• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

The Steam Deck is Awesome at 128bit 88GB/s bandwidth

GuinGuin

Banned
The Switch doesn't require you to install games at all. Games on the Switch are also specifically designed with the space limitations in mind. No game is specifically designed for the Steam Deck.



According to Valve, you're wrong. The internal drive is not user-replaceable. Educate yourself.

I seriously doubt it will be that hard to access. The thing is massive. It will be as easy as swapping the alternator in an American made car.
 

mckmas8808

Banned
They will need to get the marketing right to make it a success and it's always going to be a problem with handhelds that they cost more for less power. I can't even imagine if a handheld PS5 equivalent was even possible how much it would end up costing.

It's quite interesting that after Stadia, which promised gaming on any screen via streaming, that the newest things to come out have been the major home consoles and now a PC handheld. I would have expected that something like Steam Deck would be a streaming device instead of what it is shaping up to be.

I think most reasonable minded people realize that streaming isn't "THE" future in the sense that some want to believe. It'll be something that rides along side physical hardware.
 
EDIT:
- The rumors were wrong. The memory controller is 128bit wide.
- The Steam Deck is Awesome at 88GB/s.
- Sorry for the false concerns.

You shouldn't apologise and your concerns weren't false but were well-founded and real concerns.

Valve shouldn't have allowed incorrect technical information to be posted on their own website for one.

And secondly, deluded gamers who are too afraid to question the information their favourite gaming company is putting in front of them, because it might besmirch the honour of their glorious gaming idol company, shouldn't be allowed to bully more rational and informed minds into not asking those important questions either.

Otherwise this forum becomes a tyrannical echo chamber where any thought that even appears to deviate from the insecure populist opinion of the forum's most ardent fanboys is driven off the forum... we do not want to become Era 2.0.
 
You shouldn't apologise and your concerns weren't false but were well-founded and real concerns.

Valve shouldn't have allowed incorrect technical information to be posted on their own website for one.

And secondly, deluded gamers who are too afraid to question the information their favourite gaming company is putting in front of them, because it might besmirch the honour of their glorious gaming idol company, shouldn't be allowed to bully more rational and informed minds into not asking those important questions either.

Otherwise this forum becomes a tyrannical echo chamber where any thought that even appears to deviate from the insecure populist opinion of the forum's most ardent fanboys is driven off the forum... we do not want to become Era 2.0.

dj khaled GIF
 

poodaddy

Gold Member
This is going to be the true question that Valve is going to need to answer.
Nah, it's pretty easy to figure out. Massive Steam library at your finger tips, handheld PC with enough horse power to handle 720 p more than adequately, I'm sold at just that honestly. If you're not a PC gamer at all though, I suppose it'd be a harder sell, but I still think it'd be a worthy purchase. For me though? This is the first handheld I will have ever bought that started with a library of thousands of games that I've built up over the past decade and change. I'm sold.
 
Last edited:

mckmas8808

Banned
Nah, it's pretty easy to figure out. Massive Steam library at your finger tips, handheld PC with enough horse power to handle 720 p more than adequately, I'm sold at just that honestly. If you're not a PC gamer at all though, I suppose it'd be a harder sell, but I still think it'd be a worthy purchase. For me though? This is the first handheld I will have ever bought that started with a library of thousands of games that I've built up over the past decade and change. I'm sold.

The bolded is the potential problem though.
 

gamer82

Member
hmmm, $50 more than the switch for double the performance?
whats the switch got to do with the steam deck, no ones buying a steam deck to play Nintendo games so there's no comparrsion , yes it will probably plat Multiplatform games better but then so will a ps5/xbox.

i get this I a portable but it is still a big device how many kids are y ou going to see rocking this device . I have one pre ordered but I shall wait to see as I already have a ps5 currently and was in market for a switch depending on what other games releases this year that haven't ben dropped.

after watching the interviews on this device the looks on they're faces where like they were not happy and seem a bit cagey but that's just me. I'd really want to see what sort of running this does on a big screen because they are making a dock separately so they want you to also play it on a big screen .

mind you they did say you can buy any run of the mill dock and it ill works so I don't know why they are going to the effort of a rock being sold separately what is the actual need for them to waste their time and resources.
 
Last edited:

MadViking

Member
170 MB/s is a proprietary SanDisk thing, and SanDisk specifies that you need a special card reader to hit those speeds. It doesn't work in a standard UHS-I slot.
Ok, I didn't know. Looks like around 90MB/s is the real limit. So no point in buying these expensive cards. Samsung Evo Plus and Kingston Select Plus seem to be the best with these limits and reasonably priced.
 

