Interesting question about the L1 texture cache size of G200 used in David's review

[littlehead]

David mentioned the L1 texture cache size is 8K for one TCP. so the total L1 on the chip should be 8K * 8 on G80 or 8K * 10 on G200.
Ok. If the L1 is private, usually be true on CPU side. So the max L1 size which could be accessed by single pixel shading thread is 8K.
So I use the pointer chasing approach to verify this. the data looks quite interesting. and the point has big L1 miss latency is appeared on 20Kb.
That thread will loop on an given working set and will create significant latency if the workset larger than the cache.
The data:
workset latency
8192Byte 0.174301
12288Byte 0.180389
16384Byte 0.187164
20480Byte 0.289429
24576Byte 58.757538
28672Byte 58.551575
32768Byte 59.707016
36864Byte 59.930389


here's the shader code,
ps.3.0;
dcl_texcoord0 v0.xy;
dcl_2d s0;
dcl_2d s1;
dcl_2d s2;
dcl_2d s3;
mov r1.xy, v0.xy;
rep i0;
rep i1;
rep i2;
texld r1.rgba , r1, s0;
texld r1.rgba , r1, s1;
texld r1.rgba , r1, s2;
texld r1.rgba , r1, s3;
endrep;
endrep;
endrep;
nrm r2, r1;
mov oC0.rgba ,r2 ;

 

[Farhan]

I think dkanter says it's 8KB per SM, or 24KB per TCP, which is correct as i understand it.

 

[littlehead]

It's TPC

Based on the the diagram draw by David,
The 8K L1 texture cache is exactly below the SM controller, the SM controller is accept request from SM[0] SM[1] and SM[2]. So the 8K is shared by 3 SM and won't be 24K.
I raise the question because 8K for each cluster is reasonable. But the data show different indicatation. So this is why I'm confusing.
Or David draw the diagram wrong?

 

[Farhan]

Have you looked at the updated diagram?

 

[dkanter]

Have you looked at the updated diagram?

To clarify - my first diagram was wrong and had each L1 Texture cache as 8KB shared by 2 or 3 SMs.

That has now been corrected. Perhaps I should keep a change log in my articles, but I don't (yet).

The L1 tex caches are located in the tex pipeline in each TPC. However, 8KB is allocated to each SM (statically). So G80 has 16KB, GT200 has 24KB.

David

 

[littlehead]

Why 8K per SM

Why 8K per SM, cut them on low end GPU of the same series?
Based on my measurement, it seems G80 and G200 both use 20K per TPC. There's no L2 cache on low end card like 8500GT.
Fix me if i'm wrong.

 

[littlehead]

And L2 cache for G80 I measured is 192KB.