Wow, 44GB SRAM, not HBM3 or HBM3e, but actual SRAM ...
1x MI300x has 192GB HBM3.
1x MI325x has 256GB HBM3e.
They cost less, you can fit more into a rack and you can buy/deploy at least the 300's today and 325's early next year. AMD and library software performance for AI is improving daily [0].I'm still trying to wrap my head around how these companies think they are going to do well in this market without more memory.
> I'm still trying to wrap my head around how these companies think they are going to do well in this market without more memory.
Cerebras and Groq provide the fastest (by an order of magnitude) inference. This is very useful for certain workflows, which require low-latency feedback: audio chat with LLM, robotics, etc.
Outside that narrow niche, AMD stuff seems to be the only contender to NVIDIA, at the moment.
> Cerebras and Groq provide the fastest (by an order of magnitude) inference.
Only on smaller models, their numbers are all 70b in the article.
Those numbers also need to be adjusted for the comparable amounts of capex+opex costs. If the costs are so high that they have to subsidize the usage/results, then they are just going to run out of money, fast.
> Only on smaller models, their numbers are all 70b in the article.
No, they are 5x-10x faster for all the model sizes (because it's all just running from SRAM and they have more of it than NVIDIA/AMD), even though they benchmarked just up to 70B.
> Those numbers also need to be adjusted for the comparable amounts of capex+opex costs. If the costs are so high that they have to subsidize the usage/results, then they are just going to run out of money, fast.
True. Although, for some workloads, fast enough inference is a strict prerequisite and GPUs just don't cut it.
You are right assuming that model capabilities are determined only by model size. But consider that OpenAI is saying they have a way of scaling intelligence with inference time compute, not just model size. If that proves out, reducing latency per output token potentially becomes as valuable as or even possibly more valuable than scaling model size. Speed becomes intelligence. And Cerebras has 1/10 the latency per token of anything else.
On Google Cloud a server with 8 TPU v5e will do 2175 token/seconds on Llama2 70B.
https://cloud.google.com/blog/products/compute/updates-to-ai...
From https://cloud.google.com/tpu/pricing and https://cloud.google.com/vertex-ai/pricing#prediction-prices (search for ct5lp-hightpu-8t on the page) the cost for that appears to be $11.04/hr which is just under $100k for a year. Or half that on a 3-year commit.
That seems like a better deal than millions for a few CS-3 nodes.
And they've just announced the v6 TPU:
Compared to TPU v5e, Trillium delivers:
Over 4x improvement in training performance
Up to 3x increase in inference throughput
A 67% increase in energy efficiency
An impressive 4.7x increase in peak compute performance per chip
Double the High Bandwidth Memory (HBM) capacity
Double the Interchip Interconnect (ICI) bandwidth
https://cloud.google.com/blog/products/compute/trillium-sixt...You're correct on $/bandwidth. The point about low latency continues to be ignored, though.
That's a benchmark or shower thought?
> If models keep lasting ~year timescales could we ever see people going with ROM chips for the weights instead of memory?
Before ROM, there's a step where HBM for weights is replaced with Flash or Optane (but still high bandwidth, on top of the chip) and KV cache lives in SRAM - for small batch sizes, that would actually be decently cheap. In this case, even if weights change weekly, it's not a big deal at all.
How is this a narrow niche?
Chain of thought type operations is in this "niche".
Also anything where the value is in the follow up chat not the one shot.
Groq and Cerebras only make sense at massive scale which is why I guess they pivoted to being API providers so they can amortize the hardware over many customers.
Correct except that massive scale doesn't work cause it just uses up exponentially more power/space/resources.
They also have a very limited use case... if things ever shift away from LLM's and into another form of engineering that their hardware does not support, what are they going to do? Just keep deploying hardware?
Slippery slope.
The article explains in depth the issues with memory, did you read through ?
2x 80GB A100 is better in all the metrics than MI300x while being cheaper.
clickbait title: inference is not training
The value proposition of Cerebras is that they can compile existing graphs to their hardware and allow inference at lower costs and higher efficiencies. The title does not say anything about creating or optimizing new architectures from scratch.
the title says "Cerebras Trains Llama Models"...
That's correct and if you read the whole thing you will realize that it is followed by "... to leap over GPUs" which indicates that they're not literally referring to optimizing the weights of the graph on a new architecture or freshly initialized variables on existing ones.
What are you confused about? Their value proposition is very simple and obvious, custom hardware with a compiler that transforms existing graphs into a format that can run at lower cost and higher efficiency because it utilizes a special instruction set only available on Cerebras silicon.
Title is about training.... article about inference
Why is nobody mentioning that there is no such thing as Llama 3.2 70B
"It would be interesting to see what the delta in accuracy is for these benchmarks."
^ the entirety of it
did they release MLPerf data yet or wouldn't help their IPO?
"So, the delta in price/performance between Cerebras and the Hoppers in the cloud when buying iron is 2.75X but for renting iron it is 5.2X, which seems to imply that Cerebras is taking a pretty big haircut when it rents out capacity. That kind of delta between renting out capacity and selling it is not a business model, it is a loss leader from a startup trying to make a point."
As always, it is about TCO, not who can make the biggest monster chip.