"Alan: Sure, yep, so one of the things that we felt like on MI350 in this timeframe, that it's going into the market and the current state of AI... we felt like that FP6 is a format that has potential to not only be used for inferencing, but potentially for training. And so we wanted to make sure that the capabilities for FP6 were class-leading relative to... what others maybe would have been implementing, or have implemented. And so, as you know, it's a long lead time to design hardware, so we were thinking about this years ago and wanted to make sure that MI350 had leadership in FP6 performance. So we made a decision to implement the FP6 data path at the same throughput as the FP4 data path. Of course, we had to take on a little bit more hardware in order to do that. FP6 has a few more bits, obviously, that's why it's called FP6. But we were able to do that within the area of constraints that we had in the matrix engine, and do that in a very power- and area-efficient way.
the main question is going to be software stack. NVIDIA is already shipping NVFP4 kernels and perf is looking good. It took a really long time after MI300X's that the FP8 kernels were OK (not even good, compared to almost perfect FP8 support in NVIDIA side of things).
I will doubt that they will be able to reach %60-70 of the FLOPs in majority of the workloads (unless they hand craft and tune a specific GEMM kernel for their benchmark shape). But would be happy to be proven wrong, and go buy a bunch of them
"We've been negotiating a $2M contract to get AMD on MLPerf, but one of the sticking points has been confidentiality. Perhaps posting the deliverables on X will help legal to get in the spirit of open source!"
"Contract is signed! No confidentiality, AMD has leadership that's capable of acting. Let's make this training run happen, we work in public on our Discord.
It still amazes me that George/Tinycorp somehow seems to get AMD on board every time, and being blissfully unaware that they are a very small player. See for example top comment here [0].
Don't get me wrong, I think it's impressive what he achieved so far, and I hope tiny can stay competitive in this market.
Most of those willing to work with AMD are very small players (with some notable exceptions). They are likely hopeful that the small players will grow.
People get on board with George Hotz because they share the frustration of using ROCm on consumer GPUs, where the experience has been insultingly dreadful to the point where I decided to postpone buying new AMD GPUs for at least a decade.
I'm not quite sure why he decided to pivot to datacenter GPUs where AMD has shown at least some commitment to ROCm. The intersection between users of tinygrad and people who use MI350s should essentially be George himself and no one else.
Does this also ship only in x8 batches? I really liked MI300 and could afford
one of them for my research, but they only come in batches of x8 in a server rack, so I decided to buy an RTX Pro 6000.
These ppl are very loud online, but they don't make decisions for hyperscalers which are biggest spenders on AI chips.
AMD is doing just fine, Oracle just announced an AI cluster with up to 131,072 of AMD's new MI355X GPUs.
AMD needs to focus on bringing rack-scale mi400 as quickly as possible to market, rather than those hobbyists always find something to complain instead of spending money.
we're talking about the majority of open source developers (I'm one of them). if researchers don't get access to hardware X, they write their paper using hardware Y (Nvidia). AMD isn't doing fine because most low level research on AI is done purely on CUDA.
So nVidia has a huge software lead because of open source developers like you? Or because people employed by nVidia write closed source high performance drivers and kernels? Or because the people employed by Meta and Google that wrote Torch and Tensorflow built it on nVidia?
I am really sympathetic to the complaints. It would just be incredibly useful to have competition and options further down the food chain. But the argument that this is a core strategic mistake makes no sense to me.
Neither their revenue nor their market share in the space looks like just fine. What exactly in trailing the market for years is “just fine”?
AMD is very far behind, and their earnings are so low that even with a nonsensical pe ratio they’re still less than a tenth of nvidia. No, they are not doing anywhere near fine.
Are hobbyists the reason for this? I’m not sure. However, what AMD is doing is clearly failing.
When you design software for N customers, where N is very small, and you expect to hold each customers' hand individually, the software is basically guaranteed to be hot garbage that doesn't generalize or actually work except in exactly the use cases you supported (there are exceptions to this, but it requires having exceptional software engineers and leaders that care about doing things correctly and not just closing the next ticket, and in my experience, they are extremely rare).
If you design software for N00000 customers, it can't be shit, because you can't hold the hands of that many people, it's just not possible. By intending to design software for a wide variety of users, it forces you to make your software not suck, or you'll drown in support requests that you cannot possibly handle.
> These ppl are very loud online, but they don't make decisions for hyperscalers which are biggest spenders on AI chips.
this guy gets it - absolutely no one cares about the hobby market because it's absolutely not how software development is done (nor is it how software is paid for).
