Nvidia: Difference between revisions
Added some sources and corrected some imprecision |
In 2014, Nvidia added an on-die security processor called Falcon to the GTX 900 series that enforced vBIOS signature checks, killing a longstanding practice of cross-flashing GeForce cards with Quadro firmware. |
||
| Line 23: | Line 23: | ||
===Artificial FP64 Restriction on Kepler Consumer GPUs (''2013 - 2014'')=== | ===Artificial FP64 Restriction on Kepler Consumer GPUs (''2013 - 2014'')=== | ||
Nvidia's Kepler-generation GK110 GPU contains 64 double precision (FP64) CUDA cores per SMX block on all variants of the chip, including consumer GeForce cards, the GTX Titan, and the enterprise Tesla K40.<ref>{{Cite web |last=Angelini |first=Chris |date=2013-02-19 |title=GK110: The True Tank |url=https://www.tomshardware.com/reviews/geforce-gtx-titan-gk110-review,3438-2.html |url-status=live |access-date=2026-03-02 |website=Tom's Hardware}}</ref> On the GTX 780 and GTX 780 Ti, Nvidia forced the FP64 units to run at one eighth of the GPU's clock rate, yielding an effective FP64 throughput of 1/24 that of FP32, despite the hardware being fully present.<ref>{{Cite web |last=Angelini |first=Chris |date=2013-02-19 |title=Compute Performance And Striking A Balance |url=https://www.tomshardware.com/reviews/geforce-gtx-titan-gk110-review,3438-3.html |url-status=live |access-date=2026-03-02 |website=Tom's Hardware}}</ref> The GeForce GTX Titan, also based on GK110, shipped with a toggle in the Nvidia control panel under "CUDA - Double precision" that restored the full 1/3 FP64 rate at the cost of disabling GPU Boost, confirming the limitation on other GK110 consumer cards was a deliberate restriction rather than a hardware incapability.<ref>{{Cite web |last=Angelini |first=Chris |date=2017-05-25 |title=Nvidia Titan Xp 12GB Review |url=https://www.tomshardware.com/reviews/nvidia-titan-xp,5066-14.html |url-status=live |access-date=2026-03-02 |website=Tom's Hardware}}</ref> The Tesla K40, built on the same GK110 die, delivered the full 1/3 FP64 rate and retailed at several times the price of consumer equivalents. Beginning with the Maxwell architecture in 2014, Nvidia dropped native FP64 throughput to 1/32 of FP32 by physically reducing the number of FP64 cores on consumer dies, ending the firmware restriction but continuing the segmentation through hardware design.<ref>{{Cite web |last=Smith |first=Ryan |date=2015-03-17 |title=GM200: All Graphics, Hold The Double Precision |url=https://www.anandtech.com/show/9059/the-nvidia-geforce-gtx-titan-x-review/2 |url-status=live |access-date=2026-03-02 |website=AnandTech}}</ref> | Nvidia's Kepler-generation GK110 GPU contains 64 double precision (FP64) CUDA cores per SMX block on all variants of the chip, including consumer GeForce cards, the GTX Titan, and the enterprise Tesla K40.<ref>{{Cite web |last=Angelini |first=Chris |date=2013-02-19 |title=GK110: The True Tank |url=https://www.tomshardware.com/reviews/geforce-gtx-titan-gk110-review,3438-2.html |url-status=live |access-date=2026-03-02 |website=Tom's Hardware}}</ref> On the GTX 780 and GTX 780 Ti, Nvidia forced the FP64 units to run at one eighth of the GPU's clock rate, yielding an effective FP64 throughput of 1/24 that of FP32, despite the hardware being fully present.