the latter at 1.5GHz, 3x with the latter at 1GHz).

The future looks bright for Rembrandt if you ask me personally.

 

[JasonLD]

I'm gonna be a bit lazy and just post my Twitter post instead of downloading both images onto my phone and then uploading them here.

https://twitter.com/i/web/status/1423696744165126147

+50-60% gains over Vega at 15W (AKA would even surpass a fully maxed out stock 5700G iGPU by 10-20%). And as a reminder, Rembrandt boasts 2-3x the compute throughput of Van Gogh here (2x with the latter at 1.5GHz, 3x with the latter at 1GHz).

The future looks bright for Rembrandt if you ask me personally.

Doom Eternal is probably the title that will show the biggest performance difference between Vega and RDNA2. Doom Eternal didn't really like Vega while it runs exceptionally well on RDNA2.

 

[mikk]

It's using LPDDR5-5500 Quadchannel right? Over the typical DDR4-3200 Dualchannel it's a major bandwidth uplift over Vega iGPU, this needs to be said.

 

[uzzi38]

It's using LPDDR5-5500 Quadchannel right? Over the typical DDR4-3200 Dualchannel it's a major bandwidth uplift over Vega iGPU, this needs to be said.

Aya NEO here uses quad-channel LPDDR4X-4266.

In both instances bus width is 128b, difference in memory bandwidth is like 30%.

 

[uzzi38]

Doom Eternal is probably the title that will show the biggest performance difference between Vega and RDNA2. Doom Eternal didn't really like Vega while it runs exceptionally well on RDNA2.

It also runs a bit worse on Linux vs Windows in general.

FlightlessMango

flightlessmango.com

 

[blckgrffn]

Nice

In three years, you'll be able to buy a Next Gen Switch mobile console that is less powerful than this one.

I'll buy that too.

That's my prediction. It's worth a lot

 

[mikk]

Aya NEO here uses quad-channel LPDDR4X-4266.

In both instances bus width is 128b, difference in memory bandwidth is like 30%.

Ok but isn't it using a Ryzen 5 4500U only which uses Vega 6 running at 1500 Mhz?

 

[uzzi38]

Ok but isn't it using a Ryzen 5 4500U only which uses Vega 6 running at 1500 Mhz?

Yeah, and? It's running at 15W (or potentially lower, but there's a rather slim chance of that) and spitting out 25W+ Vega 8 iGPU w/LPDDR4X performance.

Or in other words, nearly matching Tiger Lake at 28W+ whilst needing half that power itself. It's a plenty good showing either way. And again, we already know Rembrandt will come with vastly higher compute throughput.

 

[mikk]

Yeah, and? It's running at 15W (or potentially lower, but there's a rather slim chance of that) and spitting out 25W+ Vega 8 iGPU w/LPDDR4X performance.

Or in other words, nearly matching Tiger Lake at 28W+ whilst needing half that power itself. It's a plenty good showing either way. And again, we already know Rembrandt will come with vastly higher compute throughput.

Steam Deck has 40% more tflops +30% more bandwidth than the Vega powered device and gets 50% more performance out of it according to you. It doesn't look that great to me considering it's RDNA-2 based. A comparison with dualchannel based iGPU devices isn't fair against quadchannel, such a comparison is flawed.

 

[scineram]

If AV1 is really missing what a colossal failure. What a joke for a top 2022 mobile processor.

 

[uzzi38]

Steam Deck has 40% more tflops +30% more bandwidth than the Vega powered device and gets 50% more performance out of it according to you. It doesn't look that great to me considering it's RDNA-2 based. A comparison with dualchannel based iGPU devices isn't fair against quadchannel, such a comparison is flawed.

40% more TFLOPs? Not only is it 6CUs vs 8CUs and identical max boost clocks - aka 33% higher theoretical TFLOPs, you're assuming that the 1.5GHz max GPU boost clock on both devices is sustained, which is a huge mistake of an assumption to be making for a 15W iGPU. Even at a low clock of 1500MHz. The 4500U typically sustains 1200-1450MHz at this power.

I don't expect this at all. I expect that VGH's iGPU will require more power per clock vs Vega, but also net significantly more performance as a result. I also expect the actual difference in TFLOPs here to be more akin to a ~20% difference.

Also, at this level of performance memory bandwidth should not be a problem. This should frankly be obvious by the fact that Rembrandt has 50% more CUs that will probably clock up to 2GHz, yet only 16% more memory bandwidth.

Also, I don't get why you're so fixated on dual vs quad channel. Tiger Lake runs it's LPDDR4X in "quad channel" as well. How do I know this? Because LPDDR memory is either 16b or 32b per memory channel, and Tiger Lake has a 128b memory bus. The same as Renoir, Cezanne, Ice Lake, Comet Lake and basically every other mainstream consumer-focused CPU released for the last decade.

 

[mikk]

40% more TFLOPs? Not only is it 6CUs vs 8CUs and identical max boost clocks - aka 33% higher theoretical TFLOPs, you're assuming that the 1.5GHz max GPU boost clock on both devices is sustained, which is a huge mistake of an assumption to be making for a 15W iGPU. Even at a low clock of 1500MHz. The 4500U typically sustains 1200-1450MHz at this power.