The hobby market should be considered as a pipeline to future customers. It doesn't mean AMD should drop everything and cater specifically to them, but they'd be foolish to ignore them altogether.
no we're not broke! we constantly write grants and receive funding from various sources. guess what hardware we recommend the University to purchase? it's 99.9% Nvidia, and sometimes Mac Studio just to play with MLX.
It has gone through many boom and busy cycles. If you go far back enough, it was very well funded. In particular, I recall reading about the US government investing 1 to 2 billion dollars during the Cold War into AI research to translate Russian into English. It had some very impressive demos on preselected Russian texts that had justified the investments. However, it failed to yield results on arbitrary texts. The translation problem has only been mostly solved in recent years.
If MI350 employs CDNA, which is based on the VEGA (GCN) architecture, does that imply that MI400, when introduced next year, will skip the 2020 GCN and directly transition to RDNA 5 equivalent?
There will be no RDNA 5, but a unified UDNA, replacing both CDNA and RDNA.
AMD has not disclosed how they will achieve the unification, but it is far more likely that the unified architecture will be an evolution of CDNA 4, i.e. an evolution of the old GCN, than an evolution of RDNA, because basing the unified architecture on CDNA/GCN, will create less problems in software porting than basing it on RDNA 4 or 3. The unified architecture will probably take some features from RDNA only when they are hard to emulate on CDNA.
While the first generation of RDNA has been acclaimed for having a good performance increase in games over the previous GCN-based Vega, it is not clear how much of that performance increase was due to RDNA being better for games and how much to the fact that the first RDNA GPUs happened to have double-width vector pipelines in comparison with the previous GCN GPUs, thus double throughput per clock cycle and per CU (32 FP32 operations/cycle vs. 16 FP32 operations/cycle).
It is possible that RDNA was not really a better architecture, but omitting some of the hardware that was rarely used in games from GCN allowed the implementation of the wider pipelines that were more useful for games. So RDNA was a better compromise for the technology available at that time, not necessarily better in other circumstances.
The identification of the AMD GPU architectures has always been extremely confusing, with tons of different names meaning the same thing and with some names, like GCN, used for several very different things.
The table linked by you is good for revealing the meaning of a part of the many AMD code names.
Not for me, I was burned twice buying laptops with AMD only to battle with their software, and even the FOSS drivers on GNU/Linux weren't that great versus the Windows experience.
While on Windows it has been hit and miss with their SDKs and shader tooling, anyone remembers RenderMonkey?
the only way for them to have any chance at catch up is to fire all the software VPs and all SW middle management, and 90% of the engineers and build the software team from ground up.
cause the team they have the last decade is clearly retarded.
On the consumer side, almost certainly not. Nvidia is a HUGE brand name, it doesn't matter how good and cheap AMD makes their consumer GPUs, people will buy Nvidia GPUs for the brand and prebuilts will stick with Nvidia for the name.
For AI chips... also probably not, unless AMD can compete with CUDA (or CUDA becomes irrelevant)
> On the consumer side, almost certainly not. Nvidia is a HUGE brand name, it doesn't matter how good and cheap AMD makes their consumer GPUs, people will buy Nvidia GPUs for the brand and prebuilts will stick with Nvidia for the name.
I think that AMD could do it, but they choose not to. If you look at their most recent lineup of cards (various SKUs of 9070 and 9060), they are not so much better than Nvidia at each price point that they are a must buy. They even released an outright bad card a few weeks ago (9060 8 GB). I assume that the rationale is that even if they could somehow dominate the gamer market, that is peanuts compared to the potential in AI.
FP6:
the main question is going to be software stack. NVIDIA is already shipping NVFP4 kernels and perf is looking good. It took a really long time after MI300X's that the FP8 kernels were OK (not even good, compared to almost perfect FP8 support in NVIDIA side of things).
I will doubt that they will be able to reach %60-70 of the FLOPs in majority of the workloads (unless they hand craft and tune a specific GEMM kernel for their benchmark shape). But would be happy to be proven wrong, and go buy a bunch of them
(related)
Tinygrad:
" https://x.com/__tinygrad__/status/1935364905949110532It still amazes me that George/Tinycorp somehow seems to get AMD on board every time, and being blissfully unaware that they are a very small player. See for example top comment here [0].
Don't get me wrong, I think it's impressive what he achieved so far, and I hope tiny can stay competitive in this market.
[0] https://news.ycombinator.com/item?id=36193625
Most of those willing to work with AMD are very small players (with some notable exceptions). They are likely hopeful that the small players will grow.
