Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > sci.physics > #896219

Mac Neo: The dwarf with giant GPU muscles [macOS IOGPUFamily] (Re: The new MIMD Warp is the cherry on top)

Show all headers | View raw

Hi,

Interestingly you can half the AI Laptop
Budget now. The new Mac Neo is only $500 ,
half of my discount AI laptop. Here is

some measurement if this AI Laptop zoo,
including the Mac Neo dwarf (in milliseconds):

AI Laptop    example63
Ryzen    542.0
Neo    950.0
Yoga    1475.0
Think    2798.0

AI Laptop    example64
Ryzen    1141.0
Neo    2022.0
Yoga    2118.0
Think    7439.0

Since my test doesn't use much memory, it also
fits into into the Mac Neo dwarf 8 GB, and the
Mac Neo dwarf is extrem since the GPU can access

all of 8 GB, not only 50% and the memory is
integrated directly into the processor die,
it is still DRAM not yet HBM. The ARM instruction

set has been streamlined to better support
common GPU interfaces. The CPU acts as a pure
scheduler via native kernel extensions,

appending command buffers straight into memory
queues that the GPU's command processor
reads without a middleman.

Bye

Mild Shock schrieb:
> Hi,
> 
> But I do grouping of Hack VMs for my pi-WAM in
> 32 wide work groups. And the there are 4096
> / 32 = 128 such work groups.
> 
> Traditionally work groups were executed lockstep:
> 
> In GPU architecture, a warp (or wavefront in AMD
> terminology) is the fundamental unit of execution,
> typically comprising 32 scalar threads. Warps
> execute in a SIMT (Single Instruction, Multiple
> Thread) fashion, where all 32 threads execute
> the same instruction in synchronized lockstep
> over different data
> 
> New GPUs offer independent thread scheduling:
> 
> Modern GPU architectures (such as NVIDIA's Volta
> and later) feature Independent Thread Scheduling,
> maintaining independent execution states even for
> threads within the same warp. This allows the GPU
> to yield and resume threads dynamically, essentially
> acting as a hardware-level MIMD processor running
> on SIMD execution lanes.
> 
> I guess MIMD drastically increases the arithmetic
> bandwidth for control flow based WGSL code, while
> some group arrangement can also increase
> 
> the memory bandwidth. By kind of concurrently
> flushing L1/L2 caches and reloading L1/L2 caches,
> creating some simple sequential systolic computing.
> 
> At least I have used a memory layout where threads
> from a workgroup are adjacent. Standard processors
> continuously fetch data from memory, they suffer from the
> 
> "Von Neumann bottleneck". Systolic computing bypasses
>   this by feeding data into an array of Processing
> Elements (PEs) in a wave-like flow.
> 
> Bye
> 
> Mild Shock schrieb:
>>
>> Hi,
>>
>> You are a fucking moron, arent you?
>>
>> Most of the stuff in my pi-WAM happens
>> inside the L1 and L2 caches of the GPU.
>> Which is faster than normal RAM and has
>>
>> a wider von Neumann Neck. You can try
>> yourself, in case you find an AI Laptop
>> with similary specs as the Radeon 860M.
>>
>> The example is open source:
>>
>> 11.4 Giga Lips with a Budget Laptop
>> https://github.com/Jean-Luc-Picard-2021/gigabudget
>>
>> I already wrote pi-WAM is designed to
>> not use memory contention. In particalur
>> it also uses local variables like pc and
>>
>> accu of the idependent thread states,
>> which seems to be also pretty speedy.
>>
>> Bye
>>
>> Tanner Babadzhan schrieb:
>>  > Mild Shock wrote:
>>  >
>>  >> These novel GPUs , that are part of AI Laptops, feature Unified 
>> Memory
>>  >> Architecture (UMA).
>>  >> In the case of my Ryzen the main memory is 32 GB,
>>  >
>>  > you are in error talking bullshit, the bottleneck there is the max 
>> 4 GB/s,
>>  > 4 times by paralleling, however the proper gddr5/6 gpu arrays goes 
>> up to
>>  > 4,000 GB/s by parallel design
>>  >
>>  > GDDR7 (2025–2026 standard)
>>  >      Max Bandwidth: 1,792 GB/s (RTX 5090, 32GB)
>>  >
>>
> 

Back to sci.physics | Previous | Next — Previous in thread | Next in thread | Find similar | Unroll thread

Thread

Implementing Gas for a Compute Shader [Avoid TDR] (Re: AI dooms day escape: Güttinger Wald) Mild Shock <janburse@fastmail.fm> - 2026-07-09 10:49 +0200
  FYI: Unified Memory Architecture (UMA) (Re: Implementing Gas for a Compute Shader [Avoid TDR]) Mild Shock <janburse@fastmail.fm> - 2026-07-09 19:45 +0200
    he Wider von Neumann Neck (Re: FYI: Unified Memory Architecture (UMA)) Mild Shock <janburse@fastmail.fm> - 2026-07-10 07:49 +0200
      Mac Neo: The dwarf with giant GPU muscles [macOS IOGPUFamily] (Re: The new MIMD Warp is the cherry on top) Mild Shock <janburse@fastmail.fm> - 2026-07-10 08:29 +0200
      Dumbwit, just run it on your RTX 5070 trash (Re: he Wider von Neumann Neck) Mild Shock <janburse@fastmail.fm> - 2026-07-10 15:47 +0200
        We are waiting Dumbwit: 1 Month, 3 Months, .. [node.js dawn] (Re: Dumbwit, just run it on your RTX 5070 trash) Mild Shock <janburse@fastmail.fm> - 2026-07-10 16:29 +0200
          We are still waiting for results! [Confused Dumbwit] (Re: We are waiting Dumbwit: 1 Month, 3 Months, .. [node.js dawn]) Mild Shock <janburse@fastmail.fm> - 2026-07-10 22:47 +0200
            404 Brain not Found [GPU saturation] (Re: We are still waiting for results! [Confused Dumbwit]) Mild Shock <janburse@fastmail.fm> - 2026-07-10 23:42 +0200
              I nowhere talked about 10GB/s (Re: 404 Brain not Found [GPU saturation]) Mild Shock <janburse@fastmail.fm> - 2026-07-10 23:43 +0200
                I nowhere said something about AI (Re: I nowhere talked about 10GB/s) Mild Shock <janburse@fastmail.fm> - 2026-07-10 23:51 +0200
                Re: I nowhere said something about AI (Re: I nowhere talked about 10GB/s) Mild Shock <janburse@fastmail.fm> - 2026-07-11 00:09 +0200

csiph-web