GPU chips

Introduction

Each nvidia GPU has several identifying numbers that can be used to determine supported features, the engines it contains, and the register set. The most important of these numbers is an 8-bit number known as the “GPU id”. If two cards have the same GPU id, their GPUs support identical features, engines, and registers, with very minor exceptions. Such cards can however still differ in the external devices they contain: output connectors, encoders, capture chips, temperature sensors, fan controllers, installed memory, supported clocks, etc. You can get the GPU id of a card by reading from its PMC area.

The GPU id is usually written as NVxx, where xx is the id written as uppercase hexadecimal number. Note that, while cards before NV10 used another format for their ID register and don’t have the GPU id stored directly, they are usually considered as NV1-NV5 anyway.

Nvidia uses “GPU code names” in their materials. They started out identical to the GPU id, but diverged midway through the NV40 series and started using a different numbering. However, for the most part nvidia code names correspond 1 to 1 with the GPU ids.

The GPU id has a mostly one-to-many relationship with pci device ids. Note that the last few bits [0-6 depending on GPU] of PCI device id are changeable through straps [see PSTRAPS: straps readout and override]. When pci ids of a GPU are listed in this file, the following shorthands are used:

1234
PCI device id 0x1234
1234*
PCI device ids 0x1234-0x1237, choosable by straps
123X
PCI device ids 0x1230-0x123X, choosable by straps
124X+
PCI device ids 0x1240-0x125X, choosable by straps
124X*
PCI device ids 0x1240-0x127X, choosable by straps

The GPU families

The GPUs can roughly be grouped into a dozen or so families: NV1, NV3/RIVA, NV4/TNT, Celsius, Kelvin, Rankine, Curie, Tesla, Fermi, Kepler, Maxwell and Pascal. This aligns with big revisions of PGRAPH, the drawing engine of the card. While most functionality was introduced in sync with PGRAPH revisions, some other functionality [notably video decoding hardware] gets added in GPUs late in a GPU family and sometimes doesn’t even get to the first GPU in the next GPU family. For example, NV11 expanded upon the previous NV15 chipset by adding dual-head support, while NV20 added new PGRAPH revision with shaders, but didn’t have dual-head - the first GPU to feature both was NV25.

Also note that a bigger GPU id doesn’t always mean a newer card / card with more features: there were quite a few places where the numbering actually went backwards. For example, NV11 came out later than NV15 and added several features.

Nvidia’s card release cycle always has the most powerful high-end GPU first, subsequently filling in the lower-end positions with new cut-down GPUs. This means that newer cards in a single sub-family get progressively smaller, but also more featureful - the first GPUs to introduce minor changes like DX10.1 support or new video decoding are usually the low-end ones.

Whenever a range of GPUs is mentioned in the documentation, it’s written as “NVxx:NVyy”. This is left-inclusive, right-noninclusive range of GPU ids as sorted in the following list. For example, G200:GT218 means GPUs G200, MCP77, MCP79, GT215, GT216. NV20:NV30 effectively means all NV20 family GPUs.

The full known GPU list, sorted roughly according to introduced features, is:

  • NV1 family: NV1
  • NV3 (aka RIVA) family: NV3, NV3T
  • NV4 (aka TNT) family: NV4, NV5
  • Celsius family: NV10, NV15, NV1A, NV11, NV17, NV1F, NV18
  • Kelvin family: NV20, NV2A, NV25, NV28
  • Rankine family: NV30, NV35, NV31, NV36, NV34
  • Curie family:
    • NV40 subfamily: NV40, NV45, NV41, NV42, NV43, NV44, NV44A
    • G70 subfamily: G70, G71, G73, G72
    • the IGPs: C51, MCP61, MCP67, MCP68, MCP73
    • the special snowflake: RSX
  • Tesla family:
    • G80 subfamily: G80
    • G84 subfamily: G84, G86, G92, G94, G96, G98
    • G200 subfamily: G200, MCP77, MCP79
    • GT215 subfamily: GT215, GT216, GT218, MCP89
  • Fermi family:
    • GF100 subfamily: GF100, GF104, GF106, GF114, GF116, GF108, GF110
    • GF119 subfamily: GF119, GF117
  • Kepler family: GK104, GK107, GK106, GK110, GK110B, GK208, GK208B, GK20A, GK210
  • Maxwell family: GM107, GM108, GM200, GM204, GM206, GM20B
  • Pascal family: GP100, GP102, GP104, GP106, GP107, GP108

NV1 family: NV1

gpu-gen NV1

The first generation of nVidia GPUs. Includes only one GPU – the NV1. It has semi-legendary status, as it’s very rare and hard to get. The GPU is also known by its SGS-Thomson code number, STG-2000. The most popular card using this GPU is Diamond EDGE 3D.