KAL2006

Banned
whats the switch got to do with the steam deck, no ones buying a steam deck to play Nintendo games so there's no comparrsion , yes it will probably plat Multiplatform games better but then so will a ps5/xbox.

i get this I a portable but it is still a big device how many kids are y ou going to see rocking this device . I have one pre ordered but I shall wait to see as I already have a ps5 currently and was in market for a switch depending on what other games releases this year that haven't ben dropped.

after watching the interviews on this device the looks on they're faces where like they were not happy and seem a bit cagey but that's just me. I'd really want to see what sort of running this does on a big screen because they are making a dock separately so they want you to also play it on a big screen .

mind you they did say you can buy any run of the mill dock and it ill works so I don't know why they are going to the effort of a rock being sold separately what is the actual need for them to waste their time and resources.

If there is a market for this time will tell. I feel like Valve have made this because handheld gaming is popular with the Switch. Many indie developers are having huge success on Switch. Even swaying some people who used to buy indie games on Steam now buying on Switch. This isn't supposed to replace PC gaming. This isn't supposed to be a way to play future next gen only games. However this device is for current Steam owners to play their library on the go and engage with the Steam store more. Portable gaming has become powerful enough to allow you to play PS4 power level games and pre PS4 generation games portably. This isn't only restricted to older gen gaming there are countless indie games and AA games as well as GaaS and cross gen games that will target lower powered laptops and PCs so this will run many even newer games coming out. I can see a successor to this handheld if it's success with AMD DLSS technology improving to allow even more demanding game to run. This is only phase one if this is a massive success with current Steam owners I can see Valve putting more marketing behind this and sell it through retail and other online stores.
 

Rob_27

Member
Did I read somewhere that the emmc memory is inserted into some converter and then in to what ever the nvme interface is on the bigger versions? Basically all use the same interface but the 64gb uses some type of piggyback May have dreamt it. Lol.
 

rnlval

Member
So at 88 GB/s its got
- more bandwidth than the Xbox One's DDR3 (68) which is usually ~ 900-720p
- over 3x the bandwidth of the Switch's LPDDR4 (25) which is usually 720-480p
- but half the bandwidth of the PS4's GDDR5 (176) which is usually ~ 1080-900p

Should be good enough for ~ 720p gaming but hurts the max frame rate potential. More bw = more fps.
SGM8nlb.jpg

RDNA v1 has "DCC Everywhere" improvements over VEGA and Polaris GCNs.

PS4 GPU does NOT have "DCC Everywhere"
PS4 GPU does NOT have ROPS being connected to a very fast L2 cache.
PS4 GPU does NOT have 1 MB L2 cache since R9-290X has 1 MB L2 cache. Vega 8 has 1 MB L2 cache.
 

rnlval

Member
You win the prize, it's 88 GB/s xD

Apparently they just clarified it. Take away maybe 20 GB/s for the CPU and that's about 42.5 GB/s per TF for the GPU.

For reference, for PS5 it'd be about 41.65 GB/s per TF (assuming the CPU is 20 GB/s; it's likely more considering it's higher-clocked and has more cores & threads), and for Series X it's about 46 GB/s per TF, possibly closer to 44.55 GB/s per TF with similar 20 GB/s for CPU usage factored out (though like PS5, the CPU will probably consume more than 20 GB/s since it is clocked higher and has more cores & threads).

That's very good overall for the SteamDeck, even if actual bandwidth might dip closer to (spitballing here) 36 GB/s per TF (if the CPU is going to need more than 20 GB/s, tho in SteamDecks' case I can't see that happening as it's only 4-core/8-thread and likely lower-clocked than PS5's CPU by a decent amount).

EDIT: Since the CPU is rated at 448 GFLOPs at max clock (3.5 GHz, higher than I was expecting), it should only pull around 24 GB/s - 25 GB/s under max load.



I'm guessing he's doing the Steam Deck #s given its range of 1 TF - 1.6 TF but looks like CPU usage isn't factored into this, so they're theoretical bandwidth peaks. Goes for all the systems listed there.

