Complete Accelerator Emporium
Intel — graphics
Early / fixed-function
- Intel i740
- Intel 810 / 815 graphics
- Intel 830M
- Intel Extreme Graphics
- Extreme Graphics 2
Gen graphics
- Gen4 / GMA 900
- Gen3 / GMA 950
- GMA 3000
- GMA X3000
- GMA X3100
- GMA 4500
Gen5–7
- Sandy Bridge
- Ivy Bridge
- Haswell
- Broadwell
- Skylake
- Kaby Lake
- Coffee Lake
- Ice Lake
Gen9–12
- Gen9
- Gen9.5
- Gen11
- Gen12
- UHD Graphics
- Iris Xe
Xe
- Xe-LP
- Xe-HP
- Xe-HPG
- Xe-HPC
- Arc Alchemist
- Arc Battlemage
- Data-center Max GPUs
Intel’s newer architecture is particularly interesting because Xe is simultaneously a graphics architecture and a compute architecture, making it a natural bridge between GPU emulation and accelerator simulation.
Intel Xeon Phi
This should not be buried under the Intel GPU section.
Give it a separate building.
Knights Corner
KNC
- Xeon Phi 5110P
- Xeon Phi 7120P
- 60+ x86 cores
- 512-bit vector units
- MIC architecture
Knights Landing
KNL
- Xeon Phi 7210
- 7230
- 7250
- 7290
Major architectural concepts:
x86 cores
│
├── 512-bit AVX-512
├── vector units
├── MCDRAM
├── mesh interconnect
└── many-core execution
Knights Mill
KNM
Specialized for machine learning.
This is one of the most fascinating emulator targets in the entire collection, because you’re no longer emulating a conventional GPU.
You’re emulating:
a many-core x86 vector accelerator.
That makes Xeon Phi the bridge between:
CPU ↔ vector processor ↔ GPU
AMD CPU / APU / GPU
AMD should actually have three wings.
AMD Radeon
We’ve already got:
R100
R200
R300
R400
R500
R600
R700
Evergreen
Northern Islands
Southern Islands
GCN
Polaris
Vega
RDNA
RDNA2
RDNA3
RDNA4
AMD APU
Don’t overlook these.
- Llano
- Trinity
- Richland
- Kaveri
- Carrizo
- Bristol Ridge
- Raven Ridge
- Picasso
- Renoir
- Cezanne
- Rembrandt
- Phoenix
- Hawk Point
- Strix Point
- Strix Halo
These are particularly valuable because the emulator has to deal with:
CPU
│
├── memory controller
│
├── system fabric
│
└── integrated GPU
That’s a much more interesting machine than an isolated PCI GPU.
AMD compute accelerators
Separate room:
AMD FirePro
- FirePro V-series
- FirePro W-series
- FirePro S-series
AMD Instinct
- MI25
- MI50
- MI60
- MI100
- MI200
- MI250
- MI250X
- MI300
- MI300X
- MI325
- MI350-class architectures
These are ideal architectural simulation targets, rather than trying to reproduce every physical device.
Intel compute accelerators
I’d add:
Intel GNA
Gaussian & Neural Accelerator.
Intel Movidius
- Myriad 1
- Myriad X
Intel NPU
Modern Core Ultra NPU generations.
Intel Data Center GPU
- Ponte Vecchio
- Max Series
- Flex Series
- Rialto Bridge
The Ponte Vecchio / Xe-HPC architecture deserves its own experimental emulator/simulator target.
ARM / Apple / Qualcomm — eventually
If the goal is an emporium rather than merely x86 GPU collection, I’d eventually add:
ARM Mali
- Mali-400
- Mali-T600
- Mali-T700
- Mali-G
- Mali-G7x
- Immortalis
Qualcomm Adreno
- Adreno 2xx
- 3xx
- 4xx
- 5xx
- 6xx
- 7xx
Apple
- Apple GPU
- A-series GPU generations
- M-series GPU generations
Those are especially interesting because their architectures don’t map cleanly onto the NVIDIA/AMD model.
And then the REALLY interesting section
I’d create a “Non-GPU Accelerator Hall.”
ACCELERATOR EMPORIUM
│
├── NVIDIA
│ ├── RIVA
│ ├── GeForce
│ ├── Tesla
│ ├── Fermi
│ ├── Kepler
│ ├── Maxwell
│ ├── Pascal
│ ├── Volta
│ ├── Turing
│ ├── Ampere
│ ├── Ada
│ └── Blackwell
│
├── AMD
│ ├── ATI
│ ├── Radeon
│ ├── FirePro
│ ├── GCN
│ ├── Vega
│ ├── RDNA
│ └── Instinct
│
├── INTEL
│ ├── i740
│ ├── i8xx
│ ├── GMA
│ ├── Gen
│ ├── Xe
│ └── Arc
│
├── INTEL PHI
│ ├── Knights Corner
│ ├── Knights Landing
│ └── Knights Mill
│
├── OTHER
│ ├── ARM Mali
│ ├── Qualcomm Adreno
│ ├── PowerVR
│ ├── Apple GPU
│ └── Vivante
│
└── COMPUTE
├── CUDA
├── ROCm
├── OpenCL
├── SYCL
├── Vulkan Compute
├── OpenMP offload
└── oneAPI
And I’d distinguish three different artifacts for every architecture:
1. Device emulator
PCI → MMIO → VRAM → registers → command processor
2. ISA simulator
shader/kernel
↓
instruction decoder
↓
execution model
3. Architectural simulator
cores
cache
memory
interconnect
scheduler
latency
bandwidth
occupancy
That distinction is important for Xeon Phi in particular. A KNL simulator doesn’t need to pretend it’s a graphics card; it needs to reproduce the many-core/vector machine.
The ultimate collection would therefore be less “GPU emulator museum” and more:
An executable history of heterogeneous computing — from VGA and i740, through GeForce/Radeon, through GCN/Xe, all the way to Xeon Phi, Instinct, CUDA, and modern accelerator architectures.