People get on board with George Hotz because they share the frustration of using ROCm on consumer GPUs, where the experience has been insultingly dreadful to the point where I decided to postpone buying new AMD GPUs for at least a decade.
I'm not quite sure why he decided to pivot to datacenter GPUs where AMD has shown at least some commitment to ROCm. The intersection between users of tinygrad and people who use MI350s should essentially be George himself and no one else.
For anyone interested in tracking max achievable matmul FLOPS for hardware and unaware, I highly recommend tracking Stas Bekman's mamf-finder results: https://github.com/stas00/ml-engineering/tree/master/compute...
Will 1.58 bits be in the MI400? Or is it not established as a widely useful technology yet?
See https://arxiv.org/abs/2402.17764
Does this also ship only in x8 batches? I really liked MI300 and could afford one of them for my research, but they only come in batches of x8 in a server rack, so I decided to buy an RTX Pro 6000.
Of course not.
AMD stubbornly refuses to recognise the huge numbers of low- or medium- budget researchers, hobbyists, and open source developers.
This ignorance of how software development is done has resulted in them losing out on a multi-trillion-dollar market.
It's incredible to me how obstinate certain segments of the industry (such as hardware design) can be.
These ppl are very loud online, but they don't make decisions for hyperscalers which are biggest spenders on AI chips.
AMD is doing just fine, Oracle just announced an AI cluster with up to 131,072 of AMD's new MI355X GPUs.
AMD needs to focus on bringing rack-scale mi400 as quickly as possible to market, rather than those hobbyists always find something to complain instead of spending money.
> these people
we're talking about the majority of open source developers (I'm one of them). if researchers don't get access to hardware X, they write their paper using hardware Y (Nvidia). AMD isn't doing fine because most low level research on AI is done purely on CUDA.
So nVidia has a huge software lead because of open source developers like you? Or because people employed by nVidia write closed source high performance drivers and kernels? Or because the people employed by Meta and Google that wrote Torch and Tensorflow built it on nVidia?
I am really sympathetic to the complaints. It would just be incredibly useful to have competition and options further down the food chain. But the argument that this is a core strategic mistake makes no sense to me.
> if researchers don't get access to hardware X, they write their paper using hardware Y (Nvidia).
There are plenty of research institutions that can easily spend >$250k on computational resources. Many routinely spend multiples of that volume.
They'll be fine.
Serious researchers use HPC clusters not desktop workstations. Currently the biggest HPC cluster in North America has AMD GPUs. I think it'll be fine.
Before they became serious researchers they were once upon a time students learning with what their laptops were capable of.
Neither their revenue nor their market share in the space looks like just fine. What exactly in trailing the market for years is “just fine”?
AMD is very far behind, and their earnings are so low that even with a nonsensical pe ratio they’re still less than a tenth of nvidia. No, they are not doing anywhere near fine.
Are hobbyists the reason for this? I’m not sure. However, what AMD is doing is clearly failing.
When you design software for N customers, where N is very small, and you expect to hold each customers' hand individually, the software is basically guaranteed to be hot garbage that doesn't generalize or actually work except in exactly the use cases you supported (there are exceptions to this, but it requires having exceptional software engineers and leaders that care about doing things correctly and not just closing the next ticket, and in my experience, they are extremely rare).
If you design software for N00000 customers, it can't be shit, because you can't hold the hands of that many people, it's just not possible. By intending to design software for a wide variety of users, it forces you to make your software not suck, or you'll drown in support requests that you cannot possibly handle.
Now imagine you don't have the resources to satisfy N00000 customers. What do you do?
> These ppl are very loud online, but they don't make decisions for hyperscalers which are biggest spenders on AI chips.
this guy gets it - absolutely no one cares about the hobby market because it's absolutely not how software development is done (nor is it how software is paid for).
The hobby market is how you get 'market default' N years later
This probably does work for the first mover. It's not clear that it can work for the underdog.
citation please
Nvidia's success.
The hobby market should be considered as a pipeline to future customers. It doesn't mean AMD should drop everything and cater specifically to them, but they'd be foolish to ignore them altogether.
startups and researchers are broke, just like Geoff Hinton in 2006 - https://blog.waqasrana.me/assets/papers/hinton2006.pdf
no we're not broke! we constantly write grants and receive funding from various sources. guess what hardware we recommend the University to purchase? it's 99.9% Nvidia, and sometimes Mac Studio just to play with MLX.
I mean broke compared to Meta or OpenAI.
AI research used to be fringe and not well funded.