This GPU is unusual for multiple reasons:

  • It has a builtin sound mixer with a MIDI synthetizer (aka PAUDIO). It is supposed to be paired with an audio codec (AD1848) for full integrated soundcard functionality.
  • It is not fully VGA-compatible – there is some VGA emulation, but it’s quite rough and many features are not supported.
  • It has no integrated DAC or clock generators – it has to be paired with an accompanying external DAC, the STG-1732 or STG-1764 that will convert raw framebuffer contents to display pixels. It is also charged with generating the clocks for the GPU.
  • The accompanying DAC chip also contains game port functionality, for a complete soundcard replacement.
  • As if the game port was not enough, the DAC also supports two Sega Saturn controller ports.
  • The so-called 3D engine renders textured quadratic surfaces, instead of triangles (as opposed to all later GPUs). Rendering triangles with it is pretty much impossible.

The GPU was jointly manufactured by SGS-Thomson and nVidia, and uses SGS’ PCI vendor ID (there are apparently variants using nVidia’s vendor id, but not much is known about these).

There’s also NV2, which has even more legendary status. It was supposed to be another card based on quadratic surfaces, but it got stuck in development hell and never got released. Apparently it never got to the stage of functioning silicon. The device id of NV2 was supposed to be 0x0010.

NV3 (RIVA) family: NV3, NV3T

gpu-gen NV3

The first [moderately] sane GPUs from nvidia, and also the first to use AGP bus. There are two chips in this family, and confusingly both use GPU id NV3, but can be told apart by revision. The original NV3 is used in RIVA 128 cards, while the revised NV3, known as NV3T, is used in RIVA 128 ZX. NV3 supports AGP 1x and a maximum of 4MB of VRAM, while NV3T supports AGP 2x and 8MB of VRAM. NV3T also increased number of slots in PFIFO cache. These GPUs were also manufactured by SGS-Thomson and bear the code name of STG-3000.

The NV3 GPU is made of the following functional blocks:

  • host interface, connected to the host machine via PCI or AGP
  • two PLLs, to generate video pixel clock and memory clock
  • memory interface, connected to 2MB-8MB of external VRAM via 64-bit or 128-bit memory bus, shared with an 8-bit parallel flash ROM
  • PFIFO, controlling command submission to PGRAPH and gathering commands through DMA to host memory or direct MMIO submission
  • PGRAPH, the 2d/3d drawing engine, supporting windows GDI and Direct3D 5 acceleration
  • VGA-compatible CRTC, RAMDAC, and associated video output circuitry, enabling direct connection of VGA analog displays and TV connection via an external AD722 encoder chip
  • i2c bus to handle DDC and control mediaport devices
  • double-buffered video overlay and cursor circuitry in RAMDAC
  • mediaport, a proprietary interface with ITU656 compatibility mode, allowing connection of external video capture or MPEG2 decoding chip

NV3 introduced RAMIN, an area of memory at the end of VRAM used to hold various control structures for PFIFO and PGRAPH. On NV3, RAMIN can be accessed in BAR1 at addresses starting from 0xc00000, while later cards have it in BAR0. It also introduced DMA objects, a RAMIN structure used to define a VRAM or host memory area that PGRAPH is allowed to use when executing commands on behalf of an application. These early DMA objects are limitted to linear VRAM and paged host memory objects, and have to be switched manually by host. See NV3 DMA objects for details.

NV4 (TNT) family: NV4, NV5

gpu-gen NV4

Improved and somewhat redesigned NV3. Notable changes:

  • AGP x4 support
  • redesigned and improved DMA command submission
  • separated core and memory clocks
  • DMA objects made more orthogonal, and switched automatically by card
  • redesigned PGRAPH objects, introducing the concept of object class in hardware
  • added BIOS ROM shadow in RAMIN
  • Direct3D 6 / multitexturing support in PGRAPH
  • bumped max supported VRAM to 16MB
  • [NV5] bumped max supported VRAM to 32MB
  • [NV5] PGRAPH 2d context object binding in hardware

This family includes the original NV4, used in RIVA TNT cards, and NV5 used in RIVA TNT2 and Vanta cards.

Celsius family: NV10, NV15, NV1A, NV11, NV17, NV1F, NV18

gpu-gen Celsius

The notable changes in this generation are:

  • NV10:
    • redesigned memory controller
    • max VRAM bumped to 128MB
    • redesigned VRAM tiling, with support for multiple tiled regions
    • greatly expanded 3d engine: hardware T&L, D3D7, and other features
    • GPIO pins introduced for ???
    • PFIFO: added REF_CNT and NONINC commands
    • added PCOUNTER: the performance monitoring engine
    • new and improved video overlay engine
    • redesigned mediaport
  • NV15:
    • introduced vblank wait PGRAPH commands
    • minor 3d engine additions [logic operation, ...]
  • NV1A:
    • big endian mode
    • PFIFO: semaphores and subroutines
  • NV11:
    • dual head support, meant for laptops with flat panel + external display
  • NV17:
    • builtin TV encoder
    • ZCULL
    • added VPE: MPEG2 decoding engine
  • NV18:
    • AGP x8 support
    • second straps set

Todo

what were the GPIOs for?

pciid GPU pixel pipelines and ROPs texture units date notes
0100* NV10 4 4 11.10.1999 the first GeForce card [GeForce 256]
0150* NV15 4 8 26.04.2000 the high-end card of GeForce 2 lineup [GeForce 2 Ti, ...]
01a0* NV1A 2 4 04.06.2001 the IGP of GeForce 2 lineup [nForce]
0110* NV11 2 4 28.06.2000 the low-end card of GeForce 2 lineup [GeForce 2 MX]
017X NV17 2 4 06.02.2002 the low-end card of GeForce 4 lineup [GeForce 4 MX]
01fX NV1F 2 4 01.10.2002 the IGP of GeForce 4 lineup [nForce 2]
018X NV18 2 4 25.09.2002 like NV17, but with added AGP x8 support

NV1A and NV1F are IGPs and lack VRAM, memory controller, mediaport, and ROM interface. They use the internal interfaces of the northbridge to access an area of system memory set aside as fake VRAM and BIOS image.

