v0.8.2025.53-preview

[JHRobotics]

• VESA driver
• fixed page fault in VBox VGA driver and QEMU std vga driver
• small stability improvements
• fixed corrupted driver *.inf in some cases

---

This discussion was created from the release v0.8.2025.53-preview.

[Nislaco]

Tested vanilla qemu-9.0.4 with q35 and windows95B and new build is working. Using SMD hack to enable sse3 binaries works also. Thank you!

[comchris]

wow, this magic is astonishing,

so now that VESA mode is unlocked,
I can ignore the red warning and apply it on the qemu/qxl device (spice in PVE) or virtio,
and I don’t have to use my start script on proxmox anymore to force "std + spice" to get audio streaming ;D nice!

3D Mark 2001SE b330 - default settings on Qemu9.2/PVE/intel N150 with

• vga std-wine/wine/wine-Win98SE SOFTGPUv0.8.53 score: 347
• virtio-wine/wine/wine-Win98SE SOFTGPUv0.8.53vesa score: 347
• VirGL-wine/wine/wine-Win98SE SOFTGPUv0.8.53vesa score: 349

I don't know if humanity really needs Virglrenderer support in Win98 on Qemu10,...
but it would make winamp2 visuals look smoother I guess... so ...

on virtualbox7.1.12/win11/Ryzen4GHz/RTX4070@~20% with 128MB-VMVGA-wine/wine/wine-Win98SE
SOFTGPUv0.6+SSE3 score: 4722
SOFTGPUv0.8.51 score: 5585
SOFTGPUv0.8.51+VMSVGA10=1 score: 4679
SOFTGPUv0.8.51+SSE3 simd95 score: 5778
SOFTGPUv0.8.51+globalgamma score: 4132
SOFTGPUv0.8.53+SSE3 simd95 score: 6174

for ref:
intel i5-6500T HD530 iGPU baremetal with Win11 score: 28289
same with Qemu/PVE iGPU via GVTg-mdev passthrough + Win11x64 score: 10181
same with Qemu/PVE iGPU via GVTg-mdev passthrough + Win10x32 score: 6981
same with Qemu9/PVE Virtio-gl "vGPU" + Proton exp on LinuxMint score: 4978
same with virtualbox/Cachy SOFTGPUv0.8.51 128MB-VMVGA-Win98SE score: 3701

Wyse VIA Eden C7 @ 1.2GHz S3 UniChrome™ Pro IGP Win98SE score: 2287
DELL Wyse DX0D silent 20$/20W thin client with HD6250 and WinXP score: 8833
ATI RADEON 9700pro AGP/Pentium4 and WinXP score: 11975
AMD HD8490 PCIe2.0-1x pass-through to WinXP VM on Qemu/PVE score: 25854

SOFTGPUv0.8.53vesa on HardGPU lacking win98 drivers:
AMD G-T44R (FujitsuS700) boots with strange/low colors, DDHelp executed illegal instruction in VMHAL9X.DLL, resolution change bluescreens - the VBE miniport vesa driver works* for this APU
AMD HD8490 PCIe2.0-1x pass-through to Win98SE VM on Qemu/PVE, (non-EFI mode/x-vga=1/rombar/SeaBios),
trayapp shows only 1MB VRAM and resolution change >1024 bluescreens

The question might be which VESA option to choose from in the drop down if using a VM with a real GPU.
I'll play around with it some more - might be a problem with my Displayport/HDMI/Monitor/EDID

Thanks!

[JHRobotics]

Thanks for benchmark!

I’m afraid, that combination of VESA driver + GPU in PCI passthrough is worst possible combination. My observation with PCI passthrough testing was that calling the BIOS of a given device usually ended by disaster – safe usually was initialization and attach protected mode driver as soon as possible. But my VESA driver calls video BIOS very often (about 1 time per every frame when application or game using double/triple buffering). Next limitation is relative slow access to device memory when is PCI passthrough active VESA driver is fastest when can very quickly copy between CPU and GPU memory (and needs both direction, GPU to CPU is usually super slow). Another problem can be in ATI/AMD card simply because I didn’t test them too much, I have got for testing a relatively varied collection of NVIDIA card and lots of motherboards/CPU with integrated Intel iGPU but only few AMD discrete graphics cards (I tested only for curiosity PCI ATI Rage XL with 8 MB VRAM). But when you can snap BSOD or rewrite from screen the failure address I can maybe trace the error.

