GPUs on-demand: Not serverless, not reserved, but some third thing

By Ray Thai|6/3/2024

Tl;dr

• On-demand GPUs are a great option for scaling companies who need reliability and speed but cannot yet commit to long-term enterprise reservations of GPUs
• “Graduating” from serverless to on-demand deployments starts to make sense economically when you are running ~100k+ tokens per minute
• On identical H100 hardware, Fireworks’ software advantage provides 53% cost reduction and 60% latency reduction compared to running vLLM with GPUs on competing platforms. This allows you to simultaneously serve more users, at lower cost, with faster response times
• Fireworks on-demand deployments require no software installation. Simply choose your model and GPU configuration and get started in seconds, without paying for boot times. Your GPU even scales to zero when you are not using it!

Intro

One of the most rewarding things at Fireworks is being a part of the scaling journey for many, exciting AI start-ups. Over the last few months, we’ve seen an explosion in the number of companies beginning to productionize generative AI. A question that we commonly get is:

“How should I think about serving LLMs via (1) A Serverless, token-based options vs (2) A GPU, usage time-based option? “

We’re writing this post to help explain the tradeoffs of serverless vs dedicated GPU options.

Fireworks Offerings

Fireworks has 3 offerings for LLM serving: (1) Serverless (2) On-demand (3) Enterprise.

Serverless

Fireworks hosts our most popular models “serverless”, meaning that we operate GPUs 24/7 to serve these models and provide an API for any of our users to use the models. Our serverless offering is the fastest, widely available platform and we’re proud of its production-readiness. Serverless is the perfect option for running limited traffic and experimenting with different LLM set-ups. However, serverless has limitations:

1. Serverless is not personalized for you Your LLM serving can be made faster, higher-quality or lower cost based on personalization on several levels:

   • GPU configuration - The quantity, type and configuration of GPUs can be tailored based on business needs, like optimizing for speed vs cost or providing more capacity for product launches
   • Serving stack - The software itself can be tuned based on factors like prompt length and UX goals (latency vs throughput, etc)
   • Models - Serverless model selection is curated by Fireworks, so niche models may not be available

   Our serverless platform is designed to excel at a variety of goals and support hundreds of base models and fine-tuned models. However, a personally-tailored stack still may provide better experiences.

2. Serverless performance is affected by others - Other Fireworks users share our serverless deployment, so speeds vary depending on overall usage. If you happen to use the deployment at the emptiest hours, you’ll experience the fastest speeds and vice versa. Our public platform is still independently benchmarked to have the lowest variation in latency, but consistent performance is paramount for certain use cases, like live voice chat agents.

3. Serverless has volume constraints - Generally, when businesses have the volume to use significant GPU capacity, it doesn’t make sense to use serverless because:

   • Rate limits don’t allow it - Since GPU deployments are shared, we employ rate limits to ensure that a few actors can’t greatly affect everyone else’s experience. Our rate limits (600 RPM by default, 2000 for Business Tier) are significantly higher than rate limits of other serverless inference providers (often below 50 RPM) but businesses with significant volume could struggle to run their app entirely on Fireworks
   • Private GPUs may be cheaper with large volume - Given (a) Efficiency improvements from personalized set-ups and (b) GPU pricing that builds in “bulk discounts” compared to serverless, large businesses generally receive cheaper pricing by reserving their own GPU

Enterprise Reserved GPUs

Given these constraints, companies with large usage volumes often reserve their own private GPU(s) for set periods of time. This commitment also enables Fireworks to help companies personally configure their serving set-up and provide SLAs and guaranteed support. However, many scale-up companies in the midst of prototyping are unable to commit to enterprise reserved capacity.