Kelvin family: NV20, NV2A, NV25, NV28

gpu-gen Kelvin

The first cards of this family were actually developed before NV17, so they miss out on several features introduced in NV17. The first card to merge NV20 and NV17 additions is NV25. Notable changes:

  • NV20:
    • no dual head support again
    • no PTV, VPE
    • no ZCULL
    • a new memory controller with Z compression
    • RAMIN reversal unit bumped to 0x40 bytes
    • 3d engine extensions:
      • programmable vertex shader support
      • D3D8, shader model 1.1
    • PGRAPH automatic context switching
  • NV25:
    • a merge of NV17 and NV20: has dual-head, ZCULL, ...
    • still no VPE and PTV
  • NV28:
    • AGP x8 support

The GPUs are:

pciid GPU vertex shaders pixel pipelines and ROPs texture units date notes
0200* NV20 1 4 8 27.02.2001 the only GPU of GeForce 3 lineup [GeForce 3 Ti, ...]
02a0* NV2A 2 4 8 15.11.2001 the XBOX IGP [XGPU]
025X NV25 2 4 8 06.02.2002 the high-end GPU of GeForce 4 lineup [GeForce 4 Ti]
028X NV28 2 4 8 20.01.2003 like NV25, but with added AGP x8 support

NV2A is a GPU designed exclusively for the original xbox, and can’t be found anywhere else. Like NV1A and NV1F, it’s an IGP.

Todo

verify all sorts of stuff on NV2A

Rankine family: NV30, NV35, NV31, NV36, NV34

gpu-gen Rankine

The infamous GeForce FX series. Notable changes:

  • NV30:
    • 2-stage PLLs introduced [still located in PRAMDAC]
    • max VRAM size bumped to 256MB
    • 3d engine extensions:
      • programmable fragment shader support
      • D3D9, shader model 2.0
    • added PEEPHOLE indirect memory access
    • return of VPE and PTV
    • new-style memory timings
  • NV35:
    • 3d engine now supports depth bounds check
  • NV31:
    • no NV35 changes, this GPU is derived from NV30
    • 2-stage PLLs split into two registers
    • VPE engine extended to work as a PFIFO engine
  • NV36:
    • a merge of NV31 and NV35 changes from NV30
  • NV34:
    • a comeback of NV10 memory controller!
    • NV10-style mem timings again
    • no Z compression again
    • RAMIN reversal unit back at 16 bytes
    • 3d engine additions:
      • ???

Todo

figure out NV34 3d engine changes

The GPUs are:

pciid GPU vertex shaders pixel pipelines and ROPs date notes
030X NV30 2 8 27.01.2003 high-end GPU [GeForce FX 5800]
033X NV35 3 8 12.05.2003 very high-end GPU [GeForce FX 59X0]
031X NV31 1 4 06.03.2003 low-end GPU [GeForce FX 5600]
034X NV36 3 4 23.10.2003 middle-end GPU [GeForce FX 5700]
032X NV34 1 4 06.03.2003 low-end GPU [GeForce FX 5200]

The pci vendor id is 0x10de.

Curie family

gpu-gen Curie

This family was the first to feature PCIE cards, and many fundamental areas got significant changes, which later paved the way for G80. It is also the family where GPU ids started to diverge from nvidia code names. The changes:

  • NV40:
    • RAMIN bumped in size to max 16MB, many structure layout changes
    • RAMIN reversal unit bumped to 512kB
    • 3d engine: support for shader model 3 and other additions
    • Z compression came back
    • PGRAPH context switching microcode
    • redesigned clock setup
    • separate clock for shaders
    • rearranged PCOUNTER to handle up to 8 clock domains
    • PFIFO cache bumped in size and moved location
    • added independent PRMVIO for two heads
    • second set of straps added, new strap override registers
    • new PPCI PCI config space access window
    • MPEG2 encoding capability added to VPE
    • FIFO engines now identify the channels by their context addresses, not chids
    • BIOS uses all-new BIT structure to describe the card
    • individually disablable shader and ROP units.
    • added PCONTROL area to... control... stuff?
    • memory controller uses NV30-style timings again
  • NV41:
    • introduced context switching to VPE
    • introduced PVP1, microcoded video processor
    • first natively PCIE card
    • added PCIE GART to memory controller
  • NV43:
    • added a thermal sensor to the GPU
  • NV44:
    • a new PCIE GART page table format
    • 3d engine: ???
  • NV44A:
    • like NV44, but AGP instead of PCIE