Well, technically they are very near real bandwidth usage cases but only when CPU is idle, for gaming the CPU is usually doing at least some type of transfer on the bus any portion of a given second, but the numbers are still right. I just took lower ends which thinking for a second might be too low because they assume CPU's maximum bandwidth which would mean GPU isn't accessing system memory, which is also not likely to be the case :p



TBF, the GPU in Steam Deck isn't necessarily a beast; it's 8 CUs and if they have to use, say, 4 CUs for decompression, that can bring a hit to graphical performance. DirectStorage does not need dedicated decompression hardware, but it would only help.

Also TBF, has it been confirmed there is no decompression block in this APU? We know it's custom between Valve and AMD, I'd like to think some type of decompressor has been built into it knowing what a hit the GPU would take it if it needed to do that.
Steam Deck also includes embedded DSP. Steam Deck is effectively handheld XSS/XSX/PS5.

60 hz would put a limit on the CPU usage.

Sony/MS could have released PS5 and XSS/XSX portable.
 

rnlval

Member
MS indeed plan to add GPU decompression (they are in talk with manufatures to hardware specialized decompression units in the future) but it is not implemented yet.
MS released the preview 3 days ago... you can test it if you want.
RTX 3080/3080 Ti/3090 has excess compute power for the given raster hardware.
 

Allandor

Member
So 88gb/s is not really much. If games would be developed for this device than it would be ok, but as this is just an ordinary PC in handheld form factor this won't be much.
Yes I Kno 720p but this device hast effectively less bandwidth than an xb1, because the xb1 at least had the esram where at least some bandwidth-hungry stuff could happen. Also as this is just a PC software will just copy stuff over to the "main memory" and duplicate the stuff in "GPU memory" which means effectively less bandwidth and memory.

This is really significant but thx to the small screen sacrifices won't be to prominent but don't expect wonders from the GPU.
A big cache would have done wonders for the GPU, but because of price and form factor this wouldn't be feasible.
I really don't know for whom this handheld is made and why valve did invest in such a thing. I really don't think that this has a chance on the market.
 

rnlval

Member
So 88gb/s is not really much. If games would be developed for this device than it would be ok, but as this is just an ordinary PC in handheld form factor this won't be much.
Yes I Kno 720p but this device hast effectively less bandwidth than an xb1, because the xb1 at least had the esram where at least some bandwidth-hungry stuff could happen. Also as this is just a PC software will just copy stuff over to the "main memory" and duplicate the stuff in "GPU memory" which means effectively less bandwidth and memory.

This is really significant but thx to the small screen sacrifices won't be to prominent but don't expect wonders from the GPU.
A big cache would have done wonders for the GPU, but because of price and form factor this wouldn't be feasible.
I really don't know for whom this handheld is made and why valve did invest in such a thing. I really don't think that this has a chance on the market.
XBO GCN doesn't have RDNA "DCC Everywhere".
Radeon VII GCN doesn't have RDNA "DCC Everywhere".

Instruction processing latency to retirement
GCN Vega = 12 clock cycles wave64

RDNA = 8 clock cycles for wave64
RDNA = 7 clock cycles for wave32

RDNA gains about 25% to 33% efficiency over Vega GCN

RX 5700 XT's 9.6 TFLOPS (RDNA) with 448 GB/s memory bandwidth nearly rivals Radeon VII's 13.44 TFLOPS (GCN) with 1,024 GB/s memory bandwidth,

The statement "Also as this is just a PC software will just copy stuff over to the "main memory" and duplicate the stuff in "GPU memory" which means effectively less bandwidth and memory." is obsolete.

Read https://www.slideshare.net/DICEStudio/framegraph-extensible-rendering-architecture-in-frostbite

framegraph-extensible-rendering-architecture-in-frostbite-52-638.jpg

DX11


framegraph-extensible-rendering-architecture-in-frostbite-53-638.jpg

DX12 PC


framegraph-extensible-rendering-architecture-in-frostbite-54-638.jpg

PS4


framegraph-extensible-rendering-architecture-in-frostbite-55-638.jpg

XBO

Notice DX12 PC's resource memory usage is similar to game consoles.
 
Last edited:

rnlval

Member
So 88gb/s is not really much. If games would be developed for this device than it would be ok, but as this is just an ordinary PC in handheld form factor this won't be much.
Yes I Kno 720p but this device hast effectively less bandwidth than an xb1, because the xb1 at least had the esram where at least some bandwidth-hungry stuff could happen. Also as this is just a PC software will just copy stuff over to the "main memory" and duplicate the stuff in "GPU memory" which means effectively less bandwidth and memory.

