This video is pure gold: thanks so much for uploading I've learnt so much from it. I may have to watch it several times though!!! A great overview and introduction to so many areas for further study.
This is so good. As beginner Cuda programmer I get the most important concepts and terminology seeds right into good soil. I know what to except. First thing to do is fully accept parallel threads being the king.
One thing that's confusing: if reading from a memory location in a different row is 3x slower than reading from a memory location in the same row - how come we get 13x slowdown? Worst case (if you're deliberately reading from a different row each time) - one would expect a 3x slowdown? What am I missing out on? Is it the burst mode? 2) You're using float2 type so that means your thread is loading 4 bytes (for 2 points) not 8 bytes? Which would put the 4 warps into 512B loading territory instead of the optimal 1024? -> EDIT: ok, I just saw that p1 & p2 are actually float pointers so that does make sense. 3) How can we guarantee that p1 & p2 arrays (holding the points) are adjacent, i.e. in the same physical row in memory? Great video! The sound quality is a bit off though.
It's 3x slower for reading a single value, but it gets worse when reading many contiguous values where the burst column read can read many values in one operation. For example, let's say that we're reading two sets of 10 values, one set of which are all contiguous in a row, and one set that are all on different rows. And you have the three ops in the video: LOAD a row, READ a column, STORE the row back. For the contiguous values: time = LOAD + BURST READ + STORE = 3 ops For the disjoint values: time = (LOAD + READ + STORE)*10 = 30 ops That's how you get the 10x speed-up.
Excellent. For the matrix multiply, you’re reusing the same row multiple times but the columns would have to be loaded in every time. So how do you increase compute intensity of the columns?
"Occupancy is the most powerful tool that you have for tuning a program. **Once you're doing your best for memory access patterns** there's pretty much no algorithmic optimization that you can do that'll speed your program up by as much as 33%" I thought FlashAttention's major contribution was optimizing memory access patterns, namely reducing the number of HBM loads/stores.
He literally said "Once you are doing your best for memory access paterns" and Flash Attention is a MEMORY ACCESS algorithm, it reduces the memory access to GPU HMB RAM.
They are getting better in terms of energy efficiency and performance www.cnbc.com/2024/04/09/intel-unveils-gaudi-3-ai-chip-as-nvidia-competition-heats-up-.html
@@codinginrust Limitation of Gaudi is that it is a less flexible fixed function matrix math accelerator. General purpose compute engine in Hopper/Blackwell architecture can better support rapidly evolving LLM algos. Another issue is interconnect bandwidth: NVLINK5 absolutely crushes PCIE5
This video is pure gold: thanks so much for uploading I've learnt so much from it. I may have to watch it several times though!!! A great overview and introduction to so many areas for further study.
This is so good. As beginner Cuda programmer I get the most important concepts and terminology seeds right into good soil. I know what to except. First thing to do is fully accept parallel threads being the king.
this video is so clear and great
One thing that's confusing: if reading from a memory location in a different row is 3x slower than reading from a memory location in the same row - how come we get 13x slowdown? Worst case (if you're deliberately reading from a different row each time) - one would expect a 3x slowdown?
What am I missing out on? Is it the burst mode?
2) You're using float2 type so that means your thread is loading 4 bytes (for 2 points) not 8 bytes? Which would put the 4 warps into 512B loading territory instead of the optimal 1024? -> EDIT: ok, I just saw that p1 & p2 are actually float pointers so that does make sense.
3) How can we guarantee that p1 & p2 arrays (holding the points) are adjacent, i.e. in the same physical row in memory?
Great video! The sound quality is a bit off though.
It's 3x slower for reading a single value, but it gets worse when reading many contiguous values where the burst column read can read many values in one operation.
For example, let's say that we're reading two sets of 10 values, one set of which are all contiguous in a row, and one set that are all on different rows. And you have the three ops in the video: LOAD a row, READ a column, STORE the row back.
For the contiguous values: time = LOAD + BURST READ + STORE = 3 ops
For the disjoint values: time = (LOAD + READ + STORE)*10 = 30 ops
That's how you get the 10x speed-up.
Thank you!
Excellent. For the matrix multiply, you’re reusing the same row multiple times but the columns would have to be loaded in every time. So how do you increase compute intensity of the columns?
Christopher, do you think the long time it takes for ram to be accessed could be decreased by embedding a basic cpu in those ram modules?
Good question, I don't know!
33:10 FlashAttention proved this wrong
"Occupancy is the most powerful tool that you have for tuning a program. **Once you're doing your best for memory access patterns** there's pretty much no algorithmic optimization that you can do that'll speed your program up by as much as 33%"
I thought FlashAttention's major contribution was optimizing memory access patterns, namely reducing the number of HBM loads/stores.
can you please explain this a bit more? I'm trying to teach myself flash attention's cuda code.
He literally said "Once you are doing your best for memory access paterns" and Flash Attention is a MEMORY ACCESS algorithm, it reduces the memory access to GPU HMB RAM.
Is there a chance you can do a video about Why AMD's version isnt as good as NVIDIA ?
I've not got a AMD gfx card ZLUDA means it does not really matter www.phoronix.com/review/radeon-cuda-zluda
AMD's issue is tooling and the general software ecosystem. The hardware is reasonably close.
Look like Intel is out of the question here.
They are getting better in terms of energy efficiency and performance www.cnbc.com/2024/04/09/intel-unveils-gaudi-3-ai-chip-as-nvidia-competition-heats-up-.html
@@codinginrust Limitation of Gaudi is that it is a less flexible fixed function matrix math accelerator. General purpose compute engine in Hopper/Blackwell architecture can better support rapidly evolving LLM algos. Another issue is interconnect bandwidth: NVLINK5 absolutely crushes PCIE5