Todo

more changes

Todo

figure out 3d engine changes

The GPUs are:

pciid GPU id GPU names vertex shaders pixel shaders ROPs date notes
004X 021X 0x40/0x45 NV40/NV45/NV48 6 16 16 14.04.2004 AGP
00cX 0x41/0x42 NV41/NV42 5 12 12 08.11.2004  
014X 0x43 NV43 3 8 4 12.08.2004  
016X 0x44 NV44 3 4 2 15.12.2004 TURBOCACHE
022X 0x4a NV44A 3 4 2 04.04.2005 AGP
009X 0x47 G70 8 24 16 22.06.2005  
01dX 0x46 G72 3 4 2 18.01.2006 TURBOCACHE
029X 0x49 G71 8 24 16 09.03.2006  
039X 0x4b G73 8 12 8 09.03.2006  
024X 0x4e C51 1 2 1 20.10.2005 IGP, TURBOCACHE
03dX 0x4c MCP61 1 2 1 ??.06.2006 IGP, TURBOCACHE
053X 0x67 MCP67 1 2 2 01.02.2006 IGP, TURBOCACHE
053X 0x68 MCP68 1 2 2 ??.07.2007 IGP, TURBOCACHE
07eX 0x63 MCP73 1 2 2 ??.07.2007 IGP, TURBOCACHE
- 0x4d RSX ? ? ? 11.11.2006 FlexIO bus interface, used in PS3

Todo

all geometry information unverified

Todo

any information on the RSX?

It’s not clear how NV40 is different from NV45, or NV41 from NV42, or MCP67 from MCP68 - they even share pciid ranges.

The NV4x IGPs actually have a memory controller as opposed to earlier ones. This controller still accesses only host memory, though.

As execution units can be disabled on NV40+ cards, these configs are just the maximum configs - a card can have just a subset of them enabled.

Tesla family

gpu-gen Tesla

The card where they redesigned everything. The most significant change was the redesigned memory subsystem, complete with a paging MMU [see Tesla virtual memory].

  • G80:
    • a new VM subsystem, complete with redesigned DMA objects
    • RAMIN is gone, all structures can be placed arbitrarily in VRAM, and usually host memory memory as well
    • all-new channel structure storing page tables, RAMFC, RAMHT, context pointers, and DMA objects
    • PFIFO redesigned, PIO mode dropped
    • PGRAPH redesigned: based on unified shader architecture, now supports running standalone computations, D3D10 support, unified 2d acceleration object
    • display subsystem reinvented from scratch: a stub version of the old VGA-based one remains for VGA compatibility, the new one is not VGA based and is controlled by PFIFO-like DMA push buffers
    • memory partitions tied directly to ROPs
  • G84:
    • redesigned channel structure with a new layout
    • got rid of VP1 video decoding and VPE encoding support, but VPE decoder still exists
    • added VP2 xtensa-based programmable video decoding and BSP engines
    • removed restrictions on host memory access by rendering: rendering to host memory and using blocklinear textures from host are now ok
    • added VM stats write support to PCOUNTER
    • PEEPHOLE moved out of PBUS
    • PFIFO_BAR_FLUSH moved out of PFIFO
  • G98:
    • introduced VP3 video decoding engines, and the falcon microcode with them
    • got rid of VP2 video decoding
  • G200:
    • developped in parallel with G98
    • VP2 again, no VP3
    • PGRAPH rearranged to make room for more MPs/TPCs
    • streamout enhancements [ARB_transform_feedback2]
    • CUDA ISA 1.3: 64-bit g[] atomics, s[] atomics, voting, fp64 support
  • MCP77:
    • merged G200 and G98 changes: has both VP3 and new PGRAPH
    • only CUDA ISA 1.2 now: fp64 support got cut out again
  • GT215:
    • a new revision of the falcon ISA
    • a revision to VP3 video decoding, known as VP4. Adds MPEG-4 ASP support.
    • added PDAEMON, a falcon engine meant to do card monitoring and power maanagement
    • PGRAPH additions for D3D10.1 support
    • added HDA audio codec for HDMI sound support, on a separate PCI function
    • Added PCOPY, the dedicated copy engine
    • Merged PSEC functionality into PVLD
  • MCP89:
    • added PVCOMP, the video compositor engine

The GPUs in this family are:

core hda id name TPCs MPs/TPC PARTs date notes
pciid pciid              
019X - 0x50 G80 8 2 6 08.11.2006  
040X - 0x84 G84 2 2 2 17.04.2007  
042X - 0x86 G86 1 2 2 17.04.2007  
060X+ - 0x92 G92 8 2 4 29.10.2007  
062X+ - 0x94 G94 4 2 4 29.07.2008  
064X+ - 0x96 G96 2 2 2 29.07.2008  
06eX+ - 0x98 G98 1 1 1 04.12.2007  
05eX+ - 0xa0 G200 10 3 8 16.06.2008  
084X+ - 0xaa MCP77/MCP78 1 1 1 ??.06.2008 IGP
086X+ - 0xac MCP79/MCP7A 1 2 1 ??.06.2008 IGP
0caX+ 0be4 0xa3 GT215 4 3 2 15.06.2009  
0a2X+ 0be2 0xa5 GT216 2 3 2 15.06.2009  
0a6X+ 0be3 0xa8 GT218 1 2 1 15.06.2009  
08aX+ - 0xaf MCP89 2 3 2 01.04.2010 IGP

Like NV40, these are just the maximal numbers.