This is really significant but thx to the small screen sacrifices won't be to prominent but don't expect wonders from the GPU.
A big cache would have done wonders for the GPU, but because of price and form factor this wouldn't be feasible.
I really don't know for whom this handheld is made and why valve did invest in such a thing. I really don't think that this has a chance on the market.
GTX 750 Maxwell v1 with 1.11 TFLOPS FP32 and 80 GB/s memory bandwidth usually beat XBO.
GTX 950 Maxwell v2 with 1.80 TFLOPS FP32 and 105 GB/s memory bandwidth beats XBO.
GTX 1050 Pascal with 1.80 TFLOPS FP32 and 112 GB/s memory bandwidth beats XBO.

GTX 750 Maxwell v1 has the superior DCC (Delta Color Compression) and ROPS connected to 2 MB L2 Cache (missing on all GCNs prior to RX Vega).

NVIDIA has been improving DCC with each GPU generation.

AMD added DCC with Tonga GCN prior to Polaris GCN and it was garbage when compared to NVIDIA Pascal's DCC.
Polaris GCN's DCC is inferior when compared to NVIDIA Pascal's DCC.
Vega GCN's DCC is inferior when compared to NVIDIA Pascal's DCC.

RDNA DCC rivals Turing DCC. AMD implements "DCC Everywhere" with RDNA.
 
Last edited:

Allandor

Member
GTX 750 Maxwell v1 with 1.11 TFLOPS FP32 and 80 GB/s memory bandwidth usually beat XBO.
GTX 950 Maxwell v2 with 1.80 TFLOPS FP32 and 105 GB/s memory bandwidth beats XBO.
GTX 1050 Pascal with 1.80 TFLOPS FP32 and 112 GB/s memory bandwidth beats XBO.

GTX 750 Maxwell v1 has the superior DCC (Delta Color Compression) and ROPS connected to 2 MB L2 Cache (missing on all GCNs prior to RX Vega).

NVIDIA has been improving DCC with each GPU generation.

AMD added DCC with Tonga GCN prior to Polaris GCN and it was garbage when compared to NVIDIA Pascal's DCC.
Polaris GCN's DCC is inferior when compared to NVIDIA Pascal's DCC.
Vega GCN's DCC is inferior when compared to NVIDIA Pascal's DCC.

RDNA DCC rivals Turing DCC. AMD implements "DCC Everywhere" with RDNA.
In your "tests" with the GTX cards, all have a CPU that has a seperate memory pool and seperate bandwidth. The 88GB/s from the Steamdecks is shared bandwidth. This makes a big difference.
Yes, there is some more color-compression etc in the Chip but it still has not much bandwidth to work with
XBO GCN doesn't have RDNA "DCC Everywhere".
Radeon VII GCN doesn't have RDNA "DCC Everywhere".

Instruction processing latency to retirement
GCN Vega = 12 clock cycles wave64

RDNA = 8 clock cycles for wave64
RDNA = 7 clock cycles for wave32

RDNA gains about 25% to 33% efficiency over Vega GCN

RX 5700 XT's 9.6 TFLOPS (RDNA) with 448 GB/s memory bandwidth nearly rivals Radeon VII's 13.44 TFLOPS (GCN) with 1,024 GB/s memory bandwidth,

The statement "Also as this is just a PC software will just copy stuff over to the "main memory" and duplicate the stuff in "GPU memory" which means effectively less bandwidth and memory." is obsolete.

Read https://www.slideshare.net/DICEStudio/framegraph-extensible-rendering-architecture-in-frostbite

framegraph-extensible-rendering-architecture-in-frostbite-52-638.jpg

DX11


framegraph-extensible-rendering-architecture-in-frostbite-53-638.jpg

DX12 PC


framegraph-extensible-rendering-architecture-in-frostbite-54-638.jpg

PS4


framegraph-extensible-rendering-architecture-in-frostbite-55-638.jpg

XBO

Notice DX12 PC's resource memory usage is similar to game consoles.

I'm really not sure if the developers do not just mean "Graphics-memory". Normally on PC you have 2 pools and both are used and normally in those engine-memory-comaprisons you only look at the gpu memory. And the duplication of stuff from main to GPU memory might be obsolete for some engines but isn't for all so far. There are still plenty of games that use more memory on PC than the consoles have available (main memory + gpu memory combined).
Steamdeck is "just" a PC so games will not get adapted to run on this device. So the same limitations PC games have will be present on steam-deck. And yes, I know, this has just a tiny GPU so nobody is expecting ultra-details like from a GPU that has 8 GB alone. But than there is the price that is just to high. You get a decent laptopb for this price that can run games better (because of separate GDDR5 memory for the GPU).

Btw, so far RDNA 2 didn't scale that good in comparison to smaller and older GCN GPUs. GCN just gets more and more inefficient the more TF it had (e.g. Vega had much more TF (it generally had a much bigger chip design) than Polaris but couldn't bring the performance on the road). Yes, the new features (e.g. mesh shading etc) might save a lot of bandwidth etc, but it is still a very tiny GPU with not that much of bandwidth.
 

rnlval

Member
In your "tests" with the GTX cards, all have a CPU that has a seperate memory pool and seperate bandwidth. The 88GB/s from the Steamdecks is shared bandwidth. This makes a big difference.
Yes, there is some more color-compression etc in the Chip but it still has not much bandwidth to work with


I'm really not sure if the developers do not just mean "Graphics-memory". Normally on PC you have 2 pools and both are used and normally in those engine-memory-comaprisons you only look at the gpu memory. And the duplication of stuff from main to GPU memory might be obsolete for some engines but isn't for all so far. There are still plenty of games that use more memory on PC than the consoles have available (main memory + gpu memory combined).
Steamdeck is "just" a PC so games will not get adapted to run on this device. So the same limitations PC games have will be present on steam-deck. And yes, I know, this has just a tiny GPU so nobody is expecting ultra-details like from a GPU that has 8 GB alone. But than there is the price that is just to high. You get a decent laptopb for this price that can run games better (because of separate GDDR5 memory for the GPU).

Btw, so far RDNA 2 didn't scale that good in comparison to smaller and older GCN GPUs. GCN just gets more and more inefficient the more TF it had (e.g. Vega had much more TF (it generally had a much bigger chip design) than Polaris but couldn't bring the performance on the road). Yes, the new features (e.g. mesh shading etc) might save a lot of bandwidth etc, but it is still a very tiny GPU with not that much of bandwidth.
1. Those old GTX GPUs test usually has old CPUs with old DDR3 memory modules e.g.

Core i7-4770K with 16 GB DDR3-1600 (PC3-12800) with GTX 950. CPU has 25.6 GB/s. PCI-e version 3.0 has 16 GB/s IO from CPU to GPU direction.

Core i7-4770K with 16 GB DDR3-1600 with GTX 750 OC. CPU has 25.6 GB/s. PCI-e version 3.0 has 16 GB/s IO from CPU to GPU direction.

62.4 GB/s memory bandwidth for the GPU

60 hz would have limit CPU usage, hence CPU's memory bandwidth usage.

Zen 2 has 512 KB L2 cache per core, a total 2MB for quad-core.
Haswell has 256 L2 cache, per core a total 1MB for quad core.

Both Haswell quad-core and Zen 2 APU quad core have 8 MB L3 cache.

Higher cache storage lessens external bandwidth usage.

For Cyberpunk 2077 (DX12), RX Vega 8 (R4000) has beaten the GTX 1030 Desktop and rivaled GTX 1050

For Doom Eternal (Vulkan, supports Vega RPM, Async Compute), RX Vega 8 (R4000) has crushed the GTX 1030 Desktop and beaten GTX 1050.

For Metro Exodus 2019, RX Vega 8 (R4000) has beaten the GTX 1030 Desktop, and rivaled GTX 1050

Ryzen 4000s does not support 128 bit LPDDR5-5500.
RX Vega does not support "DCC Everywhere", VRS, Mesh Shaders.

------------------

2. EA DICE's resource lecture is an apple to apple comparison across API differences. PC DX12 enables the game programmer to minimize video-memory/system memory asset resource duplicates.

PC with dGPU still has shared system memory for VRAM overspill i.e. fast VRAM being used as the last level cache and desktop video memory swap storage when the game has full-screen exclusive mode.

16 GB is more than enough for DX11 era PC games.

Your argument is useless for RX Vega IGP since it's unified memory in the 1st place.

3. RX 5700 (7.7 TFLOPS with 448 GB/s memory bandwidth) slightly beaten Vega 64 (12.66 TFLOPS with 483 GB/s memory bandwidth). https://www.techpowerup.com/review/amd-radeon-rx-5700/28.html
 
Last edited:
Top Bottom