On-demand GPUs

To make it easier for scaling teams to benefit from Fireworks, we offer on-demand, dedicated GPUs with our FireAttention stack. Users pay per hour and can scale their GPU usage up or down automatically based on traffic. Configurations can automatically scale up and down from 0 GPUs, so developers pay nothing during idle periods. Compared to serverless, using your own GPU can provide:

• Guaranteed speed and reliability by using private GPUs
• More model choice - import your custom base model or use dozens of Fireworks-provided base models
• Improved speed - configure the number of set-up of GPUs based on your goals and optimize the Fireworks serving stack based on your prompt length
• Lower costs, especially with high volume. Costs reduce as volume increases.
• Capacity for higher request volume, with no hard rate limits

On-demand GPUs performance details

An important consideration in deciding to use on-demand deployments is expected performance vs price, so we have included some performance details and FAQs.

What latency improvements can I expect compared to vLLM or hosting my own GPU?

Generally, we see that Fireworks is ~40-60% faster compared to open-source solutions but performance varies depending on model and workload. We obtained the below results using the min worldsize (minimum # GPUs to host a specific model) on H100 GPUs on Fireworks vs vLLM software. The latency was calculated with heavy throughput, so the GPU on Fireworks simultaneously delivered significantly better speed and throughput.

Mixtral 8x7b

Llama 3 8B

What throughput/cost improvements can I expect compared to vLLM or hosting my own GPU?

The efficiency of the FireAttention serving stack means that the same H100/A100 chip running on FireAttention can handle significantly more volume than an H100/A100 running software like vLLM.

• GPU:GPU comparison: Generally, we observed that an H100 running FireAttention provides ~350% the capacity of an H100 running vLLM
• Cutting # GPUs: Therefore, if you were running 10 GPUS on vLLM, you could run ~3 GPUs on Fireworks.
• Cutting costs: Thus, if you were running an H100 at $4.69/hour (pricing similar to competing platforms), you’d save ~53% by switching to Fireworks H100s priced at $7.79/hour by reducing your number of total GPUs
• Detailed numbers: See below for our calculations on cost per token when running on H100(s) at capacity. We achieved the numbers using a prompt with input lengths of 4000 and output lengths of 200, but found similar results across different prompts.

How else is running a GPU on Fireworks different from provisioning my own GPU with vLLM?

• The best serving stack, managed and updated for you: Fireworks is constantly improving our serving stack and our on-demand deployments (see our June performance updates). By choosing Fireworks, you’ll get the world’s most efficient serving stack automatically managed for you
• Set-up in seconds: Get a hyper-optimized set-up in seconds instead of spending hours configuring software on your own GPU. No paying for start-up times either on Fireworks. Simply specify the model and deployment configuration and get started right away, saving you both time and money
• Production reliability: Fireworks only utilizes the most reliable Cloud infrastructure providers for our GPUs. The same infrastructure for on-demand powers companies like Uber, Doordash and Cursor
• Compatibility with tuning and other Fireworks services: Optimize your on-demand serving experience with Fireworks innovations, like our fine-tuning service, exclusive models and structured outputs

I’m using Fireworks’ serverless model hosting. When does it make sense for me to "graduate" to on-demand?

• Tokens/minute: Cost-wise, we generally see that it makes sense to switch when achieving volumes nearing ~500,000 tokens/minute.
• RPM: 500,000 tokens/minute translates to 100 RPM with 5000 input + output tokens per request, or RPM of 500 RPM with 1000 tokens per request. Generally this “break-even” point is well-below the 2000 RPM and 600 RPM Business and Developer tier limit
• Unit economics of switching: Cost per tokens when using on-demand deployments continue to decrease until approaching the figures above of $0.23/1M tokens with Mixtral and $0.092/1M tokens with Llama 3 8B

Conclusion

Using GPUs on-demand can provide a cheaper, faster and more reliable solution than serverless, especially when operating with higher volume. On-demand GPUs can act as a great bridge to a longer-term reserved GPU solution, where developers can benefit from completely personalized set-ups and enterprise SLAs. Fireworks is committed to enabling businesses to scale on our platform from testing an initial prototype on serverless through to serving millions of end users on reserved GPUs.

When you’re ready to get started, check out our docs or hear about our latest product updates for on-demand in this blog post. Looking for more performance numbers or interested in chatting directly about on-demand deployments? Feel free to directly schedule time with our PM (https://calendly.com/raythai). We can’t wait to see what you build!