Todo

geometry information not verified for G94, MCP77

Fermi/Kepler/Maxwell/Pascal family

gpu-gen Fermi

The card where they redesigned everything again.

  • GF100:
    • redesigned PFIFO, now with up to 3 subfifos running in parallel
    • redesigned PGRAPH:
      • split into a central HUB managing everything and several GPCs doing all actual work
      • GPCs further split into a common part and several TPCs
      • using falcon for context switching
      • D3D11 support
    • redesigned memory controller
      • split into three parts:
        • per-partition low-level memory controllers [PBFB]
        • per-partition middle memory controllers: compression, ECC, ... [PMFB]
        • a single “hub” memory controller: VM control, TLB control, ... [PFFB]
    • memory partitions, GPCs, TPCs have independent register areas, as well as “broadcast” areas that can be used to control all units at once
    • second PCOPY engine
    • redesigned PCOUNTER, now having multiple more or less independent subunits to monitor various parts of GPU
    • redesigned clock setting
    • ...
  • GF119:
    • a major revision to VP3 video decoding, now called VP5. vµc microcode removed.
    • another revision to the falcon ISA, allowing 24-bit PC
    • added PUNK1C3 falcon engine
    • redesigned I2C bus interface
    • redesigned PDISPLAY
    • removed second PCOPY engine
  • GF117:
    • PGRAPH changes:
      • ???
gpu-gen Kepler

An upgrade to Fermi.

  • GK104:
    • redesigned PCOPY: the falcon controller is now gone, replaced with hardware control logic, partially in PFIFO
    • an additional PCOPY engine
    • PFIFO redesign - a channel can now only access a single engine selected on setup, with PCOPY2+PGRAPH considered as one engine
    • PGRAPH changes:
      • subchannel to object assignments are now fixed
      • m2mf is gone and replaced by a new p2mf object that only does simple upload, other m2mf functions are now PCOPY’s responsibility instead
      • the ISA requires explicit scheduling information now
      • lots of setup has been moved from methods/registers into memory structures
      • ???
  • GK110:
    • PFIFO changes:
      • ???
    • PGRAPH changes:
      • ISA format change
      • ???

Todo

figure out PGRAPH/PFIFO changes

gpu-gen Maxwell
gpu-gen Pascal

GPUs in Fermi/Kepler/Maxwell/Pascal families:

core hda id name GPCs TPCs PARTs MCs ZCULLs PCOPYs HEADs UNK7 PPCs SUBPs SPOONs CE0 CE1 CE2 date
pciid pciid       /GPC     /GPC       /GPC /PART          
06cX+ 0be5 0xc0 GF100 4 4 6 [6] [4] [2] [2] - - 2 3 0 0 - 26.03.2010
0e2X+ 0beb 0xc4 GF104 2 4 4 [4] [4] [2] [2] - - 2 3 0? 0? - 12.07.2010
120X+ 0e0c 0xce GF114 2 4 4 [4] [4] [2] [2] - - 2 3 0? 0? - 25.01.2011
0dcX+ 0be9 0xc3 GF106 1 4 3 [3] [4] [2] [2] - - 2 3 3 4 - 03.09.2010
124X+ 0bee 0xcf GF116 1 4 3 [3] [4] [2] [2] - - 2 3 3 4 - 15.03.2011
0deX+ 0bea 0xc1 GF108 1 2 1 2 4 [2] [2] - - 2 1 3 4 - 03.09.2010
108X+ 0e09 0xc8 GF110 4 4 6 [6] [4] [2] [2] - - 2 3 0 0 - 07.12.2010
104X* 0e08 0xd9 GF119 1 1 1 1 4 1 2 - - 1 1 3 - - 05.01.2011
1140 - 0xd7 GF117 1 2 1 1 4 1 -[4] - 1 2 1 3 - - ??.04.2012
118X* 0e0a 0xe4 GK104 4 2 4 4 4 3 4 - 1 4 3 ? 3 3 22.03.2012
0fcX* 0e1b 0xe7 GK107 1 2 2 2 4 3 4 - 1 4 3 3 ? 3 24.04.2012
11cX+ 0e0b 0xe6 GK106 3 2 3 3 4 3 4 - 1 4 3 3 ? 3 22.04.2012
100X+ 0e1a 0xf0 GK110 5 3 6 6 4 3 4 - 2 4 3 ? ? ? 21.02.2013
100X+ 0e1a 0xf1 GK110B 5 3 6 6 4 3 4 - 2 4 3 ? 3 3 07.11.2013
???? ???? ???? GK210 ? ? ? ? ? ? ? - ? ? ? ? ? ? ?
128X+ 0e0f 0x108 GK208 1 2 1 1 4 3 4 - 1 2 2 3 ? 3 19.02.2013
128X+ 0e0f 0x106 GK208B 1 2 1 1 4 3 4 - 1? 2? 2? 3 ? 3 ???
- - 0xea GK20A 1 1 1 1 4 3 -[4] - 1 1 1 -? -? 3 ?
138X+ 0fbc 0x117 GM107 1 5 2 2 4 3 4 1 2 4 2 3 ? 3 18.02.2014
134X+ ???? 0x118 GM108 1 3 1 1 4 3 4 0 ? ? 2 3 ? 3 ?
17cX+ 0fb0 0x120 GM200 ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?
13cX+ 0fbb 0x124 GM204 ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?
140X+ 0fba 0x126 GM206 ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?
- - 0x12b GM20B ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?
158X# ???? 0x130 GP100 ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?
1b0X# 10ef 0x132 GP102 ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?
1b8X# 10f0 0x134 GP104 4 5 4 4 4 4 4 2 ? ? ? ? ? ? ?
1c0X# 10f1 0x136 GP106 ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?
1c8X# 0fb9 0x137 GP107 ? ? ? ? ? ? ? ? ? ? ? ? ? ? 10.25.2016
1d0X# 0fb8 0x138 GP108 ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?