Both install options in drop down menu are essentially the same, only if you select real HW, greater detail reduction is turned on (I assumed that the main target would be processors around socket 775 and integrated graphics in them).

Also, I think that humanity definitely needs VirGL support for Windows 9x, because I think, that VMWare SVGA3D API too limiting for performance in 3D games. I've been avoiding it for a long time, but this will probably be the next goal.

But back to PCI passthrough – maybe it's worth a try to disable MTRR and VRAM test, you must boot to safe mode and insert to this to registry (or insert it just after driver installation before reboot):

REGEDIT4

[HKEY_LOCAL_MACHINE\Software\vmdisp9x\vesa]
"MTRR"="0"
"NoMemTest"="1"

(Docs: https://github.com/JHRobotics/softgpu/blob/main/config.md#vesa-driver)

I don't know if this will help, but I'd be happy to know the results :-)

[JHRobotics]

I found X1650 and HD 7750, first on was defective (probably bad RAM) but second one looks OK and I have same result as you. BSOD can be solved by skipping VRAM test (NoMemTest=1), MTRR can stay on. Problem is bad colours in 16bpp (32bpp was OK), it looks like card reports multiple 16bpp mode, but driver pick first one which is 5:5:5 (15bpp) in real. Also, card report only 32 MB of memory (this is still OK, it’s enough for frame buffer and texture memory is emulated in RAM. Whole system was pretty unstable but for it may be various explanations (for example that GPU isn’t in 100% condition - I pulled it out of some computer wreck).

[comchris]

I wanted to do some more testing to widen the scope and get some more reliable results,
as It's hard to exclude all other factors that prevent win98 to boot nowadays...

On the APU front of the (2010?) G-T44R: 1600x1200x32 works with "NoMemTest"="1", but the error remains

DDHELP executed an invalid instruction in module VMHAL9X.DLL at 017f:b00be1a1.
Registers:
EAX=8345a000 CS=017f EIP=b00be1a1 EFLGS=00010286
EBX=83431018 SS=0187 ESP=0056f6e0 EBP=0056f7e8
ECX=0056f7f8 DS=0187 ESI=bff77716 FS=246f
EDX=b00b15c0 ES=0187 EDI=b00b0000 GS=0000
Bytes at CS:EIP: 66 0f 3a 22 c2 01 ba 10 d5 0b b0 66 0f 3a 22 ca 
Stack dump: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 bfc08410 000001d7 0056f7ac 00000001 8250ee94 000002e3 b0119546 82525578 

If I disable tray3d.exe from autostart the error will trigger later while starting 3D mark for example.

So I choose to test on a real CPU - Win98SE on Core i7-920 PCIe baremetal with SoftGPU 0.8.53 vesa:
using VGA-SUB-D or DVI to VGA-SUB-D with analog signal

2007 ATI HD 2400 (RV610) "NoMemTest"="1": 1600x1200x32 works 3D2001score: 187 3D2000score: 361
guess I'm CPU bound ... so I overclocked from 2.66 to 3.15Ghz ... score: 187 -> 217 ... linearity checks out

2008 ATI HD 4850 (RV770) 1600x1200x32 works with "NoMemTest"="1"
2008 ATI HD 4350 (RV710) bluescreens on boot without "NoMemTest"="1", 1600x1200x32 works with "NoMemTest"="1"

2006 ATI X1900 XT (R580) 1600x1200x32 works and total vram is displayed in tray app
2004 ATI X300 SE (RV370) limited to 1280x1024x32

On the Pentium 4 with AGP I tested the
2002 ATI Radeon 9700 PRO (R300) which actually has catalyst drivers available,
It's doing a blinking white on green Ascii Art for a few seconds, trayapp won't start, but no error. modes detected again limited to 1280x1024x32. After another restart the trayapp works and shows 128MB total. Strange.
But Quake or Turok: Mesa3D.dll pagefault.

btw - The older 3 cards show up as two devices for dual head - I forced the VESA driver on both -
this works for the R300 & RV370 but the R580 didn't like it. But have to test that again.
Maybe the 2nd head needs to be deactivated ?