Back in those days, 99.9% of hardware was Xeon.
It has gone through many boom and busy cycles. If you go far back enough, it was very well funded. In particular, I recall reading about the US government investing 1 to 2 billion dollars during the Cold War into AI research to translate Russian into English. It had some very impressive demos on preselected Russian texts that had justified the investments. However, it failed to yield results on arbitrary texts. The translation problem has only been mostly solved in recent years.
This 8-combo MI350 is a beauty with 2304 GB VRAM of HMB3E memory on each UBB [1].
[1] This is the AMD Instinct MI350:
https://www.servethehome.com/this-is-the-amd-instinct-mi350/
I've got the MI300x and I can't wait to deploy a bunch of the MI355's.
A solid 40% of George's questions were deemed great. (Not counting some fluff like what's your job.)
If MI350 employs CDNA, which is based on the VEGA (GCN) architecture, does that imply that MI400, when introduced next year, will skip the 2020 GCN and directly transition to RDNA 5 equivalent?
There will be no RDNA 5, but a unified UDNA, replacing both CDNA and RDNA.
AMD has not disclosed how they will achieve the unification, but it is far more likely that the unified architecture will be an evolution of CDNA 4, i.e. an evolution of the old GCN, than an evolution of RDNA, because basing the unified architecture on CDNA/GCN, will create less problems in software porting than basing it on RDNA 4 or 3. The unified architecture will probably take some features from RDNA only when they are hard to emulate on CDNA.
While the first generation of RDNA has been acclaimed for having a good performance increase in games over the previous GCN-based Vega, it is not clear how much of that performance increase was due to RDNA being better for games and how much to the fact that the first RDNA GPUs happened to have double-width vector pipelines in comparison with the previous GCN GPUs, thus double throughput per clock cycle and per CU (32 FP32 operations/cycle vs. 16 FP32 operations/cycle).
It is possible that RDNA was not really a better architecture, but omitting some of the hardware that was rarely used in games from GCN allowed the implementation of the wider pipelines that were more useful for games. So RDNA was a better compromise for the technology available at that time, not necessarily better in other circumstances.
2026 - MI400X - CDNA 5 - UALink/IF - Helios - HBM Bandwidth: 1,400 TB/s
https://www.tomshardware.com/pc-components/gpus/amd-says-ins...
RDNA is a dead-end.
AMD went down the wrong path by focusing on traditional rendering instead of machine learning.
I think future AMD consumer GPUs would go back to GCN.
it's all just called gcn now
https://llvm.org/docs/AMDGPUUsage.html#id38
The identification of the AMD GPU architectures has always been extremely confusing, with tons of different names meaning the same thing and with some names, like GCN, used for several very different things.
The table linked by you is good for revealing the meaning of a part of the many AMD code names.
NVDAs advantage is software, not just hardware. Would be amazing to have a competitive market but better hardware won't be enough to make it happen.
Will AMD catch up to Nvidia?
and as of lately, I really think AMD exists only for NVidia not to get slapped with antitrust lawsuits.
they played that part beautifully in the past decades for Intel
Not for me, I was burned twice buying laptops with AMD only to battle with their software, and even the FOSS drivers on GNU/Linux weren't that great versus the Windows experience.
While on Windows it has been hit and miss with their SDKs and shader tooling, anyone remembers RenderMonkey?
So NVidia it is.
the only way for them to have any chance at catch up is to fire all the software VPs and all SW middle management, and 90% of the engineers and build the software team from ground up.
cause the team they have the last decade is clearly retarded.
If they improve software quality and providing some low budget versions then - Yes.
On the consumer side, almost certainly not. Nvidia is a HUGE brand name, it doesn't matter how good and cheap AMD makes their consumer GPUs, people will buy Nvidia GPUs for the brand and prebuilts will stick with Nvidia for the name.
For AI chips... also probably not, unless AMD can compete with CUDA (or CUDA becomes irrelevant)
> On the consumer side, almost certainly not. Nvidia is a HUGE brand name, it doesn't matter how good and cheap AMD makes their consumer GPUs, people will buy Nvidia GPUs for the brand and prebuilts will stick with Nvidia for the name.
I think that AMD could do it, but they choose not to. If you look at their most recent lineup of cards (various SKUs of 9070 and 9060), they are not so much better than Nvidia at each price point that they are a must buy. They even released an outright bad card a few weeks ago (9060 8 GB). I assume that the rationale is that even if they could somehow dominate the gamer market, that is peanuts compared to the potential in AI.
Yes, if they can ship on time.
They don’t care to catch up.