Todo

it is said that one of the GPCs [0th one] has only one TPC on GK106

Todo

what the fuck is GK110B? and GK208B?

Todo

GK210

Todo

GK20A

Todo

GM20x, GP10x

Todo

another design counter available on GM107, another 4 on GP10x

Comparison table

Name GPU id GPU generation Release date [approximate] Bus interface PCI vendor id PCI device IDs HDA PCI device id BIOS version prefix FB type # of FB partitions # of MCs per FB partition # of SUBPs per FB partition # of XF units # of GPCs # of TPCs [per GPC for Fermi+] # of SMs per TPC # of PPCs per GPC # of CEs
NV1 - NV1 09.1995 Pci 0x104a 0x0008-0x0009 - - NV1 - - - - - - - - -
NV3 - NV3 04.1997 Pci 0x12d2 0x0018-0x0019 - - NV3 - - - - - - - - -
NV3T - NV3 23.02.1998 Pci 0x12d2 0x0018-0x0019 - - NV3T - - - - - - - - -
NV4 - NV4 23.03.1998 Pci 0x10de 0x0020 - - NV4 - - - - - - - - -
NV5 - NV4 15.03.1999 Pci 0x10de 0x0028-0x002b - 02.05 NV5 - - - - - - - - -
NV6 - NV4 15.03.1999 Pci 0x10de 0x002c-0x002f - 02.05 NV5 - - - - - - - - -
NVA - NV4 08.09.1999 Pci 0x10de 0x00a0 - ? NVA - - - - - - - - -
NV10 0x010 Celsius 11.10.1999 Pci 0x10de 0x0100-0x0103 - 02.10 NV10 - - - 1 - - - - -
NV15 0x015 Celsius 26.04.2000 Pci 0x10de 0x0150-0x0153 - 03.15 NV10 - - - 1 - - - - -
NV1A 0x01a Celsius 04.06.2001 Pci 0x10de 0x01a0-0x01a3 - 03.1a NV1A - - - 1 - - - - -
NV11 0x011 Celsius 28.06.2000 Pci 0x10de 0x0110-0x0113 - 03.11 NV10 - - - 1 - - - - -
NV17 0x017 Celsius 06.02.2002 Pci 0x10de 0x0170-0x017f - 04.17 NV10 - - - 1 - - - - -
NV1F 0x01f Celsius 01.10.2002 Pci 0x10de 0x01f0-0x01ff - 04.1f NV1F - - - 1 - - - - -
NV18 0x018 Celsius 25.09.2002 Pci 0x10de 0x0180-0x018f - 04.18 NV10 - - - 1 - - - - -
NV20 0x020 Kelvin 27.02.2001 Pci 0x10de 0x0200-0x0203 - 03.20 NV20 4 - - 1 - 2 - - -
NV2A 0x02a Kelvin 15.11.2001 Pci 0x10de 0x02a0-0x02a3 - - NV2A 4 - - 2 - 2 - - -
NV25 0x025 Kelvin 06.02.2002 Pci 0x10de 0x0250-0x025f - 04.25 NV25 4 - - 2 - 2 - - -
NV28 0x028 Kelvin 20.01.2003 Pci 0x10de 0x0280-0x028f - 04.28 NV25 4 - - 2 - 2 - - -
NV30 0x030 Rankine 27.01.2003 Pci 0x10de 0x0300-0x030f - 04.30 NV30 4 - - 2 - 2 - - -
NV35 0x035 Rankine 12.05.2003 Pci 0x10de 0x0330-0x033f - 04.35 NV35 4 - - 3 - 2 - - -
NV31 0x031 Rankine 06.03.2003 Pci 0x10de 0x0310-0x031f - 04.31 NV31 2 - - 1 - 1 - - -
NV36 0x036 Rankine 23.10.2003 Pci 0x10de 0x0340-0x034f - 04.36 NV36 2 - - 3 - 1 - - -
NV34 0x034 Rankine 06.03.2003 Pci 0x10de 0x0320-0x032f - 04.34 NV10 - - - 1 - 1 - - -
NV40 0x040 Curie 14.04.2004 Pci 0x10de 0x0040-0x004f - 05.40 NV40 4 - - 6 - 4 - - -
