Hey all,
I tested 3 new CPUs to see if I could beat the top CPU from my last test, the AMD Ryzen 9 9950x. Testing included 3 Strategy Analyzer tests and 2 Market Replay tests.
New CPUs Tested Include:
-Intel Core Ultra 9 285k
-Intel Core Ultra 7 265k
-Intel Core Ultra 7 258V
Results can be found at my YouTube channel here: https://youtu.be/2TSE8Bo7Bg0
This follows up my original CPU testing post here: NT8 speed tested on 4 different computers - General Discussions - NinjaTrader Community Forum
7 Likes
@BobbyKeefe Thank you for making yet another original, high-quality performance testing video. I’ve watched all of your NinjaTrader performance videos and really appreciate the original research. Those videos must be time-consuming to make and I hope that lots of people watch them before they make any new purchasing decisions for a trading machine.
3 Likes
If you are a scalper, and need a CPU running well…I understand this topic. But in case you need a program running 24/7, I choose an AWS server. It is so worth the money spent.
I develop on a ThinkPad 2026 model with the NVIDIA GPU, and this little beast can handle MC, Ninja, TradingView, and TS all together with multiple datafeeds and it just keeps on working. (I use it as test beds for developing)
3 Likes
Is it accurate to say that the improvement from AMD 7950X to AMD 9950X is less than 10%?
Probably not worth the time for an algo trader to spend a couple of days to setup the machine on the new processor for such little improvement of CPU performance.
Great testing @BobbyKeefe, always appreciated when someone puts in the actual work rather than speculating.
One thing worth understanding about why these results look the way they do: the core engine work NinjaTrader does during a backtest or Market Replay runs on a single thread. What actually matters is single-thread performance (i.e. clock speed / GHz), not the number of cores. More cores simply won’t help here because the engine has to process every tick (or bar) sequentially, in order. You can’t parallelize that since each data point depends on what happened before it.
This is why the gap between the processors is so small despite the newer chip being a significant generational upgrade overall. Most of those improvements are in multi-core throughput, efficiency cores, cache architecture, etc. None of that moves the needle for a single-threaded workload like a backtest.
The one place where more cores DO help is Strategy Optimization, where you’re running hundreds or thousands of parameter combinations independently of each other. Those runs have no dependency on one another so they can be farmed out across cores and you’ll see near-linear scaling. But there’s a catch even there: the number of optimization iterations needs to actually exceed the number of threads on the CPU, otherwise some cores just sit idle and you’re not getting a fair comparison between chips. If you’re only testing 10 parameter combinations on a CPU with 16 or more threads, half the chip never wakes up and the result doesn’t tell you much. To properly stress a high core-count CPU you need the workload to match or exceed its thread count.
But a single backtest pass will always be single-threaded regardless.
For anyone building a dedicated backtesting machine, the honest recommendation is to prioritize single-thread speed above everything else. By that metric, Apple’s M-series chips are actually some of the strongest performers available right now, their single-thread scores are exceptional and performance-per-watt is hard to beat. The obvious catch is NinjaTrader is Windows-only, so you’d need Parallels which introduces its own overhead.
On native Windows hardware the sweet spot remains high-clocked chips with strong IPC, which is broadly what these results confirm.
1 Like