Terascale 2 wow

I'm actually vibecoding vx855 to w98 right now, will be done in 3-5 days... perhaps I can send you my source...

mwdmeyer wrote on 2026-10-27, 01:52:

myne wrote on 2026-10-26, 08:31:

Well done.
The test pig is all yours until further notice.
I expect 20k 3dmarks by christmas :p

First run on your 5450

Nice work, perhaps my Radeon 5770 can reach 20k, as it has ten times the number of cores. 😁 Will assist you in ironing out the issues on the 5450 first of course.

myne wrote on 2026-10-27, 03:37:

5450?

Which one of us is confused?

I just snuck my own little test pig into the farm. 😆

Watch it.
I'll get my other little piggy to beat your little piggy!
R7E > teraphael2!

https://github.com/michaeldale/velocity9x/rel … ses/tag/v0.32.0 includes Radeon 5450 support and the start of HW T&L.

Any chance to include more older VLB stuff? Trident 9xxx, Cirrus Logic 54xx, S3 Vision 9xx, ATI mach32,...

AncapDude wrote on 2026-10-27, 11:33:

Any chance to include more older VLB stuff? Trident 9xxx, Cirrus Logic 54xx, S3 Vision 9xx, ATI mach32,...

Yeah 100%, if you can do remote agent. Otherwise I have TGUI9440 and ET4000/W32i.

The S3 Trio 64 VLB is already supported, although I haven't tested in recent versions so probably no improvements.

mwdmeyer wrote on 2026-10-27, 11:30:

https://github.com/michaeldale/velocity9x/rel … ses/tag/v0.32.0 includes Radeon 5450 support and the start of HW T&L.

How does T&L work roughly speaking, and what could hardware add?
Is it just basically a bunch of floating point additions? multiplications?

myne wrote on 2026-10-27, 11:40:

mwdmeyer wrote on 2026-10-27, 11:30:

https://github.com/michaeldale/velocity9x/rel … ses/tag/v0.32.0 includes Radeon 5450 support and the start of HW T&L.

How does T&L work roughly speaking, and what could hardware add?
Is it just basically a bunch of floating point additions? multiplications?

Basics.. https://en.wikipedia.org/wiki/Transform,_clip … g,_and_lighting

Edit: gah, that doesn't really get into the meat, sorry.

I unplugged testpig1 (iCore G1 (ironlake ))
Replaced it with testpig2 (R7e (GCN 3))

IF you want #1 back, let me know.

Oh, and I figured out what T&L is.
Basically wrap texture around object, stretch/shrink, calculate viewing angle, add lighting.
Fairly straightforward but extremely repetitive math.

myne wrote on Yesterday, 03:19:

I unplugged testpig1 (iCore G1 (ironlake )) Replaced it with testpig2 (R7e (GCN 3)) […]
Show full quote

I unplugged testpig1 (iCore G1 (ironlake ))
Replaced it with testpig2 (R7e (GCN 3))

IF you want #1 back, let me know.

Oh, and I figured out what T&L is.
Basically wrap texture around object, stretch/shrink, calculate viewing angle, add lighting.
Fairly straightforward but extremely repetitive math.

Yeah it was a bit complex in the driver, and not enabled for everything yet, needs more work.

testpig2 is getting some love now!

Ah good. It's working.
Headless is always a bit of a guess. Ping replies don't necessarily mean 98 is working (ndis2).

You were mostly 'done' as it were with the ironlake, right? At least until T&L is ready for alpha testing

myne wrote on Yesterday, 06:36:

Ah good. It's working.
Headless is always a bit of a guess. Ping replies don't necessarily mean 98 is working (ndis2).

You were mostly 'done' as it were with the ironlake, right? At least until T&L is ready for alpha testing

Yeah it's pretty good for now, I'm sure I'll need it again when I add new features, but lets get R7e going!

I've got a nice speed bump for HD 4600 in next release too.

I'm curious to see. It's a shit tier IGP, but it's relatively modern and never even had xp drivers.
Even without 7000 CUs the architecture should be leagues ahead of anything that will run on 9x.

Are you noticing any interesting things with scaling?
eg a low end of the same architecture is performing similarly to a high? Or is it scaling decently enough?

myne wrote on Yesterday, 07:32:

I'm curious to see. It's a shit tier IGP, but it's relatively modern and never even had xp drivers. Even without 7000 CUs the a […]
Show full quote

I'm curious to see. It's a shit tier IGP, but it's relatively modern and never even had xp drivers.
Even without 7000 CUs the architecture should be leagues ahead of anything that will run on 9x.

Are you noticing any interesting things with scaling?
eg a low end of the same architecture is performing similarly to a high? Or is it scaling decently enough?

Newer chips are more complex, the ironlake was powerful enough but a more simple design, which is why it is faster than the HD 4600 currently.

The drivers need to advance significantly to take advantage of these newer GPUs.

0.33 is now out. AMD Radeon R7E now supported and Gen HD 4600 speed improvements (3DMark 2000 went from 2344 to 3719)

Nice I was just reading the decision notes.
Do you write them or do the agents?
How do they extrapolate solutions? Any idea?

You've got a fair chunk of architectures now.
You might get more test monkeys if you add all the untested cards with tested architectures into a different inf.
Name the driver slightly different in device manager too. Just so you don't get confused with bug reports.

Ie, theoretically all of gcn3 should work now.
It would be ideal for your remote client to recognise that the architecture is tested, immediately get the testpig to a working state and run tests/reports for you to review.

AI does most of the work. Although took a bit to get it doing what I needed.

Some good ideas there.

For now I’m going to go back to improve Direct3D support, aim to get nature test in 3dmark2001se working.

Speed is still not great so need to improve that.

Fair enough.
Just going to note that Microsoft in many cases thunk things up.
They'll leave a tiny stub of an old function that manipulates the input slightly, and feeds it into the newer version.

I noticed this when I was looking at mech2 in a debugger.

Which is the long way of saying that you might find it saves time to target dx9c.
Might.
I don't know. You'd have to look at it.