NV45 0x045 Curie 14.04.2004 Pci 0x10de 0x0040-0x004f - 05.40 NV40 4 - - 6 - 4 - - -
NV41 0x041 Curie 08.11.2004 Pcie 0x10de 0x00c0-0x00cf - 05.41 NV41 4 - - 5 - 3 - - -
NV42 0x042 Curie 08.11.2004 Pcie 0x10de 0x00c0-0x00cf - 05.41 NV41 4 - - 5 - 3 - - -
NV43 0x043 Curie 12.08.2004 Pcie 0x10de 0x0140-0x014f - 05.43 NV43 2 - - 3 - 2 - - -
NV44 0x044 Curie 15.12.2004 Pcie 0x10de 0x0160-0x016f - 05.44 NV44 ? - - 3 - 2 - - -
NV44A 0x04a Curie 04.04.2005 Pci 0x10de 0x0220-0x022f - 05.44 NV44 ? - - 3 - 2 - - -
G70 0x047 Curie 22.06.2005 Pcie 0x10de 0x0090-0x009f - 05.70 G70 4 - - 8 - 6 - - -
G72 0x046 Curie 18.01.2006 Pcie 0x10de 0x01d0-0x01df - 05.72 G72 ? - - 3 - 2 - - -
G71 0x049 Curie 09.03.2006 Pcie 0x10de 0x0290-0x029f - 05.71 G70 4 - - 8 - 6 - - -
G73 0x04b Curie 09.03.2006 Pcie 0x10de 0x0390-0x039f - 05.73 G73 2 - - 8 - 3 - - -
C51 0x04e Curie 20.10.2005 Igp 0x10de 0x0240-0x024f - 05.51 C51 ? - - 1 - 1 - - -
MCP61 0x04c Curie 06.2006 Igp 0x10de 0x03d0-0x03df - 05.61 MCP61 ? - - 1 - 1 - - -
MCP67 0x067 Curie 01.02.2006 Igp 0x10de 0x0530-0x053f - 05.67 MCP61 ? - - 1 - 1 - - -
MCP68 0x068 Curie 07.2007 Igp 0x10de 0x0530-0x053f - 05.67 MCP61 ? - - 1 - 1 - - -
MCP73 0x063 Curie 07.2007 Igp 0x10de 0x07e0-0x07ef - 05.73 MCP61 ? - - 1 - 1 - - -
RSX 0x04d Curie 11.11.2006 FlexIO - - - - G72 ? - - 8 - 6 - - -
G80 0x050 Tesla 08.11.2006 Pcie 0x10de 0x0190-0x019f - 60.80 G80 6 - - - - 8 2 - -
G84 0x084 Tesla 17.04.2007 Pcie 0x10de 0x0400-0x040f - 60.84 G80 2 - - - - 2 2 - -
G86 0x086 Tesla 17.04.2007 Pcie 0x10de 0x0420-0x042f - 60.86 G80 2 - - - - 1 2 - -
G92 0x092 Tesla 29.10.2007 Pcie 0x10de 0x0600-0x061f - 62.92 G80 4 - - - - 8 2 - -
G94 0x094 Tesla 29.07.2008 Pcie 0x10de 0x0620-0x063f - 62.94 G80 4 - - - - 4 2 - -
G96 0x096 Tesla 29.07.2008 Pcie 0x10de 0x0640-0x065f - 62.94 G80 2 - - - - 2 2 - -
G98 0x098 Tesla 04.12.2007 Pcie 0x10de 0x06e0-0x06ff - 62.98 G80 1 - - - - 1 1 - -
G200 0x0a0 Tesla 16.06.2008 Pcie 0x10de 0x05e0-0x05ff - 62.00 G80 8 - - - - 10 3 - -
MCP77 0x0aa Tesla 06.2008 Igp 0x10de 0x0840-0x085f - 62.77 MCP77 1 - - - - 1 1 - -
MCP79 0x0ac Tesla 06.2008 Igp 0x10de 0x0860-0x087f - 62.79 MCP77 1 - - - - 1 2 - -
GT215 0x0a3 Tesla 15.06.2009 Pcie 0x10de 0x0ca0-0x0cbf 0x0be4 70.15 G80 2 - - - - 4 3 - 1
GT216 0x0a5 Tesla 15.06.2009 Pcie 0x10de 0x0a20-0x0a3f 0x0be2 70.16 G80 2 - - - - 2 3 - 1
GT218 0x0a8 Tesla 15.06.2009 Pcie 0x10de 0x0a60-0x0a7f 0x0be3 70.18 G80 1 - - - - 1 2 - 1
MCP89 0x0af Tesla 01.04.2010 Igp 0x10de 0x08a0-0x08bf - 70.89 MCP77 2 - - - - 2 3 - 1