<ref>{{Cite web |last=Angelini |first=Chris |date=2013-02-19 |title=Compute Performance And Striking A Balance |url=https://www.tomshardware.com/reviews/geforce-gtx-titan-gk110-review,3438-3.html |url-status=live |access-date=2026-03-02 |website=Tom's Hardware}}</ref> The GeForce GTX Titan, also based on GK110, shipped with a toggle in the Nvidia control panel under "CUDA - Double precision" that restored the full 1/3 FP64 rate at the cost of disabling GPU Boost, confirming the limitation on other GK110 consumer cards was a deliberate restriction rather than a hardware incapability.<ref>{{Cite web |last=Angelini |first=Chris |date=2017-05-25 |title=Nvidia Titan Xp 12GB Review |url=https://www.tomshardware.com/reviews/nvidia-titan-xp,5066-14.html |url-status=live |access-date=2026-03-02 |website=Tom's Hardware}}</ref> The Tesla K40, built on the same GK110 die, delivered the full 1/3 FP64 rate and retailed at several times the price of consumer equivalents. Beginning with the Maxwell architecture in 2014, Nvidia dropped native FP64 throughput to 1/32 of FP32 by physically reducing the number of FP64 cores on consumer dies, ending the firmware restriction but continuing the segmentation through hardware design.<ref>{{Cite web |last=Smith |first=Ryan |date=2015-03-17 |title=GM200: All Graphics, Hold The Double Precision |url=https://www.anandtech.com/show/9059/the-nvidia-geforce-gtx-titan-x-review/2 |url-status=live |access-date=2026-03-02 |website=AnandTech}}</ref> | ||
===vBIOS Signature Lock via Falcon Security Processor (''2014 - Present'')=== | |||
GeForce and Quadro cards frequently share identical silicon, with enterprise features differentiated by firmware and driver configuration. The GTX 680 and Quadro K5000 both used the GK104 die,<ref>{{Cite web |last=Smith |first=Ryan |date=2012-08-07 |title=NVIDIA Announces Kepler-Based Quadro K5000 & Second-Generation Maximus |url=https://www.anandtech.com/show/6140/nvidia-announces-keplerbased-quadro-k5000-secondgeneration-maximus |url-status=live |access-date=2026-03-02 |website=AnandTech}}</ref> and users could cross-flash a Quadro vBIOS onto a GeForce card or modify resistor straps on the PCB to change the PCI device ID, unlocking Quadro driver support and associated features such as 10-bit colour output, genlock, and professional OpenGL optimisations.<ref>{{Cite web |date=2013-03 |title=Hacking NVidia Cards into their Professional Counterparts |url=https://www.eevblog.com/forum/general-computing/hacking-nvidia-cards-into-their-professional-counterparts/ |url-status=live |access-date=2026-03-02 |website=EEVblog Forum}}</ref> With the GeForce 900 series Maxwell launch in 2014, Nvidia introduced an on-die security processor codenamed Falcon across all its GPUs, which enforced vBIOS signature verification and prevented cards from booting with unauthorised firmware, ending the cross-flashing practice.<ref>{{Cite web |last=Sherr |first=Mark |date=2023-08-21 |title=Nvidia vBIOS Modding Is Back After Signature Lock Broken |url=https://www.tomshardware.com/news/new-nvidia-bios-modding-tools |url-status=live |access-date=2026-03-02 |website=Tom's Hardware}}</ref> In August 2023, modders discovered that Nvidia's own NVFlash utility contained a built-in mismatch bypass and released two tools, OMGVflash and NVflashk, which re-enabled cross-flashing on cards up to the RTX 40 series.<ref>{{Cite web |date=2023-08-22 |title=NVIDIA BIOS Signature Lock breakthrough: New tools enable vBIOS modding and crossflash |url=https://www.igorslab.de/en/durchbruch-bei-nvidia-bios-signature-lock-neue-tools-ermoeglichen-vbios-modding-und-crossflash/ |url-status=live |access-date=2026-03-02 |website=igor'sLAB}}</ref> | |||
===Stagnation of Consumer GPU Offerings=== | ===Stagnation of Consumer GPU Offerings=== | ||