Yes 40% more tflops, Steam Deck GPU has 1.6 tflops according to various sources versus 1.15 tflops from Vega 6. And sure I have to assume both are running at max boost. That's a very rough estimate and the fps comparison is also a rough one.

 

[Shivansps]

He said that the NVME is not upgradable, so they really have 3 PCB versions then... I kinda get it, otherwise it would be better to just get the eMMC version with the empty M.2 slot. They are also overpricing that NVME like hell, a 1TB NVME is what? $100? a $150?

The part where we all win, is that the Steam Deck will shift a lot of attention to gaming on Linux.

BTW, seeing the performance there, i take everything i said about the screen resolution back, at 15W there is no room for more resolution.

 

[blckgrffn]

He said that the NVME is not upgradable, so they really have 3 PCB versions then... I kinda get it, otherwise it would be better to just get the eMMC version with the empty M.2 slot. They are also overpricing that NVME like hell, a 1TB NVME is what? $100? a $150?

The part where we all win, is that the Steam Deck will shift a lot of attention to gaming on Linux.

BTW, seeing the performance there, i take everything i said about the screen resolution back, at 15W there is no room for more resolution.

I had read just the opposite, that it is replaceable but is a non-typical but standard NVMe form factor and is not easy to replace.

The cost of drives in this form factor is about what the up charge is as I understand it.

But not impossible.

 

[LightningZ71]

He said that the NVME is not upgradable, so they really have 3 PCB versions then... I kinda get it, otherwise it would be better to just get the eMMC version with the empty M.2 slot. They are also overpricing that NVME like hell, a 1TB NVME is what? $100? a $150?

The part where we all win, is that the Steam Deck will shift a lot of attention to gaming on Linux.

BTW, seeing the performance there, i take everything i said about the screen resolution back, at 15W there is no room for more resolution.

I don't get where he's getting that non-upgradeable bit from. Valve themselves have stated that the M.2 drives are "standard" 2230 devices. It may not be easy to source them, but, according to them, it's possible. I'm very interested in the Steamdeck, but, if the user can't upgrade or REPLACE a failed drive, it's a hard no, as I can get similar performance from roughly similarly priced laptops that are more upgradeable and have physical keyboards.

 

[Shivansps]

I dont get it, 2230 M.2 are not so rare, they are uncommon yeah, but is not like you have to search that hard to find one. In fact, Western Digital has a 1TB model avalible in 2230 factor.
So, the basic version is the best version, it means it has a soldered eMMC + the empty M.2, you can take the 64GB one and add a cheap NVME 2230 on it and you still have the 64GB eMMC.

 

[blckgrffn]

I am not sure the Emmc is soldered. I think it too is an m2 device. It might just be SATA? All the reports I have read is that if you upgrade the storage you take out the existing storage.

The bigger storage is also faster. My guess is we will have those details when the review units head out.

It’s just not PS5 easy. It’s partial device disassembly.

 

[uzzi38]

Yes 40% more tflops, Steam Deck GPU has 1.6 tflops according to various sources versus 1.15 tflops from Vega 6. And sure I have to assume both are running at max boost. That's a very rough estimate and the fps comparison is also a rough one.

Yeah I thought the Steam Deck was 1.5GHz on max iGPU clock yesterday when it's actually 1.6GHz max clock. So that's my bad.

However assuming both iGPUs are clocked at the max boost is still a terrible assumption to be making in my opinion. I don't think that will be possible for the overwhelming majority of games at 15W, Doom Eternal included.

I'll be getting a 5800H laptop soon, so I'll do a bit more in-depth testing on the iGPU there to see what it can do.

 

[Panino Manino]

I'm the only one thinking that Valve limited the performance during these hands-on? Linus unit was with SMT disabled, and showing 1GB allocated VRAM.
I suspect that Valve tweaked some things to perform worse but acceptable, maybe to give a better impression on heat and battery life?

 

[eek2121]

I'm the only one thinking that Valve limited the performance during these hands-on? Linus unit was with SMT disabled, and showing 1GB allocated VRAM.
I suspect that Valve tweaked some things to perform worse but acceptable, maybe to give a better impression on heat and battery life?

It is possible that the units were using Engineering Samples.

 

[LightningZ71]

SMT off can improve performance in some cases while also saving power. It's very situational.

 

[uzzi38]

It is possible that the units were using Engineering Samples.

They are indeed engineering samples. In the same video you can see the name of the chip, it's the good ol AMD standard for engineering samples: AMD ENG Sample 100-000000405.

 

[Shivansps]

The comparison linus made is not really valid because the Steam Deck and the Neo were running diferent OS. And it is running behind Proton on the Steam deck. Unless the Aya Neo was also running SteamOS and i completely missed that?
Doom Eternal is known to run slower on Linux via Proton,

 

[Panino Manino]

SMT off can improve performance in some cases while also saving power. It's very situational.

But today 4T is becoming too little, that's why the AYA competitor uses a 6T(6C) CPU, those extra 2 cores are useful.