GF100 0x0c0 Fermi 26.03.2010 Pcie 0x10de 0x06c0-0x06df 0x0be5 70.00 GF100 6 1 2 - 4 4 1 - 2
GF104 0x0c4 Fermi 12.07.2010 Pcie 0x10de 0x0e20-0x0e3f 0x0beb 70.04 GF100 4 1 2 - 2 4 1 - 2
GF114 0x0ce Fermi 25.01.2011 Pcie 0x10de 0x1200-0x121f 0x0e0c 70.24 GF100 4 1 2 - 2 4 1 - 2
GF106 0x0c3 Fermi 03.09.2010 Pcie 0x10de 0x0dc0-0x0ddf 0x0be9 70.06 GF100 3 1 2 - 1 4 1 - 2
GF116 0x0cf Fermi 15.03.2011 Pcie 0x10de 0x1240-0x125f 0x0bee 70.26 GF100 3 1 2 - 1 4 1 - 2
GF108 0x0c1 Fermi 03.09.2010 Pcie 0x10de 0x0de0-0x0dff 0x0bea 70.08 GF100 1 2 2 - 1 2 1 - 2
GF110 0x0c8 Fermi 07.12.2010 Pcie 0x10de 0x1080-0x109f 0x0e09 70.10 GF100 6 1 2 - 4 4 1 - 2
GF119 0x0d9 Fermi 05.01.2011 Pcie 0x10de 0x1040-0x107f 0x0e08 75.19 GF100 1 1 1 - 1 2 1 - 1
GF117 0x0d7 Fermi 04.2012 Pcie 0x10de 0x1140-0x117f - ? GF100 1 1 2 - 1 2 1 1 1
GK104 0x0e4 Kepler 22.03.2012 Pcie 0x10de 0x1180-0x11bf 0x0e0a 80.04 GF100 4 1 4 - 4 2 1 1 3
GK107 0x0e7 Kepler 24.04.2012 Pcie 0x10de 0x0fc0-0x0fff 0x0e1b 80.07 GF100 2 1 4 - 1 2 1 1 3
GK106 0x0e6 Kepler 22.04.2012 Pcie 0x10de 0x11c0-0x11ff 0x0e0b 80.06 GF100 3 1 4 - 3 2 1 1 3
GK110 0x0f0 Kepler 21.02.2013 Pcie 0x10de 0x1000-0x103f 0x0e1a 80.10 GF100 6 1 4 - 5 3 1 2 3
GK110B 0x0f1 Kepler 07.11.2013 Pcie 0x10de 0x1000-0x103f 0x0e1a 80.80 GF100 6 1 4 - 5 3 1 2 3
GK210 ? Kepler ? Pcie 0x10de ? ? ? GF100 6 1 4 - 5 3 1 2 3
GK208 0x108 Kepler 19.02.2013 Pcie 0x10de 0x1280-0x12bf 0x0e0f 80.28 GF100 1 1 2 - 1 2 1 1 3
GK208B 0x106 Kepler ? Pcie 0x10de 0x1280-0x12bf 0x0e0f 80.28 GF100 1 1 2 - 1 2 1 1 3
GK20A 0x0ea Kepler ? Tegra - - - - GK20A 1 1 1 - 1 1 1 1 3
GM107 0x117 Maxwell 18.02.2014 Pcie 0x10de 0x1380-0x13bf 0x0fbc 82.07 GF100 2 1 4 - 1 5 1 2 3
GM108 0x118 Maxwell ? Pcie 0x10de 0x1340-0x137f ? 82.08 GF100 1 1 ? - 1 3 1 ? 3
GM204 0x124 Maxwell ? Pcie 0x10de 0x13c0-0x13ff 0x0fbb 84.04 GF100 ? 1 ? - ? ? 1 ? 3
GM200 0x120 Maxwell ? Pcie 0x10de 0x17c0-0x17ff 0x0fb0 84.00 GF100 ? 1 ? - ? ? 1 ? 3
GM206 0x126 Maxwell ? Pcie 0x10de 0x1400-0x143f 0x0fba 84.06 GF100 ? 1 ? - ? ? 1 ? 3
GM20B 0x12b Maxwell ? Tegra - - - - GK20A ? 1 ? - ? ? 1 ? 3
GP100 0x130 Pascal ? Pcie 0x10de 0x1580-0x15ff ? 86.00 GF100 ? 1 ? - ? ? 1 ? 3
GP102 0x132 Pascal ? Pcie 0x10de 0x1b00-0x1b7f 0x10ef 86.02 GF100 ? 1 ? - ? ? 1 ? 3
GP104 0x134 Pascal ? Pcie 0x10de 0x1b80-0x1bff 0x10f0 86.04 GF100 4 1 ? - 4 5 1 ? 3
GP106 0x136 Pascal ? Pcie 0x10de 0x1c00-0x1c7f 0x10f1 86.06 GF100 ? 1 ? - ? ? 1 ? 3
GP107 0x137 Pascal ? Pcie 0x10de 0x1c80-0x1cff 0x0fb9 86.07 GF100 ? 1 ? - ? ? 1 ? 3
GP108 0x138 Pascal ? Pcie 0x10de 0x1d00-0x1d7f 0x0fb8 86.08 GF100 ? 1 ? - ? ? 1 ? 3