[Irinikus]

One of the Raster managers in my Onyx 2 is defective, so I decided to pull it from the machine and check all of the ribbon connectors for corrosion, in the hope of finding the problem, sadly after reinserting it , the problem persisted!

The problem is intermittent, and when both RM's work, this is the performance I'm getting:



When the problem presents itself, this is what I see:



This is the performance that I get with only one RM-10 fitted:



So you definitely suffer a performance loss, when you only have one RM fitted, which surprised me, as Quake 3 definitely doesn't require 256MB of texture memory! The memory throughput is more than likely increased, with the addition of the second RM. (This I was unaware of!)

This is what it consists of:

[jan-jaap]

Did you run 'irsaudit'? It will tell you more. I usually consider a part passed if it survives 'irsaudit REPEAT=5', exactly because some problems are intermittent.

If the problem is with the raster memory you may be able to fix the RM using an RM9 donor. If it's texture related you're out of luck. I suspect the latter. In that case maybe you're interested in a pair of RM11's? C'mon, you know you want them

[spiroyster]

Agree with jj, the Saturday night fever tribute screenshot you posted is a classic example of corrupted texture data. InfiniteReality explicitly differentiates between raster and texture memory (i.e the framebuffer is not mapped to the same memory as texture memory, which are mappepd to their own TRAM's), so what you are seeing is corrupted portions of the texture TRAM... however... are those screen grabs or photos? They look like they have vertical scanlines (which afaik they shouldn't) which may indicate some bad raster memory too, or may simply be a scaling artifact (perfectly normal). I can't work it out, may need some glasses :(.

Unfortunately this is out of your control as afaik you cannot dictate which memory areas to use. So if you are lucky, q3 texture data will be mapped to the 'good' TRAM, if you are unlucky it would map to some dodgy memory area, corrupting it, and then causing issues when read (i.e what you see in the sky).

---

So you definitely suffer a performance loss, when you only have one RM fitted, which surprised me, as Quake 3 definitely doesn't require 256MB of texture memory! The memory throughput is more than likely increased, with the addition of the second RM. (This I was unaware of!)

RM, while it does hold the texture memory, is responsible for rasterisation. Having the RM hold the texture and raster memory means the bottleneck between getting the texture data to the rasterizer is minimized, but the final frame drawn is still limited by the speed of the rasterizer. So two are going to be better than one... even if the texture memory reads were instantaneous.

Think of it like painting a wall, we have someone mixing up the paint (analagous to texture prep and throughput) and someone else painting the wall (rasterizer)... the entire wall space to paint being the framebuffer.

If the texture throughput was slow with a single RM (i.e takes ages to mix the paint ready for use), the painter is waiting while the 'paint-mixer' mixes the paint.

If the texture throughput was optimal with a single RM (i.e 'paint-mixer' is really fast and there is always paint ready for use by the painter), the 'paint-mixer' is waiting until the painter has finished painting to mix up some more.

With optimal texture throughput (i.e fast 'paint-mixer'), and two painters (2 x RM), the whole wall is going to be painted much faster... ergo higher FPS!!!!

[Irinikus]

Agree with jj, the Saturday night fever tribute screenshot you posted is a classic example of corrupted texture data. InfiniteReality explicitly differentiates between raster and texture memory (i.e the framebuffer is not mapped to the same memory as texture memory, which are mappepd to their own TRAM's), so what you are seeing is corrupted portions of the texture TRAM... however... are those screen grabs or photos? They look like they have vertical scanlines (which afaik they shouldn't) which may indicate some bad raster memory too, or may simply be a scaling artifact (perfectly normal). I can't work it out, may need some glasses :(.

Unfortunately this is out of your control as afaik you cannot dictate which memory areas to use. So if you are lucky, q3 texture data will be mapped to the 'good' TRAM, if you are unlucky it would map to some dodgy memory area, corrupting it, and then causing issues when read (i.e what you see in the sky).

---

So you definitely suffer a performance loss, when you only have one RM fitted, which surprised me, as Quake 3 definitely doesn't require 256MB of texture memory! The memory throughput is more than likely increased, with the addition of the second RM. (This I was unaware of!)

RM, while it does hold the texture memory, is responsible for rasterisation. Having the RM hold the texture and raster memory means the bottleneck between getting the texture data to the rasterizer is minimized, but the final frame drawn is still limited by the speed of the rasterizer. So two are going to be better than one... even if the texture memory reads were instantaneous.

Think of it like painting a wall, we have someone mixing up the paint (analagous to texture prep and throughput) and someone else painting the wall (rasterizer)... the entire wall space to paint being the framebuffer.

If the texture throughput was slow with a single RM (i.e takes ages to mix the paint ready for use), the painter is waiting while the 'paint-mixer' mixes the paint.

If the texture throughput was optimal with a single RM (i.e 'paint-mixer' is really fast and there is always paint ready for use by the painter), the 'paint-mixer' is waiting until the painter has finished painting to mix up some more.

With optimal texture throughput (i.e fast 'paint-mixer'), and two painters (2 x RM), the whole wall is going to be painted much faster... ergo higher FPS!!!!

Thanks very much for your input!

I've removed the back plate from the RM and have inserted it into the machine to act as a blank until I acquire a new RM! :(

[Jammaster]

Yep, RM9 is very close to RM10.
It will be great to test an hybrid rm9 / 10 boardset to find the problem.
And ... I don't know if a person here have an Onyx 2 with IR4/RM11 working/benching ?!

[Irinikus]

I will have to source another RM10 at some point, to complete the set. (As my Onyx2 is limping along on one RM10 at the moment!)

I have a spare DG5-2 and this broken RM10, all I need to do is source a GE16-4 and then I will dismantle them and frame an IR3 board set. (It will look really cool once properly framed on a black background!)

[Jammaster]

One of the two boards of an RM9 stack is the same of RM10 I think.

[Irinikus]

I am aware of this, but just to be sure, I will order a complete RM10 in the future.