So I jumped ship and went for the glorious Nvidia FX-5200 in the AGP slot to exclude other factors.
The blinking ascii art after Bios and prior to desktop remains, but after that 1600x1200x16! works and Q3A starts!
But 3DMark or GLchecker pagefault...

GLCHECKER ... in module MESA3D.DLL bei 018f:698a0bfc.
Register:
EAX=00000000 CS=018f EIP=698a0bfc EFLGS=00210246
EBX=00000002 SS=0197 ESP=0077f688 EBP=00000000
ECX=00000000 DS=0197 ESI=00010809 FS=382f
EDX=00000002 ES=0197 EDI=00000000 GS=0000
Bytes bei CS:EIP: 0f ae 44 24 20 31 c0 f6 44 24 3c 40 0f 95 c0 89 
Stack dump: 00000000 000007fc 00000000 8250800c 82508000 756e6547 6c65746e 49656e69 00000000 00001000 00400000 7fffffff 00000001 00000001 0000024a 00000000 

I pulled the Nvidia card as the board got an igfx 845G ftw.
The VESA driver was chosen by windows. After the blinking ascii art it works but limited to 1280x1024x32.
Same as above, WGL Test works and shows llvmpipe/GL4.5... but GLCHECKER pagefaults, same as Q3A.

The P4 might be low on electrolytes and has Win98 drivers available anyway -
so back to worst possible combination with

AMD HD8490 PCIe2.0-1x passthrough to Win98SE VM on Qemu9/PVE on i5-4670 CPU @ 3.40GHz
using Displayport TO HDMI TO DVI
"NoMemTest"="1": 1600x1200x32 works 3D2001score: 283

I think it's a great result that all ATI PCIe cards work! But seems the "HDs..." need "NoMemTest"="1"
Let me know if I can test anything specific.

[DSN9990]

Hi @JHRobotics when will you put out a new version? Why your development is so slow?

[DSN9990]

Hi again @JHRobotics when you put out a new version??

[Nislaco]

Keep in mind SoftGPU is several projects working together across several system emulators. I can only imagine how much time and effort it took to get the project where it is today. If you have any issues please report them. However just asking for another version is most likely not going to help accomplish anything.

…

On Sat, Sep 20, 2025 at 10:36 PM DSN ***@***.***> wrote: Hi again @JHRobotics <https://github.com/JHRobotics> when you put out a new version?? — Reply to this email directly, view it on GitHub <#149 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJXDSW5DMB23MG7QBBGNJY33TYFJ5AVCNFSM6AAAAACCVB6ZXWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTINBWGY3DOMQ> . You are receiving this because you commented.Message ID: ***@***.***>

[steviator]

I have looked at the JH Robotics website, and noticed that they don't seem to have any products that make use of Windows 9x emulation for sale.

I'm going out on a limb when I say this, but I don't think that Windows 9x emulation enhancements are a very profitable undertaking for JH Robotics Inc., and so that could have something to do with what some people seem to find an unsatisfactory pace of development.

Perhaps if some individuals are willing to pay a fair price for their copy of SoftGPU, by which I mean the full cost in hours spent, charged at JH Robotics usual rate for custom development, divided by the number of copies sold, those people could then feel justified in complaining about the pace of development.

I humbly suggest though, that such cost might be somewhat unpalatable, and I for one, would rather not see development rushed, particularly in light of the fact that the development of softgpu is probably a leisure activity undertaken for its own sake rather than an attempt at pleasing demanding customers, the latter of which is an itch that can probably be scratched during business hours.

Just sayin'.

[hughobrien]

I have previously requested that @DSN9990 stop this behaviour. But to no avail.
1
2
3

This is bordering on harassment so I have reported it as such.

[comchris]

Well. According to public knowledge from science&&fiction,
old real mode computers will be an important tool in humanity’s war against the AGI.
So please make sure to bring a copy into the bunker –
otherwise we might be stuck on 16 colors and without glide support for the next decades!!1 ;o)