The 360 having separate packages was only possible by having direct access to each other's cache and aren't required to go through a memory bus. While I can't speak on Linux, unless they've replicated the xbox 360 cache scheme at the server level which would arguably be a larger accomplishment. It makes no sense from a performance stance and even if they did, the cost would be prohibitive for that kind of customization when there are many off the shelf options. Ignoring the fact that the cache differences don't mean they are the same chips they could have done any number of configurations to the cache as stated.

6 months.

PSNow is on Mac and PC. Remote Play is on android.

Predicting Sony will release a new higher tier of PSnow that will include PS5 games on all platforms including PS4 possibly before PS5 even releases.

At the very least PS4 owners will be able to demo PS5 games on their PS4s

@ronvalencia:

No the video is wrong, what he is suggesting is literally impossible the GPU and GPU cannot be on two separate packages while the CPU utilizes vRAM from the GPU. This configuration is only feasible when the cpu and GPU are single packages and bus with access to the same memory controller.

Edit: to be fair later in article they claim that the HBM2 is shared which contradicts what they claimed earlier but still state that they are separate packages. It sounds like there’s a mix of marketing and technical information. So will wait for more info but the fact that they contradict themselves on whether the HBM2 is just vRAM or not would be a basic spec to get.

@ronvalencia:

https://www.guru3d.com/news-story/amd-radeon-gpus-tapped-for-google-stadia-gamestreaming-platform.html

From AMD PR

Custom AMD high-performance Radeon datacenter GPUs for Google Stadia include:

Yes, the exact GPU I've been talking about for months now.

https://www.amd.com/en/products/professional-graphics/radeon-pro-v340

Radeon Pro v340

Gee, I wonder how I knew these gpus were getting rolled out in datacenters for games streaming months in advance.....?

@ronvalencia said:

@michaelmikado said:

@ronvalencia said:

@michaelmikado said:

Nah this is known. The full hardware was announced in October by AMD.

https://community.amd.com/community/radeon-pro-graphics/blog/2018/11/13/amd-server-cpus-gpus-the-ultimate-virtualization-solution

Anyone who is doing game streaming will be using these servers except for MS.

Footnote from the link.

Estimates based on SPECfp®_rate_base2017 using the GCC-02 v7.2 compiler. AMD- based system scored 201 in tests conducted in AMD labs using an “Ethanol” reference platform configured with 2 x AMD EPYC 7601 SOC’s ($4200 each at AMD 1 ku pricing), 512GB memory (16 x 32GB 2R DDR4 2666MHz), Ubuntu 17.04, BIOS 1002E. Intel- based Supermicro SYS-1029U-TRTP server scored 164 in tests conducted by AMD, configured with 2 x 8160 CPU’s (2 x $4702 each per ark.intel.com), 768GB memory (24 x 32GB 2R DDR4 2666MHz), SLES 12 SP3 4.4.92-6.18-default kernel, BIOS set to Extreme performance setting. NAP-77

CPU Specs: 7601 Epyc

# of CPU Cores32

# of Threads64

Base Clock2.2GHz

Max Boost Clock 3.2GHz

All Core Boost Speed 2.7Ghz

Socket Count1P/2P

PCI Express Versionx128

Default TDP / TDP180W

GPU Specs: v340

GPU Architecture VegaLithography 14nm Fin FETStream Processors7168Compute Units 112 = Vega 56x2 GPU Memory

Memory Size 32 GB Memory Type (GPU) HBM2

https://www.amd.com/en/products/cpu/amd-epyc-7601

Epyc 7601 with 32 CPU cores has 16MB L2 cache + 64 MB L3 cache.

Google's version

• Custom x86 processor clocked at 2.7GHz w/ AVX2 SIMD and 9.5MB of L2+L3 cache <-------- NOT Epyc 7601

Google's Stadia will be powered by the following specs:

• Custom x86 processor clocked at 2.7GHz w/ AVX2 SIMD and 9.5MB of L2+L3 cache
• Custom AMD GPU w/ HBM2 memory, 56 compute units, and 10.7TFLOPs
• 16GB of RAM (shared between CPU and GPU), up to 484GB/s of bandwidth <------- NOT a standard PC architecture
• SSD cloud storage

Reference: https://www.pcgamer.com/google-stadias-specs-and-latency-revealed/

PCGamer's half-assed reporting which they copied from DigitalFoundry is WRONG

• 16GB of RAM (shared between CPU and GPU), up to 484GB/s of bandwidth <------- NOT a standard PC architecture

Here's what Digital Foundry ACTUALLY said.

Right now, it's not clear if the 16GB of memory is for the whole system, or for GPU VRAM only

Google is being coy about their hardware. The vCPUs may be some type of special custom virtualization but the the hardware itself is anything but that. What it looks like to that they are running a virtualization layer for instances and splitting the available cache in some weird way to come up with these specs. I can almost guarantee they are essentially meaningless and "semi"-custom at best. Semi-custom generally means readily available hardware configurations with some minor tweets or pairings. My guess is they are purposely being coy because they will be upgrading to Rome Epyc CPUs as soon as they are available.

Yeah that's marketing BS trying to give an approximate per instance virtual allocation. It doesn't mean that is the actual underlying physical hardware. Digital Foundry already correctly guessed they are just running Vega 56s. I've already stated the cards are v340s which are just a pair of Vega 56s duct taped together for use in servers. These specs are for the cameras and those who wouldn't even know anything beyond counting TFlops. They aren't "technically" wrong, but they are marketing nonsense.

This is just the Google version of 3x the power of an XB1 in the cloud.

PCGamer's half-assed reporting which they copied from DigitalFoundry is WRONG

• 16GB of RAM (shared between CPU and GPU), up to 484GB/s of bandwidth <------- NOT a standard PC architecture

Here's what Digital Foundry ACTUALLY said.

Right now, it's not clear if the 16GB of memory is for the whole system, or for GPU VRAM only

Google is being coy about their hardware. The vCPUs may be some type of special custom virtualization but the the hardware itself is anything but that. What it looks like to that they are running a virtualization layer for instances and splitting the available cache in some weird way to come up with these specs. I can almost guarantee they are essentially meaningless and "semi"-custom at best. Semi-custom generally means readily available hardware configurations with some minor tweets or pairings. My guess is they are purposely being coy because they will be upgrading to Rome Epyc CPUs as soon as they are available.

@ronvalencia:

Right, that’s google’s fuzzy math at play. There’s no modern servers with less than 10MB of L1+L2 cache combined its likely some weird division they are doing of total cache and allocating it per instance. Maybe they got a 7601 with slightly less cache from AMD, who knows. The point is that’s basically the chip they are using.

No price would be more important

Because I don't think the average user of this service who is playing on a web browser on a Chromebook is looking for 4K/60. Users looking for 4K/60 will get a physical box to play in the best possible fidelity. The average Netflix user doesn't care that Infinity War is only in 1080P rather than 4K. What I'm saying is for the average user targeted user this is likely more than fine and latency will be more important to their experience than resolution.

Yeah but this is assuming at home speeds are non-pooled. As long as a user is getting 25Mbps even when peak, the speed should be fine. This service will most likely launch in NA first anyway and then hit Europe and Asia. But that time I would expect speeds to be well above 25Mbps average in home. The service isn't even launching until later this year, never mind when it